AGENTS.md git:20261026.10b7e34A

Immutable. This exact content never changes and is served at /api/v1/blob/82f89fe75b0792de.

# Cairn: start here / Agent 接手入口

This file applies to the entire repository. Any human or agent may contribute;
no private chat, particular model, account or orchestrator is required.

1. Read [coordination](docs/coordination/README.md), [current evidence](docs/coordination/CURRENT_STATE.md),
   [architecture constraints](CONTRIBUTING.md) and the live [hub](https://github.com/xcosmosbox/Cairn/issues/7).
2. Read the task Issue and its latest comments, dependencies and linked PRs.
   Refresh `main` and record its SHA; historical snapshots are not current runtime proof.
3. Follow [claim and handoff protocol](docs/coordination/PROTOCOL.md). Declare a unique
   session, branch, write scope and expiring lease. An assignee alone is insufficient.
   Claims are advisory, not an atomic distributed lock. