CLI Reference

This page is a copy-paste reference. It may lag the parser; striatum --help (and striatum <verb> --help) is authoritative.

Core lifecycle

striatum repo add <path> [--init]
striatum skills install [--profile <profile>] [--scope project|user]
striatum plugin install [--profile <profile>] [--scope project|user]
striatum workflow validate
striatum workflow generate
striatum workflow templates list
striatum workflow templates show
striatum run prepare
striatum branch confirm
striatum run start
striatum run drive
striatum operator bootstrap
striatum run summary
striatum archive create

workflow generate (below) is the way to scaffold a starter workflow tree; run prepare requires an explicit --workflow <path> and creates no runtime default. The generated tree uses a single local process lane as a valid placeholder; edit lanes and job lane_id bindings for real agent runs.

run drive --run-id <id> [--interval 15s] [--provider-auth-gate auto|required|off] [--once] [--json] is a local operator loop over existing daemon RPC methods. It reads run.detail and list.sessions, registers and supervises one fresh session per queued role/lane as the DAG unblocks, adopts already-active matching sessions, and stops terminal or superseded launched lanes before registering fresh reviewers. It is not a daemon RPC method and does not call rescue verbs or force non-fresh sessions. --provider-auth-gate is forwarded to supervise.start; when a lane-provider-auth refusal blocks launch, the driver closes the fresh session, emits a sanitized failure action, and exits nonzero instead of retrying the same doomed launch.

run pause [run-id], run resume [run-id], and run cancel [run-id] transition an existing run's lifecycle (capability admin). run retry-job [run-id] [job-id] requeues a single failed job (capability admin). run integrate [run-id] integrates completed run output into the target branch (capability apply). All five are daemon RPC methods; pass the run id (and job id) positionally or as --run-id / --job-id.

operator bootstrap [--operator-docs-root <path>] [--limit N] [--markdown|--json] is a bounded AI-operator cold-start packet. It is a local read-only composite over existing daemon reads (repo.resolve, status, doctor) plus local git, VERSION, daemon-runtime-path, MCP-endpoint-path, and docs/operator/BRIEF.md probes. It prints frontier state, doctor counts, operator-brief freshness, skill drift, exact next commands, and a bounded reading plan without embedding full status, doctor, session, verdict, or historical run arrays. --limit caps expanded lists and is bounded to 20. The command creates no live state and is not a daemon RPC method.

workflow templates list [--kind shape|lane_set|role_pack|adversary_pack] and workflow templates show <template_id> read the bundled local workflow-template catalog (embedded under go/pkg/workflowtemplates).

workflow generate --shape <shape> [--lane-set <set>] [--workflow-id <id>] [--scaffold-root <path>] [--artifact-root <path>] [--option key=value ...] compiles a concrete workflow tree from that catalog. It previews the planned repo-relative files by default and writes them only with --write; --json emits the structured envelope. The lane set defaults to the local fixture lane (which needs no real lane command), so workflow generate --shape conversation --option topic="…" scaffolds a valid starter out of the box; edit in real lanes (e.g. --lane-set author_reviewer then supply lane commands in the generated workflow.json) before a real run. --option phases=… (multi_phase) and shape-specific --option key=value values are accepted as the shape requires.

The Python-era workflow init, workflow lint, workflow plan, workflow graph, workflow upgrade, and workflow templates render-md verbs are not part of the current Go CLI (RFC 0078 ported validate, generate, and templates {list,show}); workflow-authoring lint is enforced at validate/generation time. A fuller CLI_REFERENCE audit against the Go command surface is tracked separately.

Same-model-pairing lint is enforced by workflow validate (refuse unless --allow-same-model-pairing); operational accepted-risk overrides are recorded through the daemon workflow accept-risk / workflow accepted-risks commands.

striatum repo add <path> [--init] registers a target repository with the daemon-owned PostgreSQL registry. Pass --init for a fresh target repo so the daemon creates .striatum/scratch and adds .striatum/ to .gitignore.

striatum skills install --profile <profile> writes the agent skill bundle for claude_code | codex | agy | generic | all. Use --scope user to install once in the user's agent config directory instead of a project tree.

striatum plugin install --profile <profile> writes agent plugin bundles for claude_code, codex, or agy. Project-scope plugin installs also write a local marketplace fixture when supported by the profile.

Agent / session work loop

striatum register-session
striatum claim-next
striatum ack
striatum heartbeat
striatum release
striatum send
striatum block
striatum publish-artifact
striatum complete
striatum verdict
striatum submit-review
striatum decision record

Use submit-review for the normal single-call review path: publish the finding artifact and record the verdict. Use verdict when the required review artifact is already published for the current attempt, such as a re-claimed review job after lease recovery.

For operator-authored or otherwise unattested recovery of an accepting fresh review, record an accepting run-level decision with --escape-surface review_provenance, then pass --review-provenance-decision-id <decision-id> to submit-review or verdict. The verdict event records the decision id and artifact id as durable override evidence.

publish-artifact validates lease ownership, write scope, path safety, artifact kind, front matter, and byline. Model-bylined artifacts require lane evidence: a path-specific supervised artifact_observed event when the wrapper reports one, or the legacy clean process_executions fallback. Operators can explicitly bypass missing lane evidence with --allow-no-process-execution --override-rationale <text>; the rationale is stored on the artifact row and in the provenance event. For review jobs declaring require_attested_lane: true, publish-artifact also refuses publication unless the session has an attached lane supervisor. D185 makes publish-artifact and complete share the frozen-attempt write-scope contract: if a mid-run context or workflow edit makes the current effective scope disagree with the job attempt's frozen scope, both commands must fail with compatible typed write-scope errors that point to audited recovery instead of silently widening historical scope. decision record --mark-run-compromised records an accepting decision and transitions a completed run to compromised for provenance invalidation; V1 uses this replacement-run-only path for compromised completed review jobs.

Artifact listing, detail, summary, export, and dashboard JSON includes provenance.category so operator-bylined artifacts can be separated from attested supervised-lane, daemon auto-finalized, operator-on-behalf, self-declared operator, recovery-authored, and run-level operator artifacts.

Mediated repository and process surfaces (RFC 0099)

striatum repo write <session-id> <job-id> <lease-id> <path> --content <text>
striatum repo patch-preview <session-id> <job-id> <lease-id> --patch <text>
striatum repo patch-apply <session-id> <job-id> <lease-id> --patch <text>
striatum process run <session-id> <job-id> <lease-id> [-- <command> ...]
striatum scope-check [--packet-file <packet.json>] [--allowed <path> ...] [--forbidden <path> ...]
striatum work claim-override <session-id> <job-id> <decision-id>

These are the RFC 0099 constrained-operator mutation surfaces. repo write, repo patch-preview, and repo patch-apply (capability write) require an active repo-write job lease and refuse paths outside write_scope.allowed_paths before mutating. repo write writes exact UTF-8 --content to one repo-relative path; repo patch-preview checks a unified git-style --patch for applyability and write-scope without mutating, and repo patch-apply repeats those checks before applying.

process run (capability write) executes a command array — passed after -- or via --command-json — with --timeout-seconds (default 300, cap 1800) and records process_executions evidence without storing stdout/stderr. It requires capability_requirements.process_execution: true on the job, or a matching typed escape decision whose escape_surface is process.run or shell_command.

scope-check is a daemon-free, read-only pre-complete diagnostic: it flags any changed path outside allowed_paths or inside forbidden_paths and exits nonzero on drift. Read the scope from the active work packet with --packet-file, or paste --allowed / --forbidden paths directly. It does not widen or refresh the daemon's frozen attempt scope.

work claim-override <session-id> <job-id> <decision-id> (capability admin) is the narrow escape for the fresh-review process-lineage gate: it claims a pending job for a session authorized by an accepted decision recorded with matching --subject-session-id / --subject-job-id. A broad or mismatched decision is refused; there is no normal-lane claim-next --force.

Worktree (opt-in per lane via worktree_isolation: per_job)

striatum worktree create
striatum worktree release
striatum worktree gc [--run-id <id>]
striatum worktree list

worktree release --worktree-id <id> removes reachable worktrees; --force explicitly discards unreachable worktrees and can retire a missing-on-disk row when the owning job is terminal. Missing-on-disk forced releases emit worktree.force_released with missing_on_disk: true.

worktree gc removes terminal job worktrees whose on-disk HEAD is reachable from the run branch or a refs/striatum/ pin; it also retires terminal rows whose path is already missing on disk. Skipped rows are reported with reasons. Worktrees with no-blob published artifacts that are not present in the worktree HEAD are skipped until the artifact content is durable outside the per-job worktree.

