11. CLI reference¶
Commands in this reference (subcommands indented):
agentprovagentprov info(alias sources)agentprov validateagentprov import(alias convert)agentprov fetchagentprov runagentprov modelsagentprov backendsagentprov backends buildagentprov migrateagentprov migrate verify-baselineagentprov migrate gate-aagentprov migrate gate-bagentprov migrate gate-dagentprov migrate gate-eagentprov migrate gateagentprov renderagentprov inspectagentprov purgeagentprov statsagentprov packageagentprov deriveagentprov recomputeagentprov verifyagentprov settingsagentprov settings listagentprov settings getagentprov settings setagentprov settings unsetagentprov keysagentprov keys statusagentprov keys listagentprov keys setagentprov keys deleteagentprov pricesagentprov prices listagentprov prices getagentprov prices setagentprov prices unsetagentprov prices refreshagentprov estimateagentprov jobsagentprov jobs startagentprov jobs listagentprov jobs showagentprov jobs followagentprov jobs stopagentprov jobs drainagentprov jobs resumeagentprov jobs killagentprov jobs pruneagentprov batchagentprov batch createagentprov batch startagentprov batch resumeagentprov batch runagentprov batch stopagentprov batch listagentprov batch showagentprov devrunagentprov devrun planagentprov devrun startagentprov devrun listagentprov devrun pruneagentprov assist
agentprov¶
usage: agentprov [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N] [--control-file PATH]
[--show-content] [--content-chars N] [--json] [--summary-json PATH]
[--max-listed N] [--no-colour] [--break-stale-lock]
COMMAND ...
AgentPROV: execution provenance for LLM-agent security research.
positional arguments:
COMMAND
info (sources) versions, the source registry, installed extras and tools, host and
isolation grade
validate validate AER files (tier 2) or a corpus's stored outputs (tiers aer, sem,
prov, linkage)
import (convert) ingest released artefacts into a working corpus (store, map, seal, derive,
validate)
fetch fetch pinned upstream bytes into the verified cache (needs --allow-
network)
run live capture through a backend worker (scripted or real model; --dry-run
prints the plan)
models list the models a local server serves (and one model's observed identity),
or an API provider's list
backends backend environments (build)
migrate baseline migration gates G0, A, B, D and E (J.3)
render draw selected executions' PROV graphs (needs the render extra and Graphviz
dot)
inspect look at one execution, artefact or corpus (read-only; content hidden
unless --show-content)
purge remove a leaked object from a working corpus or the fetch cache, leaving a
tombstone (irreversible)
stats generic statistics over corpora: counts, PROV statements, unknowns, field-
group coverage
package seal a working corpus (DIR.partial -> DIR), or build an offline deposit
draft from sealed corpora
derive add missing derived layers (SEM, PROV, validation, views, compat, TriG) as
a new corpus version
recompute re-derive a corpus from its raw objects or stored AERs and compare bytes
(H.8.4)
verify verify a corpus or a deposit: checksums, the raw -> AER -> SEM -> PROV
digest chain, or a re-derivation
settings list, get, set or unset the shared user settings
keys endpoint keys in the OS keychain (names only are ever shown)
prices list, set or refresh the model prices the estimate uses
estimate the cost and time of a live run before it starts (p50/p90; no model call)
jobs start, list, follow, stop or prune detached jobs
batch ordered job lists that run as one job
devrun dev runs: a few samples of every part, imported and checked
assist ask the AgentPROV Assistant how to do something (a model answers from
AgentPROV's docs and code)
options:
-h, --help show this help message and exit
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov info¶
Alias: sources.
usage: agentprov info [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N] [--control-file PATH]
[--show-content] [--content-chars N] [--json] [--summary-json PATH]
[--max-listed N] [--no-colour] [--break-stale-lock] [--source SOURCE]
[--matrix | --must-not-claim] [--format {text,markdown,json}]
[SOURCE]
Show versions, the source registry (per part, gated statuses), installation and host (N.2.9).
positional arguments:
SOURCE a registered source id: parts, profiles, pins, licence
options:
-h, --help show this help message and exit
--source SOURCE same as SOURCE
--matrix the compatibility matrix (needs the claims library; deferred)
--must-not-claim the must-not-claim items (need the claims library; deferred)
--format {text,markdown,json}
payload format (default text)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov validate¶
usage: agentprov validate [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--aer PATH] [--store DIR] [--corpus DIR]
[--tiers LIST] [--prov-spec SPEC] [--strict | --release] [--report PATH]
[--sample N] [--seed S] [--card CARD.md] [--jobs N] [--in-process]
[--no-validation-cache] [--dry-run]
[FILE ...]
Validate AgentPROV Execution Record files (tier 2), or re-run the tiers aer, sem, prov and linkage
over a corpus's stored outputs, read-only (--corpus; N.2.5). Checks that cannot run are reported
as skipped, never as passed.
positional arguments:
FILE files to validate (type detected by suffix: .aer.jsonl)
options:
-h, --help show this help message and exit
--aer PATH an AER file, whatever its suffix (repeatable)
--store DIR raw store directory (objects/<h0h1>/<sha256>); without it locator checks
are skipped
--corpus DIR validate a corpus's stored outputs, read-only (sealed, working or
.partial)
--tiers LIST comma list from aer,sem,prov,linkage,shacl,regression (default
aer,sem,prov,linkage)
--prov-spec SPEC corpus mode: the PROV spec (agentprov-prov/0.1)
--strict a skipped check fails
--release implies --strict, release-mode validator, every applicable tier
--report PATH corpus mode: write the report to PATH (outside the corpus; - = stdout)
--sample N corpus mode: validate a seeded sample of N executions
--seed S sampling seed
--card CARD.md lint a corpus card with the claims library (N.2.5(7))
--jobs N corpus mode: worker processes for tiers sem, prov and linkage (0 = host
default)
--in-process corpus mode: run every tier in this process (no workers; allowed with
--release, AD-10 row 17)
--no-validation-cache
corpus mode: always run the tier-4 PROV-O round trip; neither read nor
write the round-trip cache under $AGENTPROV_HOME/cache/tier4 (D-550).
--release never reads it either: it recomputes every round trip and audits
the entries it finds
--dry-run accepted (file validation never writes)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov import¶
Alias: convert.
usage: agentprov import [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour] [--break-stale-lock]
[--corpus DIR] [--input PATH] [--fetch] [--allow-network]
[--cache-dir DIR] [--verify-against-pin] [--i-consent-to-large-download]
[--context ROLE=PATH] [--no-context ROLE] [--profile ID] [--pin ROLE]
[--snapshot ID] [--spec SPEC.toml] [--corpus-id ID] [--suite S]
[--attack A] [--pipeline P] [--variant V] [--select KEY=VALUE[,VALUE...]]
[--limit N] [--sample N] [--sample-mode {first,spread,random}] [--seed S]
[--until {raw,aer,sem,prov,validation}] [--prov-spec SPEC]
[--retry-failed] [--release] [--skip-prov-conformance] [--jobs N]
[--in-process] [--size-hints DIR] [--allow-hardlinks]
[--no-validation-cache] [--dry-run] [--artefacts] [--max-file-bytes N]
[--max-json-depth N]
[SOURCE]
Ingest released artefacts into DIR.partial (N.2.2): raw store, AER, PROV and validation.
positional arguments:
SOURCE registry source id or alias; optional with --spec
options:
-h, --help show this help message and exit
--corpus DIR the working corpus directory (written as DIR.partial)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
origin:
--input PATH local origin: a directory or one file, read in place
--fetch read the profile's pin from the verified fetch cache (offline unless
--allow-network)
--allow-network with --fetch: fetch the pin when the cache lacks it (otherwise refused:
raw.fetch.not_cached)
--cache-dir DIR fetch cache (default $AGENTPROV_UPSTREAM_CACHE, else
$AGENTPROV_HOME/cache)
--verify-against-pin record the pinned origin for --input files the pin vouches for
--i-consent-to-large-download
with --fetch of a download pin: consent to a download of 100 MiB or more
(U-07)
--context ROLE=PATH a local copy of a declared context pin (checked against the pin); default:
the fetch cache
--no-context ROLE map without a declared context role (its per-unit artefacts are then
absent; a notice, and the role is recorded in the snapshot's
context_skipped)
profile and snapshot:
--profile ID import profile (default: the registry's profile for the part)
--pin ROLE profile pin role, e.g. 2024-06 or release
--snapshot ID snapshot id inside the corpus (default: the registry part)
--spec SPEC.toml committed corpus spec (deferred in this build)
--corpus-id ID corpus id (default: the registry's corpus id for the part)
selection:
--suite S
--attack A
--pipeline P
--variant V
--select KEY=VALUE[,VALUE...]
