artshelf 0.6.0 → 0.7.0

package/docs/agent-usage.md

@@ -1,480 +1,53 @@
 # Agent Usage
 
 Agents that support local skills can copy or reference
-[`skills/artshelf/SKILL.md`](../skills/artshelf/SKILL.md). The public docs site at
+the whole [`skills/artshelf`](../skills/artshelf) directory. The public docs site at
 <https://calvinnwq.github.io/artshelf/> explains the same contract in browsable
 form.
 
-Artshelf works best when agents register artifacts at creation time, while the
-reason is still fresh. Do not wait for a cleanup pass to infer intent from file
-age or path names.
+Artshelf is meant to be operated by agents through a small skill contract and
+read-only scheduled reviews. Agents remember and summarize. Humans approve
+mutation.
 
-## When To Register
+## Workflow Summary
 
-Treat Artshelf as a finalization trigger, not an optional cleanup habit. Before an
-agent reports a task as done, it must check whether the task created, copied,
-exported, quarantined, backed up, or preserved any non-source file or directory
-that may outlive the current command.
+Use Artshelf as a four-stage loop around agent work:
 
-Call `artshelf put` immediately after creating an eligible artifact:
+1. **Create**: register durable temp artifacts with lookup-before-put and
+   `artshelf put`, or state the skip reason.
+2. **Monitor**: run read-only checks for registry health, due records, missing
+   paths, and trash state.
+3. **Review**: turn raw output into an `ArtshelfReviewReport` decision packet
+   with exact approval targets.
+4. **Clean**: execute approved plans, clear trash only from a separate
+   reviewed purge plan, resolve confirmed ids, then verify quiet.
 
-- config backups
-- rollback copies
-- quarantine folders
-- debug output directories
-- generated reports
-- temporary repo artifacts
-- long-running task evidence
-- copied files kept so a reviewer can inspect them later
+This maps to the product loop: **Create -> Monitor -> Review -> Clean**.
 
-Do not register normal source files, committed documentation, package build
-outputs that can be regenerated cheaply, or dependency caches.
+## Child Pages
 
-If an eligible artifact is not registered, the agent should state why. Common
-valid reasons are: the artifact is source-controlled, it is a cheap
-regeneratable cache/build output, it contains secrets, it belongs to another
-durable artifact system, or the user explicitly asked not to retain it.
+The browsable docs split the workflow into focused child pages:
 
-## Command Shape
+- [Create](agent-create.html): registration triggers, lookup-before-put, skip
+  reasons, and Artshelf id footnotes.
+- [Monitor](agent-monitor.html): registry health, scheduled read-only checks,
+  and preview plans.
+- [Review](agent-review.html): decision packet schema, classifications, and
+  exact approval wording.
+- [Clean](agent-clean.html): approval-only cleanup, trash purge, resolve,
+  receipts, and verify-quiet checks.
 
-Use the installed CLI when available:
+## Operating Principles
 
-```bash
-artshelf put <path> --reason "<why this exists>" --ttl 3d --kind run-artifact --cleanup review --owner agent
-```
+- Agents remember with the portable skill.
+- Scheduled checks read and report only.
+- Review output is a decision packet, not raw counts.
+- Approval names the exact ledger, plan id, or record ids.
+- Every approved action ends with a read-only verification.
 
-If Artshelf is not installed, use the package-manager install path when
-available:
+## Portable Skill
 