Supervisor (RFC 0009)

striatum supervise start <session-id> [--provider-auth-gate auto|required|off]
striatum supervise send
striatum supervise stop
striatum supervise status
striatum supervise list
striatum supervise trajectory

supervise start owns the lane provider-auth launch gate. The default auto mode runs supported provider-auth preflights for Codex and Claude agent-loop lanes under a distinct configured lane OS user; required blocks unsupported providers and auth-negative preflight results; off is the explicit rollback path. The gate runs before scratch/FIFO creation, session-bound lane token minting, supervisor rows/events, helper/tmux, or the provider lane process. Blocking results return a safe lane_provider_auth details block. The block can include the probe name, exit code, stdout/stderr byte counts, and bounded success-signal state; it never includes raw stdout, stderr, final text, auth paths, provider account ids, environment values, token material, PTY logs, or tracebacks. For Codex, zero exit means the lane provider CLI authenticated successfully. For Claude, the preflight is an offline freshness check of the resolver-selected .credentials.json; it does not run a provider command or spend model tokens. even if the bounded --output-last-message signal is missing or mismatched; that condition is diagnostic and does not block launch.

Conversation (RFC 0086)

striatum conversation open <participant-session-ids>
striatum conversation say <session-id> <conversation-id> <body>
striatum conversation close <session-id> <conversation-id>
striatum conversation list <run-id>
striatum conversation show <conversation-id>

The RFC 0086 multi-party conversation surface runs an N-party round-robin exchange over persistent agent-loop sessions. conversation open starts a conversation across the named participant sessions; say posts a turn (both write), close ends it (write), and list / show (read) enumerate a run's conversations and replay one transcript.

Interrogation (RFC 0082 / 0083)

striatum interrogation open <session-id> <target-session-id>
striatum interrogation ask <session-id> <interrogation-id>
striatum interrogation answer <session-id> <interrogation-id>
striatum interrogation close <session-id> <interrogation-id>
striatum interrogation list <run-id>
striatum interrogation show <interrogation-id>

Interrogation windows let a consumer session question an upstream interrogation target (a completing job marked interrogable: true whose session stays available for preserved-context questioning). open / ask / answer / close (capability write) drive one window; list / show (read) enumerate a run's interrogations and replay one exchange.

Trajectory (RFC 0081)

striatum trajectory export <run-id> [profile]
striatum trajectory watch <run-id> [profile]

trajectory export returns the ordered, profile-filtered conversation trajectory for a run as replayable/diffable JSONL; trajectory watch live-tails it (both capability read). The dialogue profile is the curated adjudicator view (striatum trajectory export <run-id> dialogue). This is distinct from supervise trajectory, which reads one supervisor's operator-local PTY log.

Escalation and principal inbox (RFC 0053 / 0102)

striatum inbox
striatum escalation list
striatum escalation show <escalation-id>
striatum escalation resolve <escalation-id>

escalation list (capability read) reports the open escalations awaiting human-principal judgment; striatum inbox is the alias that routes to the same escalation.list method. escalation show (read) prints one escalation, and escalation resolve (admin) records the principal's disposition. Escalations are also resolvable through striatum decision record (RFC 0053 escalation artifacts are durable provenance for the principal inbox).

Dashboard

striatum dashboard
striatum dashboard --all

dashboard --all (RFC 0028 V1) groups registered repositories and reports daemon/Postgres-backed per-repository runs, blockers, claimable jobs, stale leases, and degraded repositories. It uses daemon-owned repository registration state and requires a daemon read capability token from the runtime client-token file.

run summary --json, dashboard --once, and status --json all emit the RFC 0157 (D251) canonical state_projection block — {run_state, jobs:[{id,state}]} — so a script reads run/job state uniformly across the three verbs (jobs[].id is the stable workflow_job_id). The block is strictly additive: existing keys (run.summary's .run/.jobs[], dashboard's .jobs_by_state, status's .runs[]/.jobs{}) are unchanged, and dashboard --once additionally carries a top-level state mirroring state_projection.run_state. For a repo-wide call (no single run in scope) run_state is null and jobs is empty. See spec.md → Operator read-surface state projection.

Records

striatum records docket <run-id> [--format markdown|json]
striatum records migration inventory [--root path]... [path ...]
striatum records migration import --manifest inventory.json [--import-batch-id id] [--dry-run] [--json]
striatum records migration verify --manifest inventory.json [--json]
striatum records migration materialize (--record-id id | --import-batch-id id | --source-path path) [--out .striatum/scratch/path] [--json]

records docket is daemon-backed and renders a compact RFC 0171 docket from artifact and generated-record indexes. records migration inventory remains a local read-only filesystem scan. The import, verify, and materialize migration commands are CLI wrappers over daemon RPC: import uploads safe manifest entries through the daemon blob client and upserts generated_records; verify compares the original manifest to daemon-fetched blob bytes; materialize writes verified bodies only under ignored .striatum/scratch. None of these commands delete tracked historical files.

Daemon and multi-repo registry (RFC 0028 V1)

striatum daemon install [--no-start] [--print-unit]
striatum daemon status
striatum daemon uninstall
striatum daemon migrate-db [--admin-url <dsn>] [--json]
striatum daemon owner-ddl apply [--owner-url <dsn>] [--json]
striatum daemon token-create <capability>
striatum doctor [--verbose] [--json]
striatum doctor --lane-provider-auth codex|claude [--run-id <id>] [--lane-id <id>] [--timeout 45s] --json
striatumd [daemon-start options]
systemctl --user start|stop|restart|status striatumd
striatum repo add <path> [--init] [--display-name <name>] [--no-migrate] [--apply-blob-creation]
striatum repo list
striatum repo remove <id>

striatum daemon install renders the systemd user unit, scaffolds daemon.toml when absent, and enables/starts striatumd unless --no-start is passed. Use systemctl --user start|stop|restart striatumd for service lifecycle after installation. On hosts without systemd user services, run striatumd -socket "${XDG_RUNTIME_DIR}/striatum/daemon-go.sock" directly as described in the daemon runbook.

If a system unit (/etc/systemd/system/striatumd.service) already owns the daemon, striatum daemon install refuses with exit code 1 and daemon_install_system_unit_present rather than writing a conflicting user-scope unit. The user-scope unit carries no owner-DB bootstrap environment, so re-creating it next to a system unit crash-loops the daemon and takes it down for every session (#509). On such a host, manage the daemon with sudo systemctl restart striatumd and use --print-unit to inspect the unit template without installing it.

striatumd is the supported foreground daemon process. Per D094 / RFC 0043 the daemon is a hard prerequisite for every Striatum verb; CLI verbs without a reachable daemon refuse with exit code 11 (daemon_unreachable) and do not fall back to direct mode.

On first successful startup, striatumd bootstraps a single admin token when daemon-owned Postgres has no clients and writes a 0600 runtime client-token file. Token secrets are never read from environment variables, never logged to audit, and never stored in the registry.

Ordinary doctor and doctor --verbose do not run provider CLIs. The provider-auth diagnostic is explicit-only because provider-specific checks may touch the network, spend provider tokens, trigger auth refresh, or hang on an interactive prompt. The Claude diagnostic is offline-only and reads credential freshness without running a provider command or spending model tokens. Its JSON result is the same private-safe block used by supervise.start, with raw_output_returned=false. A Codex zero-exit smoke with a missing or mismatched bounded success signal reports success_signal as missing, empty, or mismatch; it is not an auth failure. Authorization uses the closed daemon method capability vocabulary: read, write, review, claim, apply, admin, recovery, and surgical_recovery.

striatum daemon token-create <capability> (method daemon.token.create, capability admin) mints an additional scoped capability token against the daemon-owned Postgres. The capability argument is one of the closed vocabulary above. The secret is returned once and never re-read from the registry.

repo add canonicalizes the repository root, refuses symlink/path-traversal ambiguity, derives a realpath/inode-based repository identity, and refuses active path re-occupation by a different identity. Pass --init when no .striatum/ directory exists; it creates operational scratch only and does not create .striatum/retired-local-state. If a pre-D094 repo-local SQLite source exists, registration refuses and tells the operator to archive/remove the legacy SQLite file before registering. --display-name sets the operator-facing repository name (defaults to the directory basename); --apply-blob-creation creates the per-repository blob-storage bucket when the daemon is configured for blob storage and the bucket is absent. --no-migrate is an accepted compatibility flag — production registration never imports retired SQLite state.

repo remove is idempotent, revokes live repo-scoped capabilities, preserves audit rows, and never reuses repository_id; re-adding allocates a fresh id.

