agent-lifecycle-kit 2.43.0__tar.gz → 2.44.1__tar.gz

{agent_lifecycle_kit-2.43.0/src/agent_lifecycle_kit.egg-info → agent_lifecycle_kit-2.44.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: agent-lifecycle-kit
-Version: 2.43.0
+Version: 2.44.1
 Summary: Provider-neutral lifecycle control layer that helps coding agents plan, execute, validate, and prove software tasks to completion.
 Author: Agent Lifecycle Kit contributors
 License: Apache-2.0
@@ -42,7 +42,7 @@ Release 2.28 adds [bounded lifecycle and accounting baselines](docs/reference/pe
 ## Quick start
 Start with [install ALK and make the first run](docs/guides/install-and-first-run.md). Release 2.23 adds [compact S0/S1 authoring](docs/reference/tier-proportional-plan-packages.md) with explicit checklist completion and unchanged approval gates. Release 2.22 adds [conservative plan-review reuse](docs/reference/conservative-delta-review.md) with independent current applicability and no automatic full reruns. Release 2.21 adds [command accounting, bounded preparation and response recovery](docs/reference/source-native-workflow-accounting.md). Release 2.20 adds [target-bound initialization](docs/reference/cli.md#release-220-composed-initialization-and-digest-display), [bounded local continuation](docs/reference/workflow-continuation.md) and batched Git snapshots. Release 2.19 adds [safe artifact authoring and planning preflight](docs/reference/safe-artifact-authoring.md), [bounded review handoff](docs/reference/review-mesh.md#bounded-review-handoff) and thin skills with complete references. Release 2.18 bounds process/Git diagnostics, hardens signing and distinguishes OpenCode 1.x/2.x compatibility ([adapter notes](docs/adapters/opencode.md)); 2.17 strengthens [workflow validation and recovery](docs/reference/workflow-completeness.md). Release 2.16 introduced cumulative accounting within verified host capabilities; see [retention maintenance](docs/operations/retention-maintenance.md).
 
-It includes macOS, Linux and Windows instructions, Python 3.11-3.15, [PyPI](https://pypi.org/project/agent-lifecycle-kit/) installation with `python -m pip install agent-lifecycle-kit==2.43.0`, and common `agent-lifecycle version` errors. See [2.15 migration notes](docs/reference/cli.md#migrating-authority-artifacts-to-215) and [authority boundaries](docs/reference/authority-artifact-boundaries.md).
+It includes macOS, Linux and Windows instructions, Python 3.11-3.15, [PyPI](https://pypi.org/project/agent-lifecycle-kit/) installation with `python -m pip install agent-lifecycle-kit==2.44.1`, and common `agent-lifecycle version` errors. See [2.15 migration notes](docs/reference/cli.md#migrating-authority-artifacts-to-215) and [authority boundaries](docs/reference/authority-artifact-boundaries.md).
 
 For a quick start, run:
 

{agent_lifecycle_kit-2.43.0 → agent_lifecycle_kit-2.44.1}/README.md

@@ -19,7 +19,7 @@ Release 2.28 adds [bounded lifecycle and accounting baselines](docs/reference/pe
 ## Quick start
 Start with [install ALK and make the first run](docs/guides/install-and-first-run.md). Release 2.23 adds [compact S0/S1 authoring](docs/reference/tier-proportional-plan-packages.md) with explicit checklist completion and unchanged approval gates. Release 2.22 adds [conservative plan-review reuse](docs/reference/conservative-delta-review.md) with independent current applicability and no automatic full reruns. Release 2.21 adds [command accounting, bounded preparation and response recovery](docs/reference/source-native-workflow-accounting.md). Release 2.20 adds [target-bound initialization](docs/reference/cli.md#release-220-composed-initialization-and-digest-display), [bounded local continuation](docs/reference/workflow-continuation.md) and batched Git snapshots. Release 2.19 adds [safe artifact authoring and planning preflight](docs/reference/safe-artifact-authoring.md), [bounded review handoff](docs/reference/review-mesh.md#bounded-review-handoff) and thin skills with complete references. Release 2.18 bounds process/Git diagnostics, hardens signing and distinguishes OpenCode 1.x/2.x compatibility ([adapter notes](docs/adapters/opencode.md)); 2.17 strengthens [workflow validation and recovery](docs/reference/workflow-completeness.md). Release 2.16 introduced cumulative accounting within verified host capabilities; see [retention maintenance](docs/operations/retention-maintenance.md).
 
-It includes macOS, Linux and Windows instructions, Python 3.11-3.15, [PyPI](https://pypi.org/project/agent-lifecycle-kit/) installation with `python -m pip install agent-lifecycle-kit==2.43.0`, and common `agent-lifecycle version` errors. See [2.15 migration notes](docs/reference/cli.md#migrating-authority-artifacts-to-215) and [authority boundaries](docs/reference/authority-artifact-boundaries.md).
+It includes macOS, Linux and Windows instructions, Python 3.11-3.15, [PyPI](https://pypi.org/project/agent-lifecycle-kit/) installation with `python -m pip install agent-lifecycle-kit==2.44.1`, and common `agent-lifecycle version` errors. See [2.15 migration notes](docs/reference/cli.md#migrating-authority-artifacts-to-215) and [authority boundaries](docs/reference/authority-artifact-boundaries.md).
 
 For a quick start, run:
 

agent_lifecycle_kit-2.44.1/docs/guides/deterministic-workflow-recipes.md

@@ -0,0 +1,6 @@
+# Deterministic workflow recipes
+
+The complete commands, refusal cases and authority boundaries are canonical in
+the [lifecycle cookbook](lifecycle-cookbook.md#deterministic-workflow-recipes).
+Use that bundled section for artifact creation, inspection and bounded local
+continuation. These operations make no token or cost saving claim.

agent_lifecycle_kit-2.44.1/docs/guides/lifecycle-cookbook.md

@@ -0,0 +1,554 @@
+# Lifecycle task scenarios
+
+Raw input stays draft-only until reviewed authority permits execution.
+
+## Canonical execution route
+
+If the route is known, call it directly with bound arguments, for example:
+`agent-lifecycle start --adapter codex --mode research --file work/tasks/research.md --out work/tasks/research-start.json`.
+For an unknown implementation transition, project `workflow continue` first;
+apply only its exact revision/action digest. Read [Workflow continuation](../reference/workflow-continuation.md)
+when using that route. Existing `--until-blocked --apply` needs a ready explicit
+bundle and positive transition/I/O caps; it stops before model, host, reviewer,
+operator or plan-authority work. Direct `workflow task-*` remains valid.
+Legacy `runner` receipts are read-only compatibility journals, never authority
+to authorize, accept or finalize; attach evidence to workflow instead.
+
+Run fast feedback before mandatory acceptance checks and the full quality floor.
+`FULL_REQUIRED` requests repair/plan review, not a whole-repository audit.
+For an unchanged external `WAITING` blocker, save handoff and wait for its event;
+do not relaunch a model, repeat discovery or rerun an unchanged successful check.
+
+## Choose a recipe
+
+Read this common route, this table, [Safety rules](#safety-rules), the selected
+recipe and its required links. Inspect durable state if the stage is unclear;
+change recipe when the stage changes. Keep the complete hash-bound role contract,
+frozen plan, active packet and shared contracts. The orchestrator's
+[complete lifecycle contract](../../skills/agent-workflow-orchestrator/references/lifecycle-contract.md)
+remains required. The [local writing profile](controlled-technical-english.md)
+contains terminology, assigning-controller role suitability and measurement limits.