AER-header filter (repeatable)
--limit N map at most N new units (0 maps nothing)
--sample N dev runs: a deterministic sample of N executions, chosen before mapping
(D-529; partial corpus)
--sample-mode {first,spread,random}
with --sample: first N by key (default), round-robin over the reader's
arms, or seeded random
--seed S with --sample-mode random: the seed (recorded in the manifest)
stages:
--until {raw,aer,sem,prov,validation}
stop after this stage (default: validation)
--prov-spec SPEC PROV spec (repeatable; default agentprov-prov/0.1)
mode:
--retry-failed re-process units whose row has status failed
--release release mode: strict checks, no filters, no --limit, no --in-process
--skip-prov-conformance
dev mode only, not advisable: skip the W3C PROV conformance check (tier
4); the corpus is marked unvalidated until a full recompute --out NEW, and
cannot be sealed (G2c-5)
--jobs N worker processes (0 = host default; default 0)
--in-process debugging only: run expand, map and prov in this process
--size-hints DIR a previous version of this corpus (read-only): its stored PROV sizes set
each prov job's memory charge, so more jobs run at once (D-554; timing
only, never outputs)
--allow-hardlinks accept st_nlink > 1 files in an --input tree
--no-validation-cache
do not store the tier-4 round trips in $AGENTPROV_HOME/cache/tier4 (D-550)
--dry-run run every in-memory stage; write nothing
--artefacts with --dry-run: scan the mapped executions for source-data artefacts and
print the per-class table (D.15.4)
tier-1 overrides:
--max-file-bytes N
--max-json-depth N
agentprov fetch¶
usage: agentprov fetch [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour] [--break-stale-lock]
[--corpus DIR] [--profile ID] [--only sha256:HEX] [--pin ROLE] [--history]
[--path PREFIX] [--cache-dir DIR] [--allow-network]
[--i-consent-to-large-download] [--dry-run]
Populate the verified, extraction-free fetch cache from in-package pins (N.2.2; G.7.5).
options:
-h, --help show this help message and exit
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
what to fetch (exactly one):
--corpus DIR fetch the objects of DIR/raw/REFERENCES.json (the corpus is never written)
--profile ID fetch the pins of a shipped source profile
selection:
--only sha256:HEX with --corpus: only this object (repeatable)
--pin ROLE with --profile: a pin role (repeatable; default: the primary pins)
--history with --profile: fetch the blobless full history of the git pins and
compute the introducing-commit index (D.6.16)
--path PREFIX with --profile: a path prefix inside a git pin (repeatable)
network and cache:
--cache-dir DIR cache directory (default $AGENTPROV_UPSTREAM_CACHE, else
$AGENTPROV_HOME/cache)
--allow-network permit the network access fetching needs (required unless --dry-run)
--i-consent-to-large-download
consent to a download of 100 MiB or more (U-07)
--dry-run print the plan; no network and no writes
agentprov run¶
usage: agentprov run [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N] [--control-file PATH]
[--show-content] [--content-chars N] [--json] [--summary-json PATH]
[--max-listed N] [--no-colour] [--break-stale-lock] [--corpus DIR]
[--campaign NAME | --corpus-id ID] [--plan PLAN.toml] [--spec SPEC.toml]
[--suite S[,S...]] [--benchmark-version V] [--user-tasks SPEC]
[--attack A[,A...]] [--injection-tasks SPEC]
[--benign {include,exclude,only}] [--defence NAME] [--task NAME]
[--task-param KEY=VALUE] [--samples limit:N|ID[,ID...]] [--epochs N]
[--message-limit N] [--token-limit N] [--time-limit SECONDS]
[--max-connections N] [--cell-size N] [--judge ROLE=MODEL]
[--model SHORTHAND] [--endpoint NAME] [--model-id ID]
[--defence-endpoint NAME] [--llm-script SCRIPT.json]
[--llm-element {native,gateway_compat}] [--model-display-name NAME]
[--temperature T] [--reasoning-effort LEVEL] [--max-tokens N|harness-default]
[--temperature-policy {as-benchmark,inject}] [--allow-text-tool-calls]
[--base-url URL] [--accept-served-model ID] [--allow-lan-endpoint]
[--no-hash-weights] [--repeats N] [--max-attempts N] [--retry-inconsistent]
[--allow-mixed-config] [--limit N] [--canary N] [--preflight-only]
[--dry-run] [--plan-json] [--catalogue] [--keep-staging]
[--cell-timeout SECONDS] [--jobs N] [--start-stagger SECONDS]
[--env-file PATH] [--env-stdin] [--keychain] [--allow-custom-endpoint]
[--skip-prov-conformance] [--i-consent-to-api-costs]
[--i-consent-to-local-run] [--containment {sandbox-exec,dedicated-user,vm}]
[--autodojo-cache PATH] [--autodojo-variant K] [--allow-network]
[--max-calls N] [--max-cost-usd USD] [--price-in USD_PER_MTOK]
[--price-out USD_PER_MTOK] [--preset NAME] [--m5ultra] [--model-dir PATH]
[--model-size-gb GB] [--active-params-b B] [--context-tokens N]
[--kv-gb-per-trace GB] [--server-max-concurrency N] [--tune-jobs N[,N...]]
[--tune-out DIR]
SOURCE
Execute a live backend and capture raw evidence (N.2.1, N.2.1b). Needs a built backend environment
('agentprov backends build'); real-model runs also need .env, --containment and consent.
positional arguments:
SOURCE live source: agentdojo, agentdyn, agentharm or autodojo
options:
-h, --help show this help message and exit
--corpus DIR working corpus directory
--campaign NAME corpus id <source>-live-<NAME>
--corpus-id ID
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
cells (exclusive with --plan):
--plan PLAN.toml committed run plan
--spec SPEC.toml corpus spec with [live]
--suite S[,S...]
--benchmark-version V
--user-tasks SPEC
--attack A[,A...]
--injection-tasks SPEC
--benign {include,exclude,only}
--defence, --defense NAME
Inspect plans (an inspect-<benchmark> backend; D-663; exclusive with the cell options):
--task NAME Inspect task (the catalogue lists them), e.g. inspect_evals/agentharm
--task-param KEY=VALUE
a task parameter the catalogue declares (repeatable; VALUE as a TOML
scalar or a string)
--samples limit:N|ID[,ID...]
