AGENTS.md git:20260928.224fb2dA

Immutable. This exact content never changes and is served at /api/v1/blob/493a996ed013a85b.

# Repository Instructions

## Governing Goal: A Stable, Simple, Consistent 2.0

Our overriding product goal is a 2.0 release people can reliably use. Apply
these priorities, in order, to every implementation and release decision:

1. **Stability first.** Without a stable system, features are not useful.
   Prioritize reliable end-to-end behavior, persistence, recovery, and clear
   failure states. Fix regressions and validate real workflows before adding
   scope; green health checks alone do not establish release readiness.
2. **Simplification second.** Reduce unnecessary controls, steps, duplicated
   surfaces, and concepts. Keep primary actions obvious; group secondary
   actions in clearly separated menus. Simplification must not hide errors,
   remove needed capabilities without an explicit decision, or weaken safety,
   authorization, observability, or testing.
3. **Consistency everywhere.** Once someone learns one part of ClawMax, that
   knowledge should transfer to the rest. Reuse established components,
   wording, layout, filters, Actions menus, and desktop/mobile interaction
   patterns instead of introducing one-off alternatives.

Prefer focused stability and usability improvements over new features on the
path to 2.0. Explain any necessary exception and its release impact before
expanding scope.

## Commit Messages

- Every commit subject must start with a lowercase Conventional Commit-style
  prefix followed by a colon and a space.
- Use the narrowest applicable prefix: `feat:`, `fix:`, `docs:`, `test:`,
  `refactor:`, `build:`, `ci:`, `chore:`, `release:`, or `merge:`.
- Keep the subject concise and describe the user-visible or engineering outcome.
- Do not publish an unprefixed commit subject. Check the proposed subject before
  committing, especially for release-candidate work pushed directly to `main`.
- Keep commits focused by concern: separate dependency/security remediation,
  product behavior, tests, documentation, and release metadata when they can
  be reviewed or reverted independently. Push each focused commit so CI
  identifies the exact change that introduced or fixed a failure.
- Commit and push as soon as one coherent, directly tested concern is complete.
  Do not accumulate multiple independently reviewable concerns in the working
  tree, and do not start the full suite, an image build, or release CI with
  completed implementation work still uncommitted. If later validation finds a
  defect, fix it in a new focused commit and push that checkpoint separately.

## Repository Ownership

- Keep public host contracts, generic dashboard components, public plugins, and
  public scoring logic in this repository.
- Keep first-party proprietary plugin manifests, suggested items,
  implementation details, and plugin-specific tests in the private
  `clawmax-plugins` monorepo. Partner-owned or customer-specific plugins may
  remain separate when they require different access controls. Never copy
  private source into this repository or the public image.
- Lifecycle and Review are the public product plugins in the current 2.0
  phase. Evals, Guardrails, and Optimize are private enterprise plugins. Treat
  `PLUGINS/test/plugin-*` as synthetic contract fixtures only;
  they must not contain or become copies of private product implementations.
- Keep every other plugin private unless its ownership is explicitly changed.
- Plugin enablement must be manageable from the PLUGINS sidebar as well as
  deployment configuration. Persist an explicit instance selection, including
  the valid empty selection, and never expose synthetic contract fixtures in
  the manager.
  Public documentation may describe the generic host boundary, but must not
  enumerate private plugin items, rules, prompts, schemas, or implementation
  details.
- A plugin can represent any ClawMax extension domain. Do not design generic
  plugin APIs or UI around assumptions that all plugins are guardrails or evals.

## Product Language And State

- Treat prompt-readiness scores and generated-artifact quality scores as
  separate claims. Prompt scoring must explain the rubric, update as the user
  edits, remain advisory rather than blocking generation, and never imply that
  an untested generated artifact is correct or safe.
- Keep baseline prompt scoring local and token-free. Do not persist or export
  raw prompt text as scoring feedback; collect only the minimum disclosed
  metadata unless the user explicitly opts into a separate data-use contract.
- Reserve `template` for ClawMax agent, organization, and workflow templates.
  Plugin starter content should be presented as `Suggested`, `Use`, or another
  domain-specific action rather than `Use Template`.
- Distinguish suggested plugin items from selected workspace items. Selected
  items must be editable and their active, running, completed, failed, or
  blocked state must remain visible.
- User-visible operations must persist lifecycle evidence appropriate to the
  domain, such as run history, activation history, scores, progress, targets,
  notifications, notes, or evidence. Do not expose an action without a way to
  see what happened afterward.
- Treat the Agents page as the dashboard layout contract for plugin tabs:
  preserve its header hierarchy, view-control placement, action menus, filters,
  responsive stacking, and distinct grid, detail, list, and relationship
  presentations. View controls must affect every visible collection they imply,
  including suggested items, and each item must expose inspectable details.
- Keep release review files separated by release, but put only checks requiring
  independent human judgment or an external environment in the active reviewer
  queue. Unit, integration, browser, source-contract, CI, version, health,
  architecture, and catalog checks are engineering-owned release evidence, not
  reviewer tasks. Every current Review item must state `reviewReason` as either
  `human-judgment` or `external-environment`. Preserve superseded criteria,
  results, notes, evidence, and verifier history in Archived; do not carry
  unfinished old-RC checklists forward as tester work.