+
+| Need | Recipe |
+| --- | --- |
+| Research or plan, including S0/S2 authoring intents | [Research and planning only](#research-and-planning-only) |
+| One task file | [Review one Markdown task](#review-one-markdown-task) |
+| Plan folder | [Review a Markdown plan folder](#review-a-markdown-plan-folder) |
+| Code or PR/MR | [Review code changes](#review-code-changes) |
+| Completed work | [Audit implementation evidence](#audit-implementation-evidence) |
+| Requested fixes | [Open a remediation attempt](#open-a-remediation-attempt) |
+| Defect | [Bug Forensics repair](#bug-forensics-repair) |
+| Several reviewers | [Coordinate cross-review](#coordinate-cross-review) |
+| Active run | [View goal and progress](#view-goal-and-progress) |
+| Opted-in risk-aware frozen task | [Run a risk-aware task](#run-a-risk-aware-task) |
+| v4 task transition | [Use the v4 task-local workflow](#use-the-v4-task-local-workflow) |
+| Recovery | [Recover a run without editing state](#recover-a-run-without-editing-state) |
+| Structured JSON or authorized local transitions | [Deterministic workflow recipes](#deterministic-workflow-recipes) |
+| New capability before freeze | [Bounded capability feasibility before freeze](#bounded-capability-feasibility-before-freeze) |
+
+## Research and planning only
+
+Use this when the deliverable is analysis, architecture notes or a plan.
+
+```bash
+cat > work/tasks/research.md <<'EOF'
+# Task
+
+Research the current adapter session flow and write an implementation plan.
+Do not change code.
+
+Return:
+- architecture observations;
+- risks and open questions;
+- recommended ALK plan shape;
+- validation commands that would be needed later.
+EOF
+
+agent-lifecycle start \
+  --adapter codex \
+  --mode research \
+  --file work/tasks/research.md \
+  --out work/tasks/research-start.json
+
+agent-lifecycle review-mesh recommend \
+  --file work/tasks/research.md \
+  --out work/tasks/research-recommendation.json
+```
+
+Stop here if you only need research. To implement later, turn the accepted plan
+into a normal ALK plan package and freeze it.
+
+## Review one Markdown task
+
+Use this for a task, proposal or plan saved in one file:
+
+```bash
+agent-lifecycle start \
+  --adapter claude \
+  --mode review \
+  --file tasks/proposal.md \
+  --out work/review/proposal-start.json
+
+agent-lifecycle review-mesh recommend \
+  --file tasks/proposal.md \
+  --out work/review/proposal-recommendation.json
+```
+
+The adapter id selects the host-local context for the intake receipt. It does
+not start the host or model for raw Markdown.
+
+## Review a Markdown plan folder
+
+Use the import command when the plan is split across several Markdown files:
+
+```bash
+agent-lifecycle import plan \
+  --source tasks/release-1-40/ \
+  --dialect spec-kit \
+  --out work/review/plan-import.json
+```
+
+For reviewer-facing intake, create one task that points at the folder or at the
+import receipt:
+
+```bash
+cat > work/review/plan-review-task.md <<'EOF'
+# Task
+
+Review the Markdown plan package represented by work/review/plan-import.json.
+Do not implement.
+
+Check requirements, acceptance criteria, evidence routes, write ownership,
+security gates and release claims.
+EOF
+
+agent-lifecycle start \
+  --adapter codex \
+  --mode review \
+  --file work/review/plan-review-task.md \
+  --out work/review/plan-review-start.json
+```
+
+## Review code changes
+
+Prepare a diff with Git, GitHub CLI or GitLab refspecs, then pass one explicit
+review task to ALK:
+
+```bash
+mkdir -p work/code-review/current
+git diff origin/main...HEAD > work/code-review/current/diff.patch
+```
+
+```bash
+agent-lifecycle start \
+  --adapter codex \
+  --mode review \
+  --file work/code-review/current/review-task.md \
+  --out work/code-review/current/start.json
+```
+
+When architecture is documented, list the architecture documents in the task
+file. When it is not documented, ask the reviewer to recover module
+responsibilities first, then review the diff against that recovered map.
+
+Full examples are in [Code review workflows](code-review-workflows.md).