the samples: the first N of the split, or ids
--epochs N
--message-limit N per sample (mandatory for Inspect, D-667)
--token-limit N per sample (mandatory for Inspect, D-667)
--time-limit SECONDS per sample (mandatory for Inspect, D-667)
--max-connections N concurrent model calls of one cell
--cell-size N samples per cell (one Inspect eval per cell)
--judge ROLE=MODEL the model of a judge role of the task's scorer (repeatable; every role is
required, D-667); judges use the agent's endpoint and key
model (one of: --model, --endpoint with --model-id, --llm-script):
--model SHORTHAND endpoint shorthand from endpoints.toml
--endpoint NAME explicit endpoint (deepseek, openai, openrouter, anthropic, local, ollama,
omlx, lmstudio, mlx_lm, llamacpp, vllm)
--model-id ID
--defence-endpoint NAME
the endpoint (endpoints.toml shorthand) a defence calls for its own model
or embedding calls (D-675; MELON's embeddings); only its key enters the
worker
--llm-script SCRIPT.json
scripted transport: no network, no keys (test_only)
--llm-element {native,gateway_compat}
--model-display-name NAME
--temperature T
--reasoning-effort LEVEL
--max-tokens N|harness-default
Anthropic endpoints (required there): the reply's token cap, stated (N) or
the harness default 16000 (harness-default, recorded as a harness policy);
AgentDojo's own 1024 is never used silently (D-656)
--temperature-policy {as-benchmark,inject}
--allow-text-tool-calls
run an endpoint that returns text-encoded tool calls
local model servers (D-691..D-694):
--base-url URL a local endpoint's base URL (loopback; overrides .env and the registry
default; recorded)
--accept-served-model ID
a served model id that may differ from the requested one (repeatable;
recorded; A12)
--allow-lan-endpoint allow a local endpoint on a literal private LAN address (D-698): prompts
and replies leave this machine over plain http; keyless; recorded
--no-hash-weights record a local model's weight files by size only (weights_hashed false)
control:
--repeats N
--max-attempts N
--retry-inconsistent
--allow-mixed-config
--limit N
--canary N canary cells (default 10; 0 disables)
--preflight-only run the endpoint preflight probes and exit
--dry-run resolve and print the plan; run nothing
--plan-json with --dry-run: print the expanded plan, one row per cell, as one JSON
document on stdout (D-651; names and counts only)
--catalogue print the suites, user and injection tasks and attacks the backend offers
for --benchmark-version as one JSON document on stdout (D-652); no model
call, nothing written
--keep-staging keep each session's staging directory after it is stored (inspection;
K.8.5 live_artefacts)
--cell-timeout SECONDS
per-cell wall clock before the worker is killed and the session sealed
aborted (default 1800)
--jobs N concurrent traces (>= 1; default 1, or computed with a local preset)
--start-stagger SECONDS
spread the first cell start of each worker
environment and consent:
--env-file PATH
--env-stdin read the endpoint variables as one JSON object from stdin (D-395; for a
GUI with a keychain); exclusive with --env-file, and the default .env is
not read
--keychain read the selected endpoint's key from the OS keychain (agentprov keys;
D-702); exclusive with --env-file and --env-stdin, and the default .env is
not read
--allow-custom-endpoint
--skip-prov-conformance
not advisable: skip the W3C PROV conformance check (tier 4) of each
session; the corpus is marked unvalidated until a full recompute --out
NEW, and cannot be sealed (G2c-5)
--i-consent-to-api-costs
--i-consent-to-local-run
consent for a run against a local model server (D-726); not consent for an
API endpoint
--containment {sandbox-exec,dedicated-user,vm}
--autodojo-cache PATH
AutoDojo attack cache copied from the pinned fork (verified against its
tree; D.11.3 rule 5)
--autodojo-variant K the cache's variant index (integer >= 0)
--allow-network let T0 fetch the pinned AutoDojo commit into the fetch cache to verify
--autodojo-cache
API caps (D-725: required for an API endpoint, optional otherwise):
--max-calls N billed model calls of the corpus at most, preflight probes included; a
cell starts only if its worst case fits
--max-cost-usd USD spend of the corpus at most (an upper bound priced from the recorded
usage); needs --price-in and --price-out
--price-in USD_PER_MTOK
USD per million input-side tokens (the highest of input, cache write and
cached)
--price-out USD_PER_MTOK
USD per million output tokens
local models and presets (N.2.1a):
--preset NAME host preset from package data (v0.1: m5ultra)
--m5ultra shorthand for --preset m5ultra
--model-dir PATH read the model's config.json and weight sizes (untrusted input)
--model-size-gb GB weight size, decimal gigabytes
--active-params-b B active parameters, billions
--context-tokens N
--kv-gb-per-trace GB
--server-max-concurrency N
the local server's concurrent-request limit
--tune-jobs N[,N...] the concurrency sweep (N.2.1a(4), D-820): the same cell sample at each
--jobs value, one working corpus per value, and a report recommending the
smallest N within 90 % of the peak traces per hour with no timeouts (a
local endpoint only)
--tune-out DIR the sweep's folder (default $AGENTPROV_HOME/calibration/tune-jobs-<UTC
stamp>)
agentprov models¶
usage: agentprov models [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour] [--break-stale-lock]
[--endpoint NAME] [--base-url URL] [--model-id ID] [--model-dir PATH]
[--no-hash-weights] [--find-folders] [--allow-lan-endpoint]
[--allow-network] [--refresh-if-stale] [--keychain] [--env-file PATH]
[--env-stdin] [--format {text,json}]
Local model servers (loopback only; read-only, no inference): the models each local endpoint of
endpoints.toml serves, or one server's, and with --model-id the identity record a live run records
(D-691). API providers (D-730): the dated cached model list; --allow-network fetches it now (no
tokens, no cost).
options:
-h, --help show this help message and exit
--endpoint NAME an endpoint shorthand: local (ollama, omlx, lmstudio, mlx_lm, llamacpp,
vllm) or an API provider (openrouter, anthropic, openai, deepseek, gemini,
llama, qwen, glm)
--base-url URL a loopback base URL (overrides the endpoint's; the server kind is
fingerprinted)
--model-id ID show the observed identity of this served model
--model-dir PATH the model's directory or file, where the server does not name it (oMLX,
llama.cpp); hashed, never parsed
--no-hash-weights record weight files by size only (weights_hashed false)
--find-folders with --model-id: list the folders on this machine that may hold the
model's files (oMLX, the Hugging Face cache, LM Studio; names only,
nothing opened, no server contacted; D-828)
--allow-lan-endpoint allow --base-url on a literal private LAN address (D-698; data leaves this
machine)
--format {text,json} payload format (default text)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
API providers (D-730..D-735):
--allow-network fetch the provider's model list now (R2; the provider's registry host
only) and cache it
--refresh-if-stale fetch only when the setting models.auto_refresh is on and the cached list
is older than 24 h; otherwise show the cache
--keychain read the endpoint's key from the OS keychain (D-702): an API provider's,
or a local server's optional key (D-827); exclusive with --env-file/--env-
stdin
--env-file PATH read the endpoint's key variable from this .env file
--env-stdin read the endpoint variables as one JSON object from stdin (D-395)
agentprov backends¶
usage: agentprov backends [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock]
ACTION ...
Backend environments for live runs (N.2.10).
positional arguments:
ACTION
build build $AGENTPROV_HOME/envs/<backend>-<commit8>/ with uv sync --locked
--no-config
options:
-h, --help show this help message and exit
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov backends build¶
usage: agentprov backends build [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N]
[--json] [--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--offline-check] [--allow-network]
[--defence-assets ID[,ID...]] [--defence-source DIR]
[--env-file PATH] [--dry-run]
BACKEND
Build a backend environment from backends/<backend>/ (N.2.10). Without --allow-network the sync
runs offline from the uv cache.
positional arguments:
BACKEND agentdojo, autodojo, agentdyn, inspect-agentharm
options:
-h, --help show this help message and exit
--offline-check verify an existing build; no network, no writes
--allow-network let uv download packages (otherwise --offline)
--defence-assets ID[,ID...]
also provision these registry defences' pinned code from the fetch cache
into the environment (D-670;
backends/shared/src/agentprov_backend/defences.toml)
--defence-source DIR read --defence-assets' pinned code from this local tree instead of the
fetch cache; each file is accepted only by its pinned sha256 and git blob
id (no network)
--env-file PATH the .env file (not read by the agentdojo build)
--dry-run print the plan; run and write nothing
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov migrate¶
usage: agentprov migrate [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour] [--break-stale-lock]
GATE-COMMAND ...
Run the baseline migration gates (N.2.14; J.3.0). Implemented: verify-baseline (G0), gate-a,
gate-b, gate-d, gate-e.