The resident recovery sweep runs inside striatumd. Use the daemon-backed recovery auto family for explicit recovery diagnostics where applicable; there is no striatum daemon sweep CLI command.

RFC 0033 V2 accepts system PostgreSQL as the daemon-owned storage substrate for daemon-global state. Configure it with STRIATUM_DAEMON_DB_URL, daemon config, or an explicit --postgres-url client surface. The daemon owns schema migrations and roles, but it does not install, start, stop, or upgrade PostgreSQL. Bundled, embedded, and Dockerized Postgres distributions are deferred.

striatum doctor is the daemon-backed health check. Top-level warnings name non-red findings that still have operator work behind them; top-level notices preserve advisory historical/local-environment facts that do not name daemon recovery work. doctor --verbose includes structured problem_records, warning_records, and notice_records alongside the stable string lists, including RFC 0171 generated_record_integrity records for missing, corrupt, swapped, duplicate-source, or metadata-incomplete generated-record blobs. doctor --lane-provider-auth codex|claude [--run-id <id>] [--lane-id <id>] [--timeout 45s] --json is an explicit provider-auth diagnostic for the lane identity; ordinary doctor and doctor --verbose never run provider CLIs. striatum daemon status is the local bootstrap summary for unit state and runtime paths; it folds in read-only doctor information when the daemon is reachable.

daemon migrate-db [--admin-url <dsn>] [--json] (RFC 0079 §5) applies pending daemon PostgreSQL schema migrations using an owner/admin DSN, so DDL the runtime role (striatumd_rw) cannot perform — e.g. a migration that adds a foreign key to an owner-held table — is applied before the daemon serves. The admin DSN is resolved from --admin-url, then STRIATUM_DAEMON_ADMIN_DB_URL, then the normal daemon DSN (flag/env/daemon.toml) as a fallback for additive migrations the runtime role can apply itself. This is distinct from the retired SQLite-era daemon migrate below.

daemon migrate --from sqlite --to pg and daemon migrate-repo-local --from sqlite --to pg --repo <path> are fully removed SQLite-era import spellings. They are no longer parseable compatibility commands; stale automation receives an unknown-command parse failure. Use striatum daemon status, striatum doctor --verbose --json, and striatum repo add <path> --init for current registration/cutover diagnostics. CLI verbs against an unregistered repo refuse with exit code 12 (repo_not_migrated) and point operators to archive/remove legacy SQLite files, then register with repo add --init.

RFC 0030/0031 add the daemon V2 RPC and supervision/apply foundation on top of RFC 0033. The wire envelope is versioned JSON; daemon.hello negotiates envelope/framing, daemon.describe publishes the method registry and methods_etag, and incompatible clients refuse with exit code 10. RFC 0048 completed the production handler-port work in v1.49.0-v1.55.0: mapped production verbs are daemon/Postgres-backed and fail closed without daemon reachability, repository registration, and capability authorization. Legacy SQLite paths remain only for migration sources, golden fixtures, and explicitly gated subprocess compatibility tests under STRIATUM_DAEMON_REQUIRED=0 STRIATUM_TEST_HARNESS=1.

Daemon-required runtime

Production commands route through the daemon authority boundary by default. There is no global --daemon flag in the Go CLI; daemon connectivity is required and assumed.

striatum status [--run-id <id>]
striatum doctor
striatum why <job-id>
striatum dashboard --all

The V1 --no-daemon flag is retired (D094 / RFC 0043); parsing it returns the standard argparse "unrecognized arguments" error and exit code 2. Production mutation and read verbs do not fall back to direct repo-local mode.