+
+## Run a risk-aware task
+
+Use this only after the plan is reviewed and frozen. First project the exact
+profile without mutating workflow state:
+
+```bash
+agent-lifecycle start \
+  --adapter codex \
+  --mode implement \
+  --risk auto \
+  --file tasks/my-release/plan.manifest.json \
+  --state work/my-release/run.state.json \
+  --lock tasks/my-release/plan.lock.json \
+  --task WS-01 \
+  --operation-id start-WS-01-attempt-1 \
+  --expected-revision 3 \
+  --source-revision "$(git rev-parse HEAD)" \
+  --host-model-profile profiles/hosts/codex-live-profile.v1.json \
+  --risk-profile-out work/my-release/WS-01/risk-profile.json
+```
+
+Then authorize the attempt with the same operation id:
+
+```bash
+agent-lifecycle workflow task-start \
+  --state work/my-release/run.state.json \
+  --task WS-01 \
+  --operation-id start-WS-01-attempt-1 \
+  --expected-revision 3 \
+  --source-revision "$(git rev-parse HEAD)" \
+  --risk-profile work/my-release/WS-01/risk-profile.json \
+  --reason "start risk-aware attempt"
+```
+
+The host must later attest tokens, invocation count and wall time. Any missing,
+estimated, lineage-drifted or over-cap value blocks the result transition. See
+[Risk-aware execution](../reference/risk-aware-execution.md) for all inputs and
+failure rules.
+
+## Audit implementation evidence
+
+Use implementation audit after a worker has produced a task result and review:
+
+```bash
+agent-lifecycle audit implementation \
+  --manifest tasks/release-x/plan.manifest.json \
+  --state work/run.state.json \
+  --task WS-01 \
+  --result work/WS-01/attempt-1/task-result.json \
+  --review work/WS-01/attempt-1/task-review.json \
+  --out work/WS-01/attempt-1/implementation-audit.json
+```
+
+Add `--review-mesh-quorum <path>` only when the frozen plan requires a quorum
+receipt for that phase.
+
+Follow the frozen validation sequence. If a check fails, stop the sequence and
+diagnose it before running checks that a correction would invalidate. Retain
+failed receipts and all costs. Reuse a successful check only when its exact
+source bytes, input digests and dependencies justify reuse; retain those bytes
+before a correction. Use canonical authority artifacts as audit inputs, not
+reformatted display JSON. See [Safe artifact authoring](../reference/safe-artifact-authoring.md).
+
+## Open a remediation attempt
+
+Use this when an independent task review or implementation audit returns
+`REWORK` for findings that remain inside the frozen task scope. The plan must
+use `remediationMode: ask` or `bounded-auto` and set `maxTaskAttempts` to at
+least 2.
+
+Before submitting each task result, calculate the current Git-backed claim:
+
+```bash
+agent-lifecycle workflow task-snapshot \
+  --state work/run.state.json \
+  --task WS-01 \
+  --out work/WS-01/attempt-1/task-snapshot.json
+```
+
+The worker places the returned `claim` object in the task result as
+`changeSet`. After `task-result` and independent review, request rework with
+the exact open finding IDs. In task-review findings, use the schema values
+`open` and `closed` in lowercase; the review verdict remains `REWORK`:
+
+```bash
+agent-lifecycle workflow task-rework \
+  --state work/run.state.json \
+  --task WS-01 \
+  --operation-id rework-WS-01-attempt-1 \
+  --expected-revision 7 \
+  --source-revision <source-sha> \
+  --review work/WS-01/attempt-1/task-review.json \
+  --finding-id F-101 \
+  --reason "address independent review findings"
+```
+
+Add `--implementation-audit <path>` when the plan requires implementation
+audit or when that audit is the artifact that returned `REWORK`. The command
+archives artifact identities, not duplicate bodies. `workflow run` then
+returns the same task for another `task-start`, which opens the first unused
+attempt path. Previous result, review and audit files remain immutable.