positional arguments:
GATE-COMMAND
verify-baseline gate G0: verify the old repository, its corpora and the baseline
environment against the lock
gate-a gate A: v1 -> AER -> v1 view, A1 (JCS) and A2 (bytes), from the AER alone
gate-b gate B: adprov-v2.3-compat graph bytes (B1 from the view, B2 from the
source)
gate-d gate D: SEM and agentprov-prov/0.1 re-derived from stored AERs and
validated in every D8 tier (release mode)
gate-e gate E: the semantic diff of the compat graphs against agentprov-prov/0.1,
from semdiff-map.toml
gate run one gate by name: G0, A, B, D or E
options:
-h, --help show this help message and exit
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov migrate verify-baseline¶
usage: agentprov migrate verify-baseline [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}]
[--progress-file PATH] [--progress-interval SECONDS]
[--progress-max-bytes N] [--control-file PATH]
[--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--scope {golden,full}]
[--workers N] [--release] [--report PATH]
[--reports-dir DIR] [--config FILE] [--in-process]
gate G0: verify the old repository, its corpora and the baseline environment against the lock
options:
-h, --help show this help message and exit
--scope {golden,full}
golden: the golden set (CI); full: the baseline trees (local). Default:
golden (G0: full)
--workers N worker processes for mapping (0 = host default; default 8 or [run] workers
of the config)
--release release mode: params.strict true, skipped checks fail, T1 isolation grade
must be watchdog or better
--report PATH write the gate report to PATH (atomic); default <reports-
dir>/<gate>/<date>-<commit>.json
--reports-dir DIR reports directory (default $AGENTPROV_HOME/migration/reports); predecessor
reports are looked up here and in the repository's migration/reports
--config FILE migration config (default migration/local.toml when it exists)
--in-process debugging only: map in this process (refused with --release)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov migrate gate-a¶
usage: agentprov migrate gate-a [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N]
[--json] [--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--scope {golden,full}] [--workers N]
[--release] [--report PATH] [--reports-dir DIR] [--config FILE]
[--in-process]
gate A: v1 -> AER -> v1 view, A1 (JCS) and A2 (bytes), from the AER alone
options:
-h, --help show this help message and exit
--scope {golden,full}
golden: the golden set (CI); full: the baseline trees (local). Default:
golden (G0: full)
--workers N worker processes for mapping (0 = host default; default 8 or [run] workers
of the config)
--release release mode: params.strict true, skipped checks fail, T1 isolation grade
must be watchdog or better
--report PATH write the gate report to PATH (atomic); default <reports-
dir>/<gate>/<date>-<commit>.json
--reports-dir DIR reports directory (default $AGENTPROV_HOME/migration/reports); predecessor
reports are looked up here and in the repository's migration/reports
--config FILE migration config (default migration/local.toml when it exists)
--in-process debugging only: map in this process (refused with --release)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov migrate gate-b¶
usage: agentprov migrate gate-b [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N]
[--json] [--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--scope {golden,full}] [--workers N]
[--release] [--report PATH] [--reports-dir DIR] [--config FILE]
[--in-process]
gate B: adprov-v2.3-compat graph bytes (B1 from the view, B2 from the source)
options:
-h, --help show this help message and exit
--scope {golden,full}
golden: the golden set (CI); full: the baseline trees (local). Default:
golden (G0: full)
--workers N worker processes for mapping (0 = host default; default 8 or [run] workers
of the config)
--release release mode: params.strict true, skipped checks fail, T1 isolation grade
must be watchdog or better
--report PATH write the gate report to PATH (atomic); default <reports-
dir>/<gate>/<date>-<commit>.json
--reports-dir DIR reports directory (default $AGENTPROV_HOME/migration/reports); predecessor
reports are looked up here and in the repository's migration/reports
--config FILE migration config (default migration/local.toml when it exists)
--in-process debugging only: map in this process (refused with --release)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov migrate gate-d¶
usage: agentprov migrate gate-d [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N]
[--json] [--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--scope {golden,full}] [--workers N]
[--release] [--report PATH] [--reports-dir DIR] [--config FILE]
[--in-process] [--corpus DIR] [--limit N] [--accept-gap TREE]
gate D: SEM and agentprov-prov/0.1 re-derived from stored AERs and validated in every D8 tier
(release mode)
options:
-h, --help show this help message and exit
--scope {golden,full}
golden: the golden set (CI); full: the baseline trees (local). Default:
golden (G0: full)
--workers N worker processes for mapping (0 = host default; default 8 or [run] workers
of the config)
--release release mode: params.strict true, skipped checks fail, T1 isolation grade
must be watchdog or better
--report PATH write the gate report to PATH (atomic); default <reports-
dir>/<gate>/<date>-<commit>.json
--reports-dir DIR reports directory (default $AGENTPROV_HOME/migration/reports); predecessor
reports are looked up here and in the repository's migration/reports
--config FILE migration config (default migration/local.toml when it exists)
--in-process debugging only: map in this process (refused with --release)
--corpus DIR gate D, full scope: a corpus of stored AERs (read-only); once per tree
--limit N gate D: check N evenly spaced executions per corpus (scope sample)
--accept-gap TREE gate D, full scope: record that this tree has no stored AERs (the report
says so)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov migrate gate-e¶
usage: agentprov migrate gate-e [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N]
[--json] [--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--scope {golden,full}] [--workers N]
[--release] [--report PATH] [--reports-dir DIR] [--config FILE]
[--in-process] [--semdiff-dir DIR]
gate E: the semantic diff of the compat graphs against agentprov-prov/0.1, from semdiff-map.toml
options:
-h, --help show this help message and exit
--scope {golden,full}
golden: the golden set (CI); full: the baseline trees (local). Default:
golden (G0: full)
--workers N worker processes for mapping (0 = host default; default 8 or [run] workers
of the config)
--release release mode: params.strict true, skipped checks fail, T1 isolation grade
must be watchdog or better
--report PATH write the gate report to PATH (atomic); default <reports-
dir>/<gate>/<date>-<commit>.json
--reports-dir DIR reports directory (default $AGENTPROV_HOME/migration/reports); predecessor
reports are looked up here and in the repository's migration/reports
--config FILE migration config (default migration/local.toml when it exists)
--in-process debugging only: map in this process (refused with --release)
--semdiff-dir DIR gate E: where semdiff-<tree>.{json,md} go (default
$AGENTPROV_HOME/migration/semdiff)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov migrate gate¶
usage: agentprov migrate gate [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--scope {golden,full}] [--workers N]
[--release] [--report PATH] [--reports-dir DIR] [--config FILE]
[--in-process] [--corpus DIR] [--limit N] [--accept-gap TREE]
[--semdiff-dir DIR]
GATE
run one gate by name: G0, A, B, D or E
positional arguments:
GATE G0, A, B, D or E
options:
-h, --help show this help message and exit
--scope {golden,full}
golden: the golden set (CI); full: the baseline trees (local). Default:
golden (G0: full)
--workers N worker processes for mapping (0 = host default; default 8 or [run] workers
of the config)
--release release mode: params.strict true, skipped checks fail, T1 isolation grade
must be watchdog or better
--report PATH write the gate report to PATH (atomic); default <reports-
dir>/<gate>/<date>-<commit>.json
--reports-dir DIR reports directory (default $AGENTPROV_HOME/migration/reports); predecessor
reports are looked up here and in the repository's migration/reports
--config FILE migration config (default migration/local.toml when it exists)
--in-process debugging only: map in this process (refused with --release)
--corpus DIR gate D, full scope: a corpus of stored AERs (read-only); once per tree
--limit N gate D: check N evenly spaced executions per corpus (scope sample)
--accept-gap TREE gate D, full scope: record that this tree has no stored AERs (the report
says so)
--semdiff-dir DIR gate E: where semdiff-<tree>.{json,md} go (default
$AGENTPROV_HOME/migration/semdiff)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov render¶
usage: agentprov render [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour] [--break-stale-lock]
--corpus DIR [--ex ID [ID ...]] [--select KEY=VALUE[,VALUE...]]
[--limit N] [--bundle {core,all}] [--format LIST] [--direction {LR,TB}]
[--out DIR] [--max-edges N] [--jobs N] [--dry-run]
Render the agentprov-prov/0.1 PROV graph of selected executions to DOT and images (N.2.13).
Nothing is rendered unless asked; renders go to <corpus>-rendered/ (or --out), never into a
corpus. Needs the render extra (uv sync --frozen --group docs --extra render) and the Graphviz dot
binary, except with --format dot. --show-content adds attribute values (escaped, one line,
--content-chars long) to the node labels; without it labels hold only ids, types and rule ids.
options:
-h, --help show this help message and exit
--corpus DIR the corpus to read (sealed, working, or DIR.partial)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
selection (--ex ID..., --select with --limit N, or --limit N; there is no render-all):
--ex ID [ID ...] execution ids to render
--select KEY=VALUE[,VALUE...]