Daemon RPC method capabilities use the closed vocabulary read, write, review, claim, apply, admin, recovery, and surgical_recovery. supervise.* and apply.* are daemon RPC routes; sealed apply fails closed unless a daemon signing key and apply capability are present. The Go daemon rotates the local Ed25519 fallback signing key through the admin RPC method daemon.key.rotate; there is not yet a stable user-facing striatum keys CLI.

D270 retires the cross-repo CLI surface. Current CLI workflows are single-repository units; there is no supported striatum cross-repo ... command family. Historical RFC 0032 schema may remain in upgraded databases, but no current CLI route exposes cross-repo lifecycle, read, or recovery behavior. Daemon MCP tools/list is filtered by each token's effective capabilities and scope, and tools/call re-checks authorization and audits denials.

RFC 0036 adds no new CLI verb. Regenerate agent-facing MCP guidance with striatum skills install or striatum plugin install; chat workflow generation uses the existing local service and RFC 0034 generator paths.

Skills (RFC 0015)

striatum skills install

Codex launcher

striatum codex [codex args ...]

striatum codex launches the codex CLI pre-wired to the live daemon MCP endpoint: it injects -c mcp_servers.striatum.url=<live endpoint> (overriding ~/.codex/config.toml) and -c mcp_servers.striatum.bearer_token_env_var=STRIATUM_MCP_TOKEN, and sets STRIATUM_MCP_TOKEN in codex's environment from the runtime client-token (never printed). Extra arguments pass through to codex unchanged.

Daemon-Mounted HTTP Service (RFC 0012 / 0013 / 0085)

striatumd -mcp-http-addr 127.0.0.1:0
BASE_URL=$(sed 's#/mcp$##' "$XDG_RUNTIME_DIR/striatum/mcp-http-endpoint")
TOKEN=$(cat "$XDG_RUNTIME_DIR/striatum/client-token")
curl -H "Authorization: Bearer $TOKEN" "$BASE_URL/v1/health"

Web routes (RFC 0013 / 0022 / 0024 / 0038)

striatumd mounts the Go web service on the same localhost-only HTTP listener as daemon MCP. The endpoint file includes /mcp; strip that suffix for web routes. The loopback service requires Authorization: Bearer <client-token>; there is no separate striatum serve command and no serve --web flag. Mutations are read-only by default and require STRIATUM_DAEMON_WEB_ALLOW_MUTATIONS=1 on the daemon process before startup.

For ordinary browser access without bearer-header tooling, run the optional read-only tailnet identity listener (striatumd -web-tailscale or STRIATUM_DAEMON_WEB_TAILSCALE=1) and point tailscale serve --bg unix:$XDG_RUNTIME_DIR/striatum/web-ui.sock at the owner-only socket.

Route	Surface
/	Run list (RFC 0013/0022/0037).
/run?run_id=<id>	Server-rendered run detail/status page.
/v1/health	Service health and mutation posture.
/v1/runs	Daemon status JSON.
/v1/runs/<id>	Daemon status --run-id JSON.
/v1/runs/<id>/events	Server-Sent Events over daemon run.events.
/v1/runs/<id>/dashboard	Dashboard DTO for the run.
/v1/runs/<id>/why?id=<entity>	Daemon why JSON.
/v1/runs/<id>/artifacts	Artifact list for the run.
/v1/artifacts/<id>/raw	Raw artifact content.
/workflow-templates	Workflow template catalog list.
/workflow-templates/<id>	Workflow template catalog entry.
/workflows/generate/preview	Workflow generator preview (POST).
/workflows/generate	Workflow generator write (POST; requires web mutations).

List (read-only enumeration)

striatum list runs
striatum list sessions
striatum list jobs
striatum list artifacts
striatum list workflows

list runs includes the workflow identity triple for each run: workflow_id, workflow_version, and workflow_snapshot_id. The web run list uses the same snapshot identity to display the workflow name and link back to the workflow detail when the source path is known.