-```bash
-npm install -g artshelf
-artshelf --version
-artshelf doctor
-```
-
-With pnpm:
-
-```bash
-pnpm add -g artshelf
-artshelf --version
-artshelf doctor
-```
-
-For source installs, do not assume a repo path. Ask where the user wants the
-Artshelf repo cloned, then use the supported local path:
-
-```bash
-git clone https://github.com/calvinnwq/artshelf.git "$ARTSHELF_REPO"
-cd "$ARTSHELF_REPO"
-corepack enable
-pnpm install --frozen-lockfile
-pnpm run build
-npm link
-artshelf --version
-artshelf doctor
-```
-
-Do not create a custom shim. Use the published package or `npm link` from a
-local source checkout.
-
-Useful defaults for agents:
-
-- `--kind scratch` for temporary working directories.
-- `--kind backup` for rollback copies.
-- `--kind run-artifact` for logs, reports, and generated evidence.
-- `--kind quarantine` for files isolated because they may be unsafe or wrong.
-- `--cleanup review` when a human or future agent should inspect before moving.
-- `--cleanup trash` only when the artifact is definitely disposable after the
-  retention window.
-- `--owner <agent-or-runtime>` should name the agent, tool, CI job, or human
-  process that created the artifact.
-- `--label <project>` and `--label <task-id>` when the artifact relates to a
-  repo, PR, issue, workflow id, or run id.
-
-Use `--json` when another tool needs to capture the Artshelf entry id.
-When `artshelf put --json` succeeds, include a deterministic Artshelf footnote in
-the same handoff, status, final response, or run summary that mentions the
-artifact:
-
-```text
-Artshelf footnote: registered <artifact-path> as <artshelf-id>; reason: <short reason>; due: <YYYY-MM-DD|manual-review>; cleanup=<cleanup-mode>.
-```
-
-## Idempotent Lookup
-
-Integrations should check the ledger before creating another record for the
-same artifact. Use `find` and `get` for read-only lookup:
-
-```bash
-artshelf find --path <path> --owner <agent-or-runtime> --label <task-or-run-id> --json
-artshelf get <id> --json
-```
-
-`find` requires at least one selector: `--path`, `--owner`, `--label`, or
-`--status`. Multiple labels are an all-label match. If `find` returns an
-existing record, report that Artshelf id instead of calling `put` again. If it
-returns no entries, call `put` and record the new id.
-
-## Ledger Registry
-
-Artshelf keeps a user-level registry at `~/.artshelf/ledgers.json` so one CLI can
-review all known ledgers without moving project records into one global file.
-`put` registers the ledger it writes to. Register existing ledgers explicitly
-when adopting Artshelf for an existing project:
-
-```bash
-artshelf ledgers add --ledger <repo>/.artshelf/ledger.jsonl --name <project> --scope repo
-artshelf ledgers list --json
-```
-
-Renamed installs before `0.5.0` used `.shelf` paths. For migration, copy the old
-ledger directories into `.artshelf`, rewrite registry entries, validate the new
-registry, and keep the `.shelf` copies until rollback is no longer needed.
-
-`artshelf ledgers list --json` validates each registered ledger and reports
-ok/missing/invalid status with entry and warning/error counts, so agents can
-detect stale registry entries without a separate validate pass. Add `--plain`
-for a fast listing that skips validation.
-
-Use the registry for read-only review and discovery:
-
-```bash
-artshelf review --all --json
-artshelf status --all --json
-artshelf due --all --json
-artshelf find --all --owner <agent-or-runtime> --json
-artshelf trash list --all --json
-```
-
-`artshelf review --all --json` returns an aggregate triage summary (affected
-ledgers, due, manual-review, missing-path, executable, and skipped counts plus
-preview plan ids) alongside the per-ledger detail, and states the next safe
-action.
-
-Use global cleanup dry-run when you want Artshelf to write cleanup plans for
-registered ledgers with cleanup entries, without moving files:
-
-```bash
-artshelf cleanup --dry-run --all --json
-```
-
-Do not use `--all` as permission to mutate files. Cleanup execution remains
-ledger-specific and requires a reviewed plan id for that ledger.
-If the executable cleanup entries have not changed, dry-run reuses the existing
-plan id and refreshes the same plan file instead of creating duplicate plans.
-
-## Daily Review Workflow
-
-Use this flow when a scheduled review, recurring task, or user request reports
-Artshelf cleanup attention:
-
-1. Register artifacts early during work, or state why an eligible artifact was
-   skipped.
-2. Review state with read-only commands first:
-   `artshelf ledgers list --json`, `artshelf review --all --json`, and
-   `artshelf trash list --all --json`; for old trash on a selected ledger, run
-   `artshelf trash purge --older-than 7d --dry-run --ledger <ledger-path> --json`.
-3. Present a decision packet instead of raw counts. Include registry health,
-   affected ledgers, due/manual-review/missing-path counts, executable entries,
-   skipped entries, refused entries, trashed record counts and ages, purge
-   dry-run plan ids/skipped entries, and the next safe action.
-4. Classify each candidate:
-   - `trash-safe`: disposable after the reviewed plan moves it into Artshelf trash.
-   - `needs-human-review`: `cleanup=review`, evidence, backups, reports, or
-     anything that should be inspected before closing.
-   - `resolve-candidate`: already handled, missing, or no longer needed; use