AER-header filter, as for import --select (repeatable)
--limit N render at most the first N selected executions (required with --select)
drawing and output:
--bundle {core,all} core (default): the core bundle only; all: also the inference and
assessment bundles
--format LIST comma list from svg,png,pdf,dot (default svg); the .dot text is always
written
--direction {LR,TB} rank direction (default LR)
--out DIR output directory (default <corpus>-rendered); never inside a corpus
--max-edges N image layout limit: a graph with more edges gets its .dot text but no
image (default 800; 0 = no limit, dot still times out at 120 s)
--jobs N parallel dot processes (0 = the host default worker count; default 0)
--dry-run select and build the DOT text in memory; write and run nothing
Notation: entities are ovals, activities boxes, agents houses (W3C PROV colours). Border and line
style show the evidence class: solid thick = Observed, dashed = Derived, dotted = Inferred, thin
grey = no class. With --bundle all, purple marks the inference bundles and brown the assessment
bundles; untrusted integrity is salmon, unknown integrity or role is grey, sinks are dark red and
endorsers light blue. Examples: agentprov render --corpus C --ex ex_... | agentprov render
--corpus C --select attack=important_instructions --limit 20 --format svg,png | agentprov render
--corpus C --ex ex_... --bundle all --show-content
agentprov inspect¶
usage: agentprov inspect [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour] [--break-stale-lock]
[--corpus DIR]
[--view {summary,records,transcript,transcript-v1,timeline,evidence,sem,prov,raw,detect}]
[--record ID] [--bundle {core,inference,assessment,toplevel,all}]
[--max-chars N] [--format {text,json}]
TARGET
Look at an execution id (with --corpus), an .aer.jsonl / .sem.jsonl / PROV-JSON file, a raw file
(--view detect) or a corpus root (N.2.8). Read-only. Without --show-content only structure is
printed: ids, kinds, locators, digests, lengths, rule ids and labels; with it, content is printed
through quote_untrusted (escaped, fenced, truncated at --max-chars).
positional arguments:
TARGET an execution id, a file, or a corpus root
options:
-h, --help show this help message and exit
--corpus DIR the corpus (sealed, working or DIR.partial) for an execution id or the raw
store
--view {summary,records,transcript,transcript-v1,timeline,evidence,sem,prov,raw,detect}
what to show (default summary)
--record ID one record (records, evidence, timeline, sem, raw)
--bundle {core,inference,assessment,toplevel,all}
prov view: which bundles (default core)
--max-chars N with --show-content: characters per value (default 500)
--format {text,json} payload format (default text)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov purge¶
usage: agentprov purge [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour] [--break-stale-lock]
--reason {secret,personal_data} (--corpus DIR | --cache) [--cache-dir DIR]
[--yes] [--dry-run]
sha256:HEX
Delete one object (by sha256) from a working corpus's raw store, or one downloaded file from the
fetch cache, and write a tombstone beside the store's records (N.2.15; G.9.6). Executions that
cite the object get the register code release.withheld. A sealed corpus is never modified (exit
4). Purging is irreversible: it asks for confirmation on a terminal, and refuses (exit 4) without
--yes when not interactive.
positional arguments:
sha256:HEX the object to remove
options:
-h, --help show this help message and exit
--reason {secret,personal_data}
why the object is removed
--corpus DIR the working corpus (DIR.partial) whose raw store holds the object
--cache remove a downloaded file from the fetch cache instead
--cache-dir DIR with --cache: the fetch cache (default $AGENTPROV_UPSTREAM_CACHE, else
$AGENTPROV_HOME/cache)
--yes confirm the irreversible removal without asking (needed when not
interactive)
--dry-run print what would be removed; write nothing, take no lock
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov stats¶
usage: agentprov stats [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour] [--break-stale-lock]
--corpus DIR [--tables LIST] [--artefacts] [--per-execution]
[--prov-spec {agentprov-prov/0.1,adprov-v2.3-compat}]
[--select KEY=VALUE[,VALUE...]] [--format {text,csv,json,markdown}]
[--out PATH]
Compute the generic tables of N.2.12 over each corpus separately (never pooled): counts per
source, status and unit; PROV statements by evidence class, bundle and rule; unknown-reason
proportions per field-group pointer; and the per-source information-loss table (field-group
coverage, R-57). Reads the corpus read-only; writes only to stdout or --out (outside every
corpus). Only aggregates are printed.
options:
-h, --help show this help message and exit
--corpus DIR a corpus to read (sealed, working, or DIR.partial); repeatable, one table
set per corpus
--tables LIST comma list from counts,statements,unknowns,field_groups (default: all
four)
--artefacts also scan every selected execution with the shared artefact layer (D.15.4
table)
--per-execution with the statements table, also give it at execution grain
(statements_by_execution; O.3.5)
--prov-spec {agentprov-prov/0.1,adprov-v2.3-compat}
the provenance spec whose PROV files the statements table counts (default
agentprov-prov/0.1)
--select KEY=VALUE[,VALUE...]
index-row filter (repeatable; keys: source_id, profile, unit, status,
mode, completeness, execution_observed, adapter, snapshot, issue,
facet.<name>, cell.<member> (suite, task_id, attack_id, attack_target_id,
defense_id, ...), outcome.<claim>=true|false|none)
--format {text,csv,json,markdown}
payload format (default text)
--out PATH write the payload to PATH (atomic; never inside a corpus)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
Examples: agentprov stats --corpus C | agentprov stats --corpus C1 --corpus C2 --tables
field_groups --format csv --out fg.csv | agentprov stats --corpus C --select unit=agent_execution
--format json | agentprov stats --corpus C --artefacts --tables counts
agentprov package¶
usage: agentprov package [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour] [--break-stale-lock]
--corpus DIR [--corpus-version X.Y.Z] [--spec PATH]
[--licence-file NAME=PATH] [--release-date YYYY-MM-DD] [--deposit OUT]
[--index OUT] [--part RECEIPT] [--metadata release.toml]
[--max-archive-bytes N] [--allow-mixed-licence-classes] [--dry-run]
Seal a working corpus (N.2.11): write rules/, schemas/, LICENSES/, DATASHEET.md, CHANGELOG.md, the
full corpus.json and checksums.sha256, then swap DIR.partial to DIR and print the root digest.
With --deposit OUT, build a deposit draft (deterministic archives, DEPOSIT-MANIFEST.json, deposit-
metadata.json) from one or more sealed corpora. With --index OUT, build the index-only umbrella
deposit of a collection (COLLECTION-INDEX.json, datasheet copies, no archives) from the part
records' draft receipts and their sealed corpora. Nothing is uploaded or published.
options:
-h, --help show this help message and exit
--corpus DIR the corpus directory (seal mode: once, its working tree is DIR.partial;
deposit and index modes: a sealed corpus, repeatable)
--corpus-version X.Y.Z
seal mode (required there): the version to seal (the spec's)
--spec PATH corpora/<id>/corpus-spec.toml (CARD.md and CHANGES.toml beside it)
--licence-file NAME=PATH
an upstream licence file the spec names that the corpus does not store
(repeatable)
--release-date YYYY-MM-DD
release.created (else SOURCE_DATE_EPOCH, else null)
--deposit OUT build a deposit draft directory OUT from sealed corpora (H.10)
--index OUT build the index-only umbrella deposit OUT of a collection (H.10.7; D-947):
no archives, nothing uploaded
--part RECEIPT index mode: a part record's draft receipt (DIR.zenodo-TARGET-draft-
receipt.json, status complete), its deposit directory DIR beside it;
repeatable
--metadata release.toml
deposit and index modes: title, creators, keywords, preamble, DOIs
(H.10.2)
--max-archive-bytes N
deposit mode: split archives on shard boundaries at N bytes (default and
maximum 2^31-1, no ZIP64)
--allow-mixed-licence-classes
deposit mode: admit corpora of different licence classes in one deposit
(recorded; never -nc)
--dry-run compute and check everything, write nothing
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov derive¶
usage: agentprov derive [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour] [--break-stale-lock]
[--corpus SRC] [--out NEW] [--prov-spec SPEC] [--jobs N] [--in-process]
[--link-raw] [--dry-run] [--no-reuse] [--skip-prov-conformance]
[--no-validation-cache] [--stages S[,S...]] [--views VIEW]
[--semantics-profile ID@N] [--trig DIR]
Add missing derived layers of SRC into a new working corpus NEW.partial (N.2.3); SRC is never
written.
options:
-h, --help show this help message and exit
--corpus SRC the source corpus (sealed, working or .partial); read only
--out NEW the new corpus directory (written as NEW.partial; must not exist)
--prov-spec SPEC provenance spec (agentprov-prov/0.1; adprov-v2.3-compat writes the
regression oracle, R-55)
--jobs N worker processes (0 = host default)
--in-process debugging only: run the T1 jobs in this process
--link-raw hard-link raw objects into NEW instead of copying them (D-32)
--dry-run derive and compare in a private temporary tree; write nothing
--no-reuse recompute every execution (--out reuses unchanged outputs by default;
--dry-run never reuses)
--skip-prov-conformance
dev mode only, not advisable: a recomputed validation skips the W3C PROV
conformance check (tier 4); NEW is marked unvalidated and cannot be sealed
(G2c-5)
--no-validation-cache
always run the tier-4 PROV-O round trip; neither read nor write
$AGENTPROV_HOME/cache/tier4 (D-550)
--stages S[,S...] layers to add: sem,prov,validation,views
--views VIEW add a view (transcript-v1)
--semantics-profile ID@N
semantics profile for an added sem layer (repeatable; default: the
source's)
--trig DIR write PROV-O TriG (on demand; never part of a corpus) into DIR
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov recompute¶
usage: agentprov recompute [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--corpus SRC] [--out NEW] [--prov-spec SPEC]
[--jobs N] [--in-process] [--link-raw] [--dry-run] [--no-reuse]
[--skip-prov-conformance] [--no-validation-cache] [--stages S[,S...]]