+
+## Use the v4 task-local workflow
+
+For a new run from a reviewed frozen plan, project the initialization from the
+repository root. Use the state path and package id declared by that plan:
+
+```bash
+agent-lifecycle workflow init \
+  --manifest tasks/release-x/plan.manifest.json \
+  --state work/run.state.json \
+  --run-id run-001 \
+  --package-id release-x
+```
+
+This projection is read-only. Apply with its current full `bindingDigest`:
+
+```bash
+agent-lifecycle workflow init \
+  --manifest tasks/release-x/plan.manifest.json \
+  --state work/run.state.json \
+  --run-id run-001 \
+  --package-id release-x \
+  --apply \
+  --projected-target-digest <full-binding-digest> \
+  --operation-id initial-adoption
+```
+
+The composed route compiles packets in `<artifactRoot>/workflow/task-packets`,
+creates v4 state and adopts the plan in `AWAITING_AUTHORIZATION`. It does not
+start execution. Use the separate existing authorization route. Without
+`--manifest`, `workflow init` creates unbound state only. For unsupported
+composition policies, use the existing explicit compile/init/adopt route. See
+[Composed initialization](../reference/cli.md#release-220-composed-initialization-and-digest-display)
+for selector checks and failure rules.
+
+For a legacy v3 state, use the explicit migration command and keep the source
+revision and expected state revision bound to the migration receipt:
+
+```bash
+agent-lifecycle workflow state-migrate \
+  --state work/run.state.json \
+  --operation-id migrate-run-001 \
+  --expected-revision 1 \
+  --source-revision <source-sha>
+```
+
+After a task result and independent review exist, apply the single canonical
+task outcome. `ACCEPTED`, `REWORK`, `CONTRACT_CHANGE` and `BLOCKED` are
+task-local outcomes; an active sibling keeps running and earlier attempt
+artifacts remain immutable:
+
+```bash
+agent-lifecycle workflow task-review-apply \
+  --state work/run.state.json \
+  --task WS-01 \
+  --operation-id review-WS-01-attempt-1 \
+  --expected-revision 7 \
+  --source-revision <source-sha> \
+  --review work/WS-01/attempt-1/task-review.json \
+  --reason "apply independent review"
+```
+
+The result is already attached by `workflow task-result`; `task-review-apply`
+has no `--result` option. The example is checked against the actual parser and `--help`.
+
+
+The run enters `FINAL_AUDIT` only after every required task is accepted. The
+legacy `task-accept` and `task-rework` commands remain compatibility wrappers
+for one release and use the same transition service.
+
+## Recover a run without editing state
+
+Approval-required runs consume an exact-lineage receipt through
+`workflow authorize`; `PLAN_ONLY` runs never become executable. Pause a
+supported phase for one host-owned operation with `workflow external-pause`
+and resume it with `workflow external-resume` only after the matching receipt
+has been written. Typed blockers expose their legal route, so a generic
+`resolve-blocker` cannot clear a task or plan blocker.
+
+After the required tasks are accepted, apply the independent final-audit
+decision explicitly:
+
+```bash
+agent-lifecycle workflow final-audit-outcome \
+  --state work/run.state.json \
+  --operation-id final-audit-outcome-1 \
+  --expected-revision 12 \
+  --source-revision <source-sha> \
+  --final-audit final/final-audit.json \
+  --verdict REWORK \
+  --task-id WS-01 \
+  --finding-id F-001 \
+  --reason "address the independent finding"
+```
+
+`ACCEPTED` proceeds to the existing finalization gates. `REWORK` archives the
+named accepted attempts and preserves their files, `CONTRACT_CHANGE` requires
+a new frozen plan, and `BLOCKED` waits for the declared external action. None
+of these transitions changes a frozen plan or raises a retry budget.
+
+## View goal and progress
+
+Use this when another adapter is working and you need a compact, read-only
+status line:
+
+```bash
+agent-lifecycle goal view \
+  --record work/run/goal.json \
+  --state work/run/state.json \
+  --usage-receipt work/run/usage.json \
+  --change-summary work/run/change-summary.json \
+  --terminal
+```
+
+The command does not claim completion. It only combines the goal record,
+workflow state, progress rows, token receipts and Git-style change summary.