-     `artshelf resolve` only after confirmation.
-   - `registry-problem`: stale, missing, or invalid ledger; fix registry health
-     before touching artifacts.
-5. If cleanup execution is appropriate, generate or reuse a dry-run plan, then
-   ask for explicit approval naming the ledger path and reviewed plan id.
-6. For trashed records, require a separate reviewed purge plan before physical
-   deletion:
-
-```bash
-artshelf trash list --ledger <ledger-path>
-artshelf trash purge --older-than 7d --dry-run --ledger <ledger-path> --json
-artshelf trash purge --execute --plan-id <purge-plan-id> --ledger <ledger-path> --json
-```
-
-7. After approved cleanup execute, trash purge, or resolve, verify quiet with
-   `artshelf review --all --json`, plus `artshelf trash list --ledger <ledger-path> --json`
-   and purge receipt evidence after purge, or explain what remains.
-
-### Review Plan Report Schema
-
-When a dry-run creates or reuses a cleanup or trash purge plan, surface the plan
-in a compact human-readable report. The report should let the user approve, ask
-for changes, or request alternatives without opening the plan file.
-
-For deterministic agent integrations, construct an `ArtshelfReviewReport` JSON
-object first, then render it to text. Use
-[`schemas/artshelf-review-report.schema.json`](schemas/artshelf-review-report.schema.json)
-for the packet shape and
-[`examples/artshelf-review-report.json`](examples/artshelf-review-report.json)
-as the canonical example. The schema locks report structure; the CLI output and
-approval rules still define cleanup safety.
-
-Render the JSON packet as a compact decision card using `decisionSummary` and
-`decisionGroups`. Lead with counts, then show exactly what is ready for approval
-and what needs review first. Emojis are encouraged when the host renders them
-well because they make the groups scannable; omit them only for plain-text or
-accessibility-constrained surfaces.
-
-```text
-Artshelf daily review
-Status: <ok|attention needed>; registry <ok|attention>
-
-✅ Ready for approval: <n>
-👀 Needs review first: <n>
-⚠️ Blocked: <n>
-
-Recommendation
-<one short sentence with the next safest action>.
-
-Ready for approval
-1. <short item label>
-   Why: <short reason>
-   Action: <what approval will do>
-   approve artshelf cleanup ledger <ledger-path> plan <plan-id>
-
-2. <short item label>
-   Why: <short reason>
-   Action: <ledger-only update, no file changes>
-   approve artshelf resolve missing ledger <ledger-path> ids <id...>
-
-Needs review first
-1. <short item label>
-   Why: <short reason>
-   Suggested next step: inspect path, then choose keep, change retention, resolve, or clean up later
-   Path: <path>
-
-Blocked
-<none, or the registry/refused/missing-path blocker and next repair step>
-
-Safety
-Dry-run only. No execute, resolve, or delete ran.
-```
-
-Keep the full `ArtshelfReviewReport` JSON as the audit packet and include it as
-an attachment, linked file, or expandable detail when the host supports that.
-Do not paste the whole packet into chat unless the user asks for it. For long
-plans, show only the first 3 to 5 decisions under each visible group, then state
-the hidden count by group and classification. Do not hide refused,
-registry-problem, or missing-path items.
-
-If the host supports buttons, menus, or other interactive controls, they should
-emit exact text commands such as the approval targets below. Always include the
-exact approval target in the message body as a fallback for clients where those
-controls do not render.
-
-Approval wording should be exact:
-
-```text
-approve artshelf cleanup ledger <ledger-path> plan <plan-id>
-approve artshelf trash purge ledger <ledger-path> plan <purge-plan-id>
-approve artshelf resolve missing ledger <ledger-path> ids <id...>
-```
-
-Never execute from a read-only preview id. Never generate a fresh plan and
-execute it in the same step. `trash` moves artifacts into Artshelf trash; physical
-delete requires a separate reviewed trash purge plan.
-
-## Reasons
-
-Write reasons as small audit notes. A good reason lets a future agent decide
-whether the artifact still matters without replaying the whole conversation.
-
-Good:
-
-```text
-backup before rewriting migration order for issue-123
-```
-
-Weak:
-
-```text
-backup
-```
-
-Include the source of authority when useful: PR number, issue id, workflow id,
-command, failing check, or user request.
-
-## Reporting Artshelf IDs
-
-After registration, include the Artshelf id anywhere future cleanup context will be
-read:
-
-- handoff notes
-- PR comments
-- issue comments
-- daily memory
-- task run summaries
-- incident or debugging notes
-
-Example:
-
-```text
-Temporary parser output registered in Artshelf as shf_20260601_182800_ab12.
-Retain until 2026-06-04; cleanup=review.
-```
-
-## Cleanup Boundary
-
-Agents may run non-destructive cleanup checks:
-
-```bash
-artshelf validate --json
-artshelf validate --all --json
-artshelf due --json
-artshelf due --all --json
-artshelf review --all --json
-```
-
-Cleanup dry-run is safe to run. It writes plan files for later review only when
-there are executable cleanup entries:
-
-```bash
-artshelf cleanup --dry-run --json
-artshelf cleanup --dry-run --all --json
-```
-
-Cleanup execution is approval-only: no daemon, no auto-execute, no global
-execute, and no fresh-plan-then-execute shortcut. Agents must not run this