[--show-diff N] [--strict]
Re-run the derivation stages of SRC into NEW.partial (or in memory with --dry-run) and compare.
options:
-h, --help show this help message and exit
--corpus SRC the source corpus (sealed, working or .partial); read only
--out NEW the new corpus directory (written as NEW.partial; must not exist)
--prov-spec SPEC provenance spec (agentprov-prov/0.1; adprov-v2.3-compat writes the
regression oracle, R-55)
--jobs N worker processes (0 = host default)
--in-process debugging only: run the T1 jobs in this process
--link-raw hard-link raw objects into NEW instead of copying them (D-32)
--dry-run derive and compare in a private temporary tree; write nothing
--no-reuse recompute every execution (--out reuses unchanged outputs by default;
--dry-run never reuses)
--skip-prov-conformance
dev mode only, not advisable: a recomputed validation skips the W3C PROV
conformance check (tier 4); NEW is marked unvalidated and cannot be sealed
(G2c-5)
--no-validation-cache
always run the tier-4 PROV-O round trip; neither read nor write
$AGENTPROV_HOME/cache/tier4 (D-550)
--stages S[,S...] stages to re-run: expand,aer,sem,prov,validation,views (default: every
stage the source has)
--show-diff N show the first differences of N changed executions
--strict exit 1 when an execution is unverified
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov verify¶
usage: agentprov verify [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour] [--break-stale-lock]
--corpus PATH [--level {checksums,chain,rederive}]
[--root-digest sha256:HEX] [--against-pins] [--allow-network]
[--cache-dir DIR] [--sample N] [--seed S] [--release] [--jobs N]
[--in-process] [--no-validation-cache]
Verify corpus integrity (N.2.6; H.5.5) read-only: a corpus tree, a deposit directory or a lone
meta archive (archive mode, never extracted). Levels: checksums (closed world), chain (the
default: manifest, digests, index, register and counts, AER seals, raw objects and member CRC-32s,
SEM/PROV inputs, facets, catalogue authority), rederive (chain plus a byte-identical re-derivation
from the stored raw objects).
options:
-h, --help show this help message and exit
--corpus PATH a corpus directory, a deposit directory or a .meta.zip
--level {checksums,chain,rederive}
verification level (default chain)
--root-digest sha256:HEX
a root digest published outside the corpus (D-39; repeatable; with
--release required, and for a deposit one per corpus; D-620, D-641)
--against-pins compare every git origin with the in-package profile pins
--allow-network --against-pins: fetch a pinned tree listing not in the cache
--cache-dir DIR --against-pins: the fetch cache (default $AGENTPROV_HOME/cache)
--sample N check a seeded sample of N executions (N.1.7)
--seed S the sampling seed (with --sample)
--release release mode: unverified fails; no sampling; needs --root-digest
--jobs N rederive: worker processes (0 = host default)
--in-process rederive: run the T1 jobs in this process (debugging only)
--no-validation-cache
rederive: always run the tier-4 PROV-O round trip; neither read nor write
the round-trip cache (D-550). With --release the cache is never read; it
is audited
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov settings¶
usage: agentprov settings [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock]
ACTION ...
``agentprov settings``: the user settings the CLI and the GUI share (D-703; reverse CLI parity
D-700).
positional arguments:
ACTION
list the stored settings (or the schema)
get one setting (nothing when unset)
set store one setting (checked against the schema)
unset remove one setting
options:
-h, --help show this help message and exit
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov settings list¶
usage: agentprov settings list [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--prefix P] [--schema] [--format {text,json}]
options:
-h, --help show this help message and exit
--prefix P only keys starting with P
--schema list the key patterns, types and bounds instead
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov settings get¶
usage: agentprov settings get [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--format {text,json}]
KEY
positional arguments:
KEY
options:
-h, --help show this help message and exit
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov settings set¶
usage: agentprov settings set [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock]
KEY VALUE
positional arguments:
KEY
VALUE lists: comma or space separated; booleans: true or false
options:
-h, --help show this help message and exit
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov settings unset¶
usage: agentprov settings unset [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N]
[--json] [--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock]
KEY
positional arguments:
KEY
options:
-h, --help show this help message and exit
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov keys¶
usage: agentprov keys [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N] [--control-file PATH]
[--show-content] [--content-chars N] [--json] [--summary-json PATH]
[--max-listed N] [--no-colour] [--break-stale-lock]
ACTION ...
``agentprov keys``: endpoint keys in the OS keychain (D-702; D-395, D-655; reverse CLI parity
D-700).
positional arguments:
ACTION
status whether a secure keychain is available
list the endpoints' key variables and whether each is stored
set store a key (prompted without echo, or one line on stdin)
delete remove a stored key
options:
-h, --help show this help message and exit
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov keys status¶
usage: agentprov keys status [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--format {text,json}]
options:
-h, --help show this help message and exit
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov keys list¶
usage: agentprov keys list [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--format {text,json}]
options:
-h, --help show this help message and exit
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov keys set¶
usage: agentprov keys set [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--stdin]
NAME
positional arguments:
NAME a key variable of endpoints.toml, e.g. OPENROUTER_API_KEY, or
ZENODO_SANDBOX_TOKEN / ZENODO_TOKEN
options:
-h, --help show this help message and exit
--stdin read the value as one line from stdin (never from argv)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov keys delete¶
usage: agentprov keys delete [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--yes]
NAME
positional arguments:
NAME
options:
-h, --help show this help message and exit
--yes confirm without asking (needed when not interactive)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov prices¶
usage: agentprov prices [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour] [--break-stale-lock]
ACTION ...
``agentprov prices``: the model prices the estimate reads (D-701; D-650; reverse CLI parity
D-700).
positional arguments:
ACTION
list the price sources and every priced model per endpoint
get the price one estimate would use (nothing when unpriced)
set store a per-model override (USD per million tokens)
unset remove an override
refresh fetch OpenRouter's model list into a dated snapshot (network; confirm
first)
options:
-h, --help show this help message and exit
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov prices list¶
usage: agentprov prices list [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--endpoint NAME] [--format {text,json}]
options:
-h, --help show this help message and exit
--endpoint NAME only this endpoint of endpoints.toml
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov prices get¶
usage: agentprov prices get [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--format {text,json}]
ENDPOINT MODEL
positional arguments:
ENDPOINT
MODEL
options:
-h, --help show this help message and exit
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov prices set¶
usage: agentprov prices set [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] --input USD --output USD [--cached-input USD]
[--cache-write USD] [--as-of YYYY-MM-DD]
ENDPOINT MODEL
positional arguments:
ENDPOINT
MODEL
options:
-h, --help show this help message and exit
--input USD input tokens (uncached)
--output USD output tokens (reasoning included)
--cached-input USD cache reads (default: the input price)
--cache-write USD 5-minute cache writes (default: the input price)
--as-of YYYY-MM-DD the date of the price (default: today)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov prices unset¶
usage: agentprov prices unset [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock]
ENDPOINT MODEL
positional arguments:
ENDPOINT
MODEL
options:
-h, --help show this help message and exit
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov prices refresh¶
usage: agentprov prices refresh [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N]
[--json] [--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--allow-network | --from-file RESPONSE.json]
[--fetched YYYY-MM-DD]
options:
-h, --help show this help message and exit
--allow-network required to fetch (one public URL, no key)
--from-file RESPONSE.json
import a saved /api/v1/models response instead
--fetched YYYY-MM-DD with --from-file: the date it was fetched (default: today)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov estimate¶
usage: agentprov estimate [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--corpus DIR] [--campaign NAME | --corpus-id ID]
[--plan PLAN.toml] [--spec SPEC.toml] [--suite S[,S...]]
[--benchmark-version V] [--user-tasks SPEC] [--attack A[,A...]]