Inspection and recovery

striatum status
striatum why
striatum work packet-show
striatum doctor
striatum git snapshot
striatum git commit-apply
striatum evidence export
striatum run graph
striatum recovery auto
striatum recovery auto-publish
striatum recovery auto-finalize
striatum recovery stale-leases
striatum recovery requeue-stale
striatum recovery cancel-job
striatum recovery invalidate-job
striatum recovery process-reconcile
striatum recovery resume
striatum recovery reseal
striatum recovery complete-stalled
striatum recovery accept-quarantined
striatum recovery resolve-blocker
striatum recovery prune-debris
striatum recovery quarantine-lane
striatum checkpoint resolve
striatum override-verdict

recovery --help (and bare striatum recovery) lists the full recovery verb family above; run striatum recovery <subcommand> --help for a subcommand's flags. Each recovery verb requires the recovery capability. All recovery verbs are diagnostic/compatibility clients of the same daemon boundary — they route remediation back through the daemon's own state machine and never hand-finish a wedged lane (see AGENTS.md, "Do not paste over a broken runner").

Recovery verb family (when each applies)

The recovery verbs are the legitimate way to unstick a wedged lane or run. The non-obvious entry point is the "lane finished but the seal failed" path: a transient artifact.publish / seal failure (e.g. a DB lock timeout, SQLSTATE 55P03, under concurrent runs) can leave the work done and durable on disk but the job in state blocked, not requeued. run retry-job refuses (it would exceed max_attempts), and run drive cannot launch a blocked job — it now reports cannot advance <job> (cannot_advance_blocked): … instead of idling silently. The fix is:

striatum recovery reseal --run-id <run-id> --job-id <job-id>

recovery reseal moves the job blocked -> queued on the SAME attempt so the driver re-claims and re-seals the already-durable artifact; it does not consume a retry.

• recovery reseal <run-id> <job-id> — re-seal a job whose lane finished but whose seal/publish failed transiently; blocked -> queued, same attempt.
• recovery resolve-blocker <blocker-id> — close a dangling non-escalation, non-checkpoint blocker that no completion path cleared.
• recovery complete-stalled <run-id> <job-id> — finalize a recovery-exhausted job from its already-durable published artifacts when the lane is dead and the work is complete. Review / phase-synthesis jobs are refused by default; after inspection, --force may recover a dead or blocked verdict-capable job only from an already-published required finding or synthesis artifact whose front matter carries a recoverable verdict (verdict_intent or status).
• recovery accept-quarantined <run-id> <job-id> — resolve a quarantined job's blocker and mark it canceled-by-operator.
• recovery prune-debris <run-id> — tombstone a terminal-debris run's unrecoverable artifact debris so doctor reports ok again.
• recovery quarantine-lane <run-id> <job-id> — snapshot a terminal run's dirty lane worktree to a durable quarantine ref, then clean the worktree.

striatum doctor surfaces this wedge family directly: a non-terminal job in running / claimed / stale_lease / blocked with no live session and no recent progress is reported as a job_stuck_no_live_session warning naming the recovery verb to run. The warning is advisory (it does not flip doctor to a hard red) so it tolerates normal in-flight latency.

run graph --run-id <id> [--format mermaid|json|dot|ascii] renders the workflow graph for an existing run with each node colored by current job state. Mermaid output appends classDef/class lines; JSON adds current_state, attempt, and a latest_verdict block on review nodes; ascii reuses the dashboard's graph panel renderer (RFC 0016).

recovery resume resolves remediated process-adapter blockers with the preserved lease. For remediated write-scope dirty-path blockers, it validates that the tree is clean, resolves the blocker, and requeues the same attempt for a fresh claim before completion. It is also the named recovery destination for write-scope drift failures: recovery must audit the remediation or override and must not mutate the historical attempt scope that caused publish-artifact or complete to fail.

git snapshot --json [--ancestry-limit N] [--no-ancestry] emits the daemon read-only git.snapshot projection for the registered target repository: local branch, HEAD metadata, dirty counts, changed paths, and bounded ancestry. It does not fetch, push, commit, read remote URLs, or include diff hunks or commit bodies.

git commit-apply <commit-request-path> --confirm --confirm-request-id <id> --json emits daemon method git.commit_apply. It creates only a local commit from a striatum.commit_request.v1 artifact whose confirmation_status is already operator_confirmed or human_confirmed. It refuses base-HEAD, branch, or dirty-path mismatches, disables repository Git hooks for the commit invocation, and never pushes or calls hosted providers.