## Tests And Release Validation

- Treat 2,000 tokens per local calendar day as the user-notification threshold
  for repository work. When session-visible usage reaches that threshold, tell
  the user once and continue only as directed. This cannot include usage from
  unrelated threads or tools that do not expose token accounting.
- Minimize token overhead by using focused file reads, bounded command output,
  sparse CI polling, long waits for known-duration jobs, and concise progress
  summaries. Do not trade away required implementation or test coverage merely
  to reduce reported usage.

- Treat transcript, prompt, conversation, and user-activity export as sensitive
  data egress. It must be disabled by default, explicitly consented per affected
  user, scoped to a named destination and purpose, visibly indicated while
  active, redacted before durable storage, revocable immediately, and delivered
  asynchronously so partner failure cannot block normal product behavior.
- Configuring a partner is not user consent. Never silently reuse consent when
  the destination, purpose, scopes, identity fields, or disclosure version
  changes, and never expose a general transcript-reading capability to plugins.
- Add or update focused tests for every behavior change, including error,
  persistence, and responsive presentation paths where applicable. Improve
  coverage with each change rather than relying only on existing tests.
- Run TypeScript and the directly affected unit/contract suites while
  iterating. Before an RC, run the complete integration, validation, and
  coverage suite.
- When cutting an RC, align `CLAWMAX_VERSION` in the ignored local
  `SYSTEM/dashboard/.env` to the exact candidate being built, restart the local
  dashboard, and verify both the visible version and `/api/system`. For example:
  `perl -0pi -e 's/^CLAWMAX_VERSION=.*/CLAWMAX_VERSION=2.0.0-test-rcN/m' SYSTEM/dashboard/.env`.
  Never carry an older RC value into a new RC or stable release. The test-image
  workflow supplies this value for containers, but local source checkouts
  otherwise fall back to the latest Git tag plus the current short SHA.
- Visually audit changed pages, dialogs, pop-ups, cards, lists, and graphical
  views at desktop and mobile widths. Check long text, scrolling, sticky
  actions, progress states, and empty/error states.
- For public/private combined RCs, verify amd64 and arm64 image builds, registry
  smoke tests, packaged version identity, plugin discovery, restart persistence,
  and the public/private source boundary before reporting the release ready.
- When creating an RC image, dispatch the public `Test Container Image`
  workflow from the candidate ref first, then dispatch `Private ClawMax Plugins
  Image` in `clawmax-plugins` with the exact same public `base_tag` and private
  `image_tag`. Record both CI links and do not call the public image complete
  until the matching combined image has passed its validation and smoke jobs.
- If a combined-image smoke job fails after the image build and plugin contract
  checks pass, inspect the failed step and distinguish a dashboard readiness
  probe/connection-reset flake from a packaging or plugin failure. The image
  may be usable, but report the smoke result as unresolved and do not silently
  mark the combined RC fully green; rerun the failed smoke jobs once before
  deferring readiness-test hardening.
- Record the observed or documented duration before starting a long-running CI,
  image, deployment, or test job. Do not continuously poll it.
- Check once near half the expected duration to catch an early failure. If the
  job is still healthy, check next at the expected completion time plus one
  minute. Only use additional sparse checks when the job exceeds that window or
  a known failure needs investigation.
- When waiting on local test processes, use long yielded waits and small output
  budgets. Report the failure or final summary instead of repeatedly relaying
  routine progress output.
- Before handoff, commit and push every owning repository with an approved
  lowercase prefix, record the relevant SHAs and CI links, and verify that no
  task-created changes remain uncommitted.
- Security claims require reproducible evidence. Record dependency counts,
  command results, residual severity, owner, and deadline; do not describe an
  initial dependency/configuration baseline as the complete security audit.

## Documentation Maintenance

- Every Sunday PM or Monday AM, write a concise weekly recap in
  `SYSTEM/docs/recap/YYYY-MM-DD_WEEKLY_RECAP.md`, using the Monday that starts
  the reviewed week. Keep it to two or three readable pages. Summarize verified
  outcomes, what went well (WWR), what went wrong (WWW), lessons, open evidence,
  and a short, explicitly soft plan for the coming week. Link separate event or
  hackathon recaps instead of repeating them. Check commits, release evidence,
  active backlog, and known issues; distinguish source completion, CI/image
  validation, deployment, and user acceptance. Do not infer current runtime
  state from an old handoff.
- At each month end, add `SYSTEM/docs/recap/YYYY-MM_MONTHLY_RECAP.md` to
  synthesize that month's weekly recaps, trend in outcomes and WWR/WWW, unresolved
  patterns, and a few process changes for the next month. Keep the recap index
  ordered by date, newest first. Treat plans as suggestions that can change with
  meetings, testing, and new priorities.

- Audit active documentation when a stable release is promoted, a major release
  changes phase, or release feedback materially changes the roadmap.
- Keep `README.md`, `CHANGELOG.md`, `SYSTEM/docs/STATUS.md`,
  `SYSTEM/docs/KNOWN_ISSUES.md`, `SYSTEM/docs/BACKLOG.md`, and the documentation
  index aligned on the stable version, active development line, and release
  boundary.
- Move completed or superseded release plans and validation checklists into the
  appropriate archive. Preserve useful history, but do not leave shipped work
  presented as an active plan or backlog item.
- Update archive indexes and links as part of the same change. Do not close or
  archive a partially completed issue solely because one implementation phase
  shipped; retain the remaining concrete work in the active backlog.