[--injection-tasks SPEC] [--benign {include,exclude,only}]
[--defence NAME] [--task NAME] [--task-param KEY=VALUE]
[--samples limit:N|ID[,ID...]] [--epochs N] [--message-limit N]
[--token-limit N] [--time-limit SECONDS] [--max-connections N]
[--cell-size N] [--judge ROLE=MODEL] [--model SHORTHAND]
[--endpoint NAME] [--model-id ID] [--defence-endpoint NAME]
[--llm-script SCRIPT.json] [--llm-element {native,gateway_compat}]
[--model-display-name NAME] [--temperature T] [--reasoning-effort LEVEL]
[--max-tokens N|harness-default]
[--temperature-policy {as-benchmark,inject}] [--allow-text-tool-calls]
[--base-url URL] [--accept-served-model ID] [--allow-lan-endpoint]
[--no-hash-weights] [--repeats N] [--max-attempts N]
[--retry-inconsistent] [--allow-mixed-config] [--limit N] [--canary N]
[--cell-timeout SECONDS] [--jobs N] [--start-stagger SECONDS]
[--env-file PATH] [--env-stdin] [--keychain] [--allow-custom-endpoint]
[--containment {sandbox-exec,dedicated-user,vm}] [--autodojo-cache PATH]
[--autodojo-variant K] [--allow-network] [--max-calls N]
[--max-cost-usd USD] [--price-in USD_PER_MTOK]
[--price-out USD_PER_MTOK] [--stats-corpus DIR] [--format {text,json}]
SOURCE
``agentprov estimate``: the cost and time of a live run before it starts (D-701; design §8.3 as
amended by D-650,
positional arguments:
SOURCE live source: agentdojo, agentdyn, agentharm or autodojo
options:
-h, --help show this help message and exit
--corpus DIR working corpus directory
--campaign NAME corpus id <source>-live-<NAME>
--corpus-id ID
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
cells (exclusive with --plan):
--plan PLAN.toml committed run plan
--spec SPEC.toml corpus spec with [live]
--suite S[,S...]
--benchmark-version V
--user-tasks SPEC
--attack A[,A...]
--injection-tasks SPEC
--benign {include,exclude,only}
--defence, --defense NAME
Inspect plans (an inspect-<benchmark> backend; D-663; exclusive with the cell options):
--task NAME Inspect task (the catalogue lists them), e.g. inspect_evals/agentharm
--task-param KEY=VALUE
a task parameter the catalogue declares (repeatable; VALUE as a TOML
scalar or a string)
--samples limit:N|ID[,ID...]
the samples: the first N of the split, or ids
--epochs N
--message-limit N per sample (mandatory for Inspect, D-667)
--token-limit N per sample (mandatory for Inspect, D-667)
--time-limit SECONDS per sample (mandatory for Inspect, D-667)
--max-connections N concurrent model calls of one cell
--cell-size N samples per cell (one Inspect eval per cell)
--judge ROLE=MODEL the model of a judge role of the task's scorer (repeatable; every role is
required, D-667); judges use the agent's endpoint and key
model (one of: --model, --endpoint with --model-id, --llm-script):
--model SHORTHAND endpoint shorthand from endpoints.toml
--endpoint NAME explicit endpoint (deepseek, openai, openrouter, anthropic, local, ollama,
omlx, lmstudio, mlx_lm, llamacpp, vllm)
--model-id ID
--defence-endpoint NAME
the endpoint (endpoints.toml shorthand) a defence calls for its own model
or embedding calls (D-675; MELON's embeddings); only its key enters the
worker
--llm-script SCRIPT.json
scripted transport: no network, no keys (test_only)
--llm-element {native,gateway_compat}
--model-display-name NAME
--temperature T
--reasoning-effort LEVEL
--max-tokens N|harness-default
Anthropic endpoints (required there): the reply's token cap, stated (N) or
the harness default 16000 (harness-default, recorded as a harness policy);
AgentDojo's own 1024 is never used silently (D-656)
--temperature-policy {as-benchmark,inject}
--allow-text-tool-calls
run an endpoint that returns text-encoded tool calls
local model servers (D-691..D-694):
--base-url URL a local endpoint's base URL (loopback; overrides .env and the registry
default; recorded)
--accept-served-model ID
a served model id that may differ from the requested one (repeatable;
recorded; A12)
--allow-lan-endpoint allow a local endpoint on a literal private LAN address (D-698): prompts
and replies leave this machine over plain http; keyless; recorded
--no-hash-weights record a local model's weight files by size only (weights_hashed false)
control:
--repeats N
--max-attempts N
--retry-inconsistent
--allow-mixed-config
--limit N
--canary N canary cells (default 10; 0 disables)
--cell-timeout SECONDS
per-cell wall clock before the worker is killed and the session sealed
aborted (default 1800)
--jobs N concurrent traces (>= 1; default 1, or computed with a local preset)
--start-stagger SECONDS
spread the first cell start of each worker
environment and consent:
--env-file PATH
--env-stdin read the endpoint variables as one JSON object from stdin (D-395; for a
GUI with a keychain); exclusive with --env-file, and the default .env is
not read
--keychain read the selected endpoint's key from the OS keychain (agentprov keys;
D-702); exclusive with --env-file and --env-stdin, and the default .env is
not read
--allow-custom-endpoint
--containment {sandbox-exec,dedicated-user,vm}
--autodojo-cache PATH
AutoDojo attack cache copied from the pinned fork (verified against its
tree; D.11.3 rule 5)
--autodojo-variant K the cache's variant index (integer >= 0)
--allow-network let T0 fetch the pinned AutoDojo commit into the fetch cache to verify
--autodojo-cache
API caps (D-725: required for an API endpoint, optional otherwise):
--max-calls N billed model calls of the corpus at most, preflight probes included; a
cell starts only if its worst case fits
--max-cost-usd USD spend of the corpus at most (an upper bound priced from the recorded
usage); needs --price-in and --price-out
--price-in USD_PER_MTOK
USD per million input-side tokens (the highest of input, cache write and
cached)
--price-out USD_PER_MTOK
USD per million output tokens
estimate:
--stats-corpus DIR measure calls, tokens and time in this stored corpus (a sample of its AER
files, counts only; repeatable, at most 16)
--format {text,json} json: one agentprov-estimate/0.1 document on stdout
agentprov jobs¶
usage: agentprov jobs [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N] [--control-file PATH]
[--show-content] [--content-chars N] [--json] [--summary-json PATH]
[--max-listed N] [--no-colour] [--break-stale-lock]
ACTION ...
``agentprov jobs``: detached jobs that outlive the command that started them (D-704; GUI design
§4; reverse CLI
positional arguments:
ACTION
start start ARGV (after --) as a detached job
list the jobs, newest first
show one job, with the tails of its output
follow print a job's progress (or log) lines until it ends
stop a graceful stop (SIGTERM to the job's CLI)
drain finish the units in flight, start no new ones (needs --with-control-file)
resume start the job's request again (the CLI's resume rules decide what is
redone)
kill force kill the job's whole process tree
prune delete old ended jobs (lists them without --yes)
options:
-h, --help show this help message and exit
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov jobs start¶
usage: agentprov jobs start [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--label TEXT] [--writes DIR] [--reads DIR]
[--slot {network,backends}] [--with-control-file] [--queue-cpu N]
[--progress-level {status,steps}] [--tool tools/NAME.py] [--note TEXT]
[--meta KEY=VALUE] [--format {text,json}]
...