recovery auto emits the daemon recovery.sweep method. The sweep runs workflow-opt-in recovery.auto_finalize before lazy lease expiry, then the existing stale-lease, process-reconcile, and review-only requeue recovery pieces where policy allows. When live recovery creates escalation rows, STRIATUM_ESCALATION_NOTIFY_URL opts into a best-effort post-commit JSON POST to loopback or tailnet HTTP(S) targets only. Dry-runs do not notify, and notification success or failure is not authoritative. recovery auto-publish emits the explicit recovery.auto_publish_stale_artifacts method. The deprecated recovery.auto daemon RPC alias is retired; stale direct RPC calls audit as method_unknown. D184 allows the sweep to auto-cancel an abandoned running run after the default 24h threshold when there are no live sessions, no live supervisors or supervised processes, no active leases, and no progress or durable events in the threshold window. Any live-work evidence or inconclusive liveness probe fails closed and leaves the run non-terminal for operator inspection; there is no needs_operator intermediate when the abandonment predicate is proven.

recovery invalidate-job <job-id> <decision-id> (RFC 0118, capability recovery) invalidates a completed job's provenance against an accepting run-level decision and records a durable supersede receipt so a fresh attempt can re-drive the work.

work packet-show (capability read) inspects work-packet metadata. Select a packet with [packet-id] or --message-id / --job-id / --session-id / --run-id (at least one selector required); --limit caps rows (default 20, cap 200). It returns metadata and packet_sha256 only; pass --raw to include packet_json, which is omitted by default to avoid leaking task prose.

Corpus export (RFC 0044 V1 / RFC 0057 contract)

striatum corpus export --since <ref> --out <dir>

corpus export emits a redacted JSONL bundle of Striatum's durable provenance (RFCs, decision-log rows, operator reports, run summaries, audit-chain entries, changelog entries, ubiquitous-language terms, harness-friction patterns, recent commits) plus a verifying manifest.json with explicit state_authority metadata. Re-running over unchanged inputs produces byte-identical JSONL files and stable per-file SHA-256s; only generated_at varies, and it is excluded from the bundle digest. The bundle is operator-triggered local provenance, never streamed to any external service. Optional consumers (Engram is the first reference under RFC 0044) may ingest the bundle for retrieval, but Striatum does not call them at runtime and runs identically when no consumer is configured. The V2 contract decisions (multi-corpus identity, redaction-tier metadata, incremental watermarks, optional context-injection policy) are scoped by RFC 0057.

Run archive

striatum archive create --run-id <id> --out <dir>

archive create is a daemon/Postgres-backed read command that writes a local archive directory for one run. The V2 archive contains the run row, workflow snapshot, run-scoped rows, artifact metadata, event metadata, and a self-verifying manifest.json; it does not copy artifact contents, transcripts, or .striatum/ scratch. The current Go CLI exposes archive creation only; local archive verify / archive inspect verifier commands are not part of the active command surface.

Adapter

striatum adapter run

adapter run is retired outside the explicit legacy test-fixture compatibility environment. Use daemon-supervised process lanes instead.

Session lifecycle

striatum session close

Stable exit codes

• 0: success, including claim-next with no_work.
• 1: generic / unhandled runtime error.
• 2: CLI usage error (argparse).
• 3: missing run, session, job, message, blocker, artifact, verdict, or session target.
• 4: invalid state transition.
• 5: lease expiry or ownership mismatch.
• 6: artifact or write-scope violation.
• 7: branch confirmation required before work can be claimed.
• 8: workflow config rejected (also raised by branch confirm when a requested git operation cannot be performed).
• 9: state schema is newer than this striatum install supports (daemon PostgreSQL in production; legacy SQLite only in fixture paths).
• 10: daemon RPC transport, handshake, or version-skew refusal.
• 11: daemon_unreachable. The CLI could not reach the daemon socket; stderr names the socket path and remediation. No SQLite fallback is attempted.
• 12: repo_not_migrated. The target repository is not registered for daemon/PostgreSQL state or still has a legacy .striatum/retired-local-state; stderr and the --json hint tell the operator to archive/remove legacy SQLite files and register with repo add --init.

See also

• HOW_TO_HUMAN.md — the operator's playbook with examples per verb.
• HOW_TO_AGENT.md — the coding-agent companion to the RFC 0015 skill bundle.
• SPEC.md — the implementation contract.