+
+## Bug Forensics repair
+
+For defects, regressions, flaky failures, incidents and security bugs, start
+with task intake and let ALK add advisory-only Bug Forensics markers:
+
+```bash
+agent-lifecycle start \
+  --adapter codex \
+  --mode plan \
+  --file work/bugs/checkout-regression.md \
+  --out work/bugs/checkout-start.json
+```
+
+The advisory does not activate the workflow gate. A reviewed frozen plan must
+explicitly opt in before the gate can require reproduction, fingerprint,
+hypothesis ledger, regression proof and fix-impact receipts.
+
+See [Bug Forensics workflows](bug-forensics-workflows.md) for concrete defect
+search, regression, flaky failure and security-bug examples.
+
+## Coordinate cross-review
+
+For higher-risk research, planning or audits, create reviewer assignments with
+`review-mesh assign`, run the reviewers outside ALK, import their JSON output,
+then synthesize and check quorum.
+
+See [Review Mesh workflow scenarios](review-mesh-workflow.md) for concrete
+Codex, Claude and OpenCode/GLM examples. GLM is only an example model behind
+OpenCode; any configured model can be used by changing the host command.
+
+## Safety rules
+
+- Do not put secrets or private environment values in task files.
+- Prefer relative repository paths in portable packets.
+- Treat `review-mesh recommend` as advice until a frozen plan opts in.
+- Plugin installation alone is not proof that the ALK lifecycle ran.
+- Raw text, Markdown and imported plans never authorize implementation by
+  themselves.
+
+Use `agent-lifecycle start` for the common path. Advanced scripts may use the
+atomic `adapter task start`, `adapter run` and `adapter session resume`
+commands when they need direct control over one lifecycle primitive.
+
+
+### Bounded authoring and evidence examples
+
+- After a reviewed plan amendment, compile the new packet index in its fresh
+  artifact root before adoption. A new result path alone does not replace a
+  stale packet index. Keep old packets, failed receipts and expenses.
+- Link acceptance text to its canonical requirement. When scope changes, update
+  the linked acceptance before freeze. Structural preflight does not prove
+  that copied prose has the same meaning.
+- Check the actual lint selectors before promising unchanged test AST and clean
+  full-file lint. A production baseline does not exempt test files. Do not
+  invent a baseline or ignore after a failed check.
+- For long checks, retain selected test identities, starts and outcomes. Use an
+  existing bounded result tool or verbose output with the same selected tests.
+  A partial log does not prove that missing cases passed.
+- Keep process, validation, measurement and provider counts separate. Charge one
+  actual receipt once even when workers and the controller both reference it.
+  Preserve unknown usage and original clocks after retries or compaction.
+- A partial profile can justify a finite experiment; it cannot establish a
+  whole-workload speed result. Keep the frozen target until an explicit scope
+  decision. New same-job controls need their own bounded cost and deadline.
+- If a helper fails before its workload, preserve that distinction. An upstream
+  bug report is a hypothesis, not a confirmed local cause. Record exact platform
+  source bytes and any explicit CRLF materialization; never normalize a failed
+  proof silently or relabel it `PASS`.
+
+These are existing authoring and evidence rules, illustrated by Release 2.28.
+They require no additional audit round or native gate.
+
+## Bounded capability feasibility before freeze
+
+For a new route, portability claim or changed supported scenario, bind the claim
+to 1–3 actual entrypoints, the lower capability/authority boundary, one existing
+negative case and exact write owners before freeze. Distinguish refactoring
+existing behavior from adding a capability. UNIT_ONLY or mock evidence cannot
+establish an end-to-end route. If the existing boundary refuses the claimed
+scenario, obtain the user's scope decision before freeze: narrow the claim or
+include the required lower-layer changes and owners. If 1–3 entrypoints are
+insufficient, scope a bounded feasibility probe before freeze; do not repeat a
+whole audit. Record this in the existing specification/evidence plan and review;
+no extra audit stage, manifest schema or mandatory receipt framework is added.
+Structural plan verify/ownership/traceability checks do not prove semantic reachability.