positional arguments:
-- ARGV the agentprov arguments of the job
options:
-h, --help show this help message and exit
--label TEXT a label shown in lists
--writes DIR a corpus the job writes (one writer per corpus; repeatable)
--reads DIR a corpus the job reads (repeatable)
--slot {network,backends}
a global slot (one job at a time)
--with-control-file give the job a control file (drain, D-394)
--queue-cpu N run through the machine-wide job queue reserving N CPUs
--progress-level {status,steps}
--tool tools/NAME.py run one repository tool of this checkout instead of agentprov (ARGV are
its arguments)
--note TEXT a note recorded with the job
--meta KEY=VALUE a fact recorded with the job (KEY [a-z_][a-z0-9_]*; VALUE a JSON scalar or
text; repeatable)
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov jobs list¶
usage: agentprov jobs list [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--state {running,ended,all}]
[--format {text,json}]
options:
-h, --help show this help message and exit
--state {running,ended,all}
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov jobs show¶
usage: agentprov jobs show [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--format {text,json}]
ID
positional arguments:
ID
options:
-h, --help show this help message and exit
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov jobs follow¶
usage: agentprov jobs follow [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--timeout S] [--log]
ID
positional arguments:
ID
options:
-h, --help show this help message and exit
--timeout S stop following after S seconds (0: never)
--log the job's log records (log.jsonl) instead of its progress
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov jobs stop¶
usage: agentprov jobs stop [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock]
ID
positional arguments:
ID
options:
-h, --help show this help message and exit
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov jobs drain¶
usage: agentprov jobs drain [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock]
ID
positional arguments:
ID
options:
-h, --help show this help message and exit
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov jobs resume¶
usage: agentprov jobs resume [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--format {text,json}]
ID
positional arguments:
ID
options:
-h, --help show this help message and exit
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov jobs kill¶
usage: agentprov jobs kill [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--yes]
ID
positional arguments:
ID
options:
-h, --help show this help message and exit
--yes confirm (a forced kill can leave a stale lock)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov jobs prune¶
usage: agentprov jobs prune [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--keep N] [--max-age-days D] [--yes]
[--format {text,json}]
options:
-h, --help show this help message and exit
--keep N keep the newest N ended jobs (default: the setting jobs.retention, else
200)
--max-age-days D and none older than D days (default: the setting jobs.retention_days, else
90)
--yes delete them
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov batch¶
usage: agentprov batch [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour] [--break-stale-lock]
ACTION ...
``agentprov batch``: ordered job lists that run as one job (D-704; GUI design §4.4; user decision
2026-10-22).
positional arguments:
ACTION
create create a batch from a step list (not started)
start start the batch's driver job
resume start a new driver for a batch whose driver ended
run advance the batch in the foreground until it is done (the driver)
stop stop the driver (the steps it started keep running)
list the batches, newest first
show one batch and its steps
options:
-h, --help show this help message and exit
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov batch create¶
usage: agentprov batch create [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--stdin] [--label TEXT] [--concurrency N]
[--continue-on-failure | --no-continue-on-failure]
[--format {text,json}]
[SPEC.json]
positional arguments:
SPEC.json the step list (a file)
options:
-h, --help show this help message and exit
--stdin read the step list from stdin instead
--label TEXT
--concurrency N steps at a time (1 to 16)
--continue-on-failure, --no-continue-on-failure
go on after a step ends with exit 4 or 70 (default: the setting
jobs.continue_on_failure)
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov batch start¶
usage: agentprov batch start [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--confirm-r2] [--format {text,json}]
ID
positional arguments:
ID
options:
-h, --help show this help message and exit
--confirm-r2 confirm the batch's R2 steps (network, downloads, paid calls) for this
driver
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov batch resume¶
usage: agentprov batch resume [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--confirm-r2] [--format {text,json}]
ID
positional arguments:
ID
options:
-h, --help show this help message and exit
--confirm-r2 confirm the batch's R2 steps (network, downloads, paid calls) for this
driver
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov batch run¶
usage: agentprov batch run [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--confirm-r2] [--interval S]
ID
positional arguments:
ID
options:
-h, --help show this help message and exit
--confirm-r2 confirm the batch's R2 steps (network, downloads, paid calls) for this
driver
--interval S seconds between ticks
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov batch stop¶
usage: agentprov batch stop [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--steps]
ID
positional arguments:
ID
options:
-h, --help show this help message and exit
--steps also stop the running steps (one SIGTERM each)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov batch list¶
usage: agentprov batch list [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--format {text,json}]
options:
-h, --help show this help message and exit
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov batch show¶
usage: agentprov batch show [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--format {text,json}]
ID
positional arguments:
ID
options:
-h, --help show this help message and exit
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov devrun¶
usage: agentprov devrun [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour] [--break-stale-lock]
ACTION ...
``agentprov devrun``: a quick end-to-end check before a full import (D-705; GUI design §19; D-544,
D-529).
positional arguments:
ACTION
plan the parts and steps a start would run
start create a run folder and start its batch
list the runs, newest first
prune delete old run folders (lists them without --yes)
options:
-h, --help show this help message and exit
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov devrun plan¶
usage: agentprov devrun plan [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--format {text,json}]
options:
-h, --help show this help message and exit
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov devrun start¶
usage: agentprov devrun start [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--confirm-r2] [--format {text,json}]
options:
-h, --help show this help message and exit
--confirm-r2 confirm the imports' fetch-cache reads (R2)
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov devrun list¶
usage: agentprov devrun list [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--format {text,json}]
options:
-h, --help show this help message and exit
--format {text,json}
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov devrun prune¶
usage: agentprov devrun prune [-h] [--version] [-v] [-q] [--log-format {text,json}]
[--log-file PATH] [--log-content]
[--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour]
[--break-stale-lock] [--keep N] [--yes]
options:
-h, --help show this help message and exit
--keep N keep the newest N (default devrun.keep)
--yes delete them
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead
agentprov assist¶
usage: agentprov assist [-h] [--version] [-v] [-q] [--log-format {text,json}] [--log-file PATH]
[--log-content] [--progress {off,summary,status,steps,trace}]
[--progress-format {auto,tty,plain,jsonl}] [--progress-file PATH]
[--progress-interval SECONDS] [--progress-max-bytes N]
[--control-file PATH] [--show-content] [--content-chars N] [--json]
[--summary-json PATH] [--max-listed N] [--no-colour] [--break-stale-lock]
[--question TEXT] [--endpoint NAME] [--model-id ID] [--base-url URL]
[--server KIND] [--list-models] [--allow-network] [--keychain]
[--env-file PATH] [--context-chars N] [--show-context]
[--format {text,json}]
The desktop app's Help > Assistant (D-1010). The conversation is read from stdin as JSON
(agentprov-assist-request/0.1) or given with --question; the most relevant passages of the user
guide, the README, this CLI's help and the module docstrings are sent with it to one endpoint of
endpoints.toml. They come from the checkout this CLI runs from (or the one around the working
directory); a CLI installed with pip install agentprov, run outside a checkout, uses the copy of
the user guide its package ships (D-1030). An API provider needs --allow-network (R2); a local
server is reached on loopback only. --show-context calls no model.
options:
-h, --help show this help message and exit
--question TEXT one question (instead of a conversation on stdin)
--endpoint NAME an endpoint of endpoints.toml (default: the setting assistant.endpoint)
--model-id ID the model id (default: the setting assistant.model)
--base-url URL a local OpenAI-compatible server on loopback, for one that is not an
endpoint of endpoints.toml (default: the setting assistant.base_url)
--server KIND the local server's kind: its default URL, and how --list-models finds the
loaded model (default: the setting assistant.server)
--list-models list the models the local server serves (the loaded ones first, where it
says) and stop
--allow-network allow the request to an API provider (R2; the registry host only, the
conversation and the selected passages are sent)
--keychain read the endpoint's key from the OS keychain only (D-702)
--env-file PATH read the endpoint's key variable from this .env file only
--context-chars N characters of documentation sent with the question (default 24,000; 0
sends none)
--show-context print the passages that would be sent, and call no model
--format {text,json} payload format (default text)
global options:
--version show versions and exit
-v, --verbose log DEBUG (repeatable)
-q, --quiet log WARNING and above; implies --progress off
--log-format {text,json}
stderr log format (default text)
--log-file PATH also log to PATH (DEBUG, JSON lines)
--log-content permit escaped content excerpts in DEBUG logs
--progress {off,summary,status,steps,trace}
live feedback level on stderr (default: status on a terminal, else
summary; env AGENTPROV_PROGRESS)
--progress-format {auto,tty,plain,jsonl}
feedback rendering (default auto: tty on an interactive terminal, else
plain)
--progress-file PATH also append feedback events as agentprov-progress/0.2 JSON lines to PATH
(never below status)
--progress-interval SECONDS
heartbeat interval of --progress-file (default 2, minimum 1)
--progress-max-bytes N
rotate --progress-file to PATH.1 past N bytes (default 8 MiB, minimum 64
KiB)
--control-file PATH writers: poll PATH for an agentprov-control/0.1 request (drain or stop;
D-394)
--show-content add escaped content excerpts at steps and trace
--content-chars N excerpt length for --show-content (default 200, max 2000)
--json summary document to stdout
--summary-json PATH write the summary document to PATH (atomic)
--max-listed N cap on failures[] entries in the summary (default 100)
--no-colour, --no-color
no ANSI colour
--break-stale-lock writers only: remove a lock whose pid is dead