-without explicit human approval:
-
-```bash
-artshelf cleanup --execute --plan-id <id>
-```
-
-Approval should name the plan id. Do not generate a fresh plan and execute it in
-the same breath. Review the dry-run first, then execute the reviewed plan id.
-After cleanup execution, agents may inspect trash and create a purge dry-run for
-review:
-
-```bash
-artshelf trash list --ledger <ledger-path> --json
-artshelf trash purge --older-than 7d --dry-run --ledger <ledger-path> --json
-```
-
-Trash purge execution is separately approval-only and must name the ledger and
-reviewed purge plan id:
-
-```bash
-artshelf trash purge --execute --plan-id <purge-plan-id> --ledger <ledger-path> --json
-```
-
-No-op dry-runs report `not-created` and do not write plan files. When dry-run or
-execute creates plan or receipt artifacts, Artshelf records those artifacts in the
-ledger as `owner=artshelf`.
-
-`cleanup=delete` stays refused; execution records `cleanup-refused` instead
-of silently deleting files. Physical deletion requires a separate reviewed trash
-purge plan.
-
-Execution writes a receipt and updates touched ledger records to `trashed`,
-`review-required`, or `cleanup-refused`, so handled artifacts stop reappearing in
-future due and dry-run cleanup output.
-
-Agents may mark a ledger record manually resolved when the user confirms the
-artifact was inspected, is already missing, or is no longer needed:
-
-```bash
-artshelf resolve <id> --status resolved --reason <text>
-```
-
-Use a specific reason. `resolve` only updates the ledger; it does not move or
-delete files. For batches of missing-path records, ask for approval that names
-the exact ledger and ids:
-
-```text
-approve artshelf resolve missing ledger <ledger-path> ids <id...>
-```
-
-After resolution, verify with `artshelf review --all --json` and report whether
-the review is quiet or what remains. Resolved records stop reappearing in future
-due and dry-run cleanup output while remaining visible in
-`artshelf list --status resolved`.
-
-## Scheduled Review
-
-Agents may schedule routine Artshelf reviews for stale artifacts through their host
-runtime, such as an agent cron, CI job, or recurring task. Keep the scheduled
-job non-destructive:
-
-```bash
-artshelf validate --json
-artshelf due --json
-artshelf review --all --json
-```
-
-Read-only health and dashboard checks are also safe to schedule. Run
-`artshelf review --all --json` for aggregate triage (`summary` and `nextAction`),
-`artshelf doctor --json` to catch a broken or stale registry before relying on
-cleanup planning, and `artshelf status --all --json` for a compact cron summary:
-
-```bash
-artshelf doctor --json
-artshelf status --all --json
-```
-
-Scheduled cleanup and trash purge dry-runs may write plan files for later review
-when entries exist, but must not move or delete files:
-
-```bash
-artshelf cleanup --dry-run --json
-artshelf trash list --ledger <ledger-path> --json
-artshelf trash list --all --json
-artshelf trash purge --older-than 7d --dry-run --ledger <ledger-path> --json
-```
-
-The scheduled job should report the ledger path, due/manual-review/missing-path
-counts, cleanup dry-run plan id, executable entries, skipped entries, and refused
-entries. When reporting trash, `artshelf trash list --all --json` may discover trashed
-records across registered ledgers. Include trashed record counts and target ages;
-run purge dry-runs only for an explicit ledger and report any plan id, matching
-entries, and skipped entries. It should be
-quiet when nothing needs attention unless the user asked for a regular summary.
-
-Use explicit ledger paths when scheduling checks for a known project or user
-ledger. Do not scan arbitrary filesystem locations looking for ledgers unless
-the user has opted into that discovery scope.
-
-Scheduled jobs must not run cleanup execution or trash purge execution. They
-may only dry-run and report plans for later human review:
-
-```bash
-artshelf cleanup --execute --plan-id <id>
-artshelf trash purge --execute --plan-id <id>
-```
-
-Any later execution requires a human to review the dry-run output and approve
-that specific plan id.
-
-## Handoff Pattern
-
-When a task creates registered artifacts, add a short section like this:
-
-```text
-Artshelf artifacts:
-- shf_20260601_182800_ab12: /tmp/parser-output, debug evidence for issue-123,
-  retain until 2026-06-04, cleanup=review
-```
-
-If there are no eligible artifacts, say nothing. If eligible artifacts were
-skipped instead of registered, include the brief skip reason from the completion
-checklist. Do not invent Artshelf entries after the fact just to make a handoff look
-tidy.
-
-## Completion Checklist
-
-Before final response or handoff, agents should review their own file actions
-from the current task:
-
-1. Did I create, copy, export, quarantine, back up, or preserve any non-source
-   file or directory?
-2. Will any of those paths outlive this command?
-3. If yes, did I either register them with Artshelf or record a clear skip reason?
-
-Do not call work done while known eligible artifacts are neither registered nor
-explicitly skipped.
+The repo ships a portable skill at
+[`skills/artshelf`](../skills/artshelf). Agents that support local skills can
+copy or reference the whole directory directly, including the bundled
+`scripts/render-review-report.mjs` renderer plus schema and example copies.