+
+Example: Release 2.24 WS-03 keeps `resolve_declared_ledger_context`'s
+`declared-ledger-route-unsupported` refusal and the existing
+`test_unsupported_package_and_rebound_opening_refuse` negative case. Its typed
+usage-interface fixture is UNIT_ONLY. Isolating legacy paths is a refactor;
+a generic financial ledger or a newly qualified host route would be a separate
+capability requiring lower-layer ownership and end-to-end evidence.
+
+
+## Require a bounded decision record
+
+Select a research profile for the task and project; use DORA delivery measures
+in the external SDLC when they fit the project. Add `decisionResearch` to the
+draft workstream, with bounded frozen inputs and a separate record. Include a
+no-change alternative, criterion assessments and a finite declared budget.
+
+Follow the [research authoring sequence](research-workflow.md#check-decisions-before-task-start),
+then independent review, reviewed freeze and adoption. Existing task-start and
+continue perform one native freshness check. Reuse unchanged research; retain
+failed work, unknown usage and original clocks when correcting a refusal.
+`STOP` or `SPLIT` needs a revised reviewed plan. The record cannot waive the
+frozen quality floor or grant execution permission.
+
+<!-- deterministic-workflow-recipes:begin -->
+<a id="deterministic-workflow-recipes"></a>
+
+## Deterministic workflow recipes
+
+Use the CLI for structured files and already authorized local transitions.
+These operations do not need a model call. The calling agent still uses its
+own context and tokens. This guide makes no token or cost saving claim.
+
+## Create one canonical artifact
+
+Create `advisory.json` as a JSON object, for example:
+
+```json
+{"z": 2, "a": 1}
+```
+
+Run from the project root. Use a new output path:
+
+```sh
+agent-lifecycle task artifact-create --kind json --input advisory.json --out work/advisory.canonical.json --repository-root .
+```
+
+The command writes canonical JSON. It refuses missing or invalid input,
+duplicate keys, sensitive content, unsafe paths and existing output files.
+It does not review the content or grant execution authority.
+
+## Inspect the retained artifact
+
+```sh
+agent-lifecycle task artifact-inspect --input work/advisory.canonical.json --max-display-bytes 4096
+```
+
+Use the returned original path and digest to identify the artifact. The display
+can be redacted or truncated. Do not use the display as a replacement authority
+file. See [safe artifact authoring](../reference/safe-artifact-authoring.md).
+
+## Continue a bounded local chain
+
+First obtain the genuine reviewed plan, matching lock, native authorization,
+current state and a ready input bundle. The bundle uses the existing
+`agent-workflow-continuation-input-bundle.v1` contract. It must bind the same
+run, plan and source; do not copy another run's authority.
+
+Replace `REV` and `SOURCE_SHA` with the current full values. Select finite
+transition and I/O caps that fit the required inputs. This example allows four
+transitions and 65536 bytes; larger inputs require a separately chosen finite
+cap and must never be truncated to fit.
+
+```sh
+agent-lifecycle workflow continue \
+  --state work/run.state.json \
+  --manifest plans/feature/plan.manifest.json \
+  --lock plans/feature/plan.lock.json \
+  --expected-revision REV --source-revision SOURCE_SHA \
+  --reason "Apply authorized local steps" \
+  --input-bundle work/continuation-inputs.json \
+  --until-blocked --local-only --max-transitions 4 --max-io-bytes 65536 \
+  --apply --out work/continuation-receipt.json
+```
+
+The batch applies only existing validated transitions. `--local-only` stops
+before a model, host, worker, reviewer or operator action and does not consume
+a supplied authorization decision. Missing inputs, stale revisions, missing
+authority and exhausted caps stop the chain. Read the full receipt and its
+`stopReason`; a successful CLI exit alone does not mean the task is accepted.
+
+Keep the receipt, prior attempts, original deadline and expenses. Repair the
+named input or obtain the required external action before continuing. Never
+edit workflow state, invent an accepting review, or redispatch after an
+uncertain effect. See [workflow continuation](../reference/workflow-continuation.md)
+and the [lifecycle cookbook](lifecycle-cookbook.md).
+<!-- deterministic-workflow-recipes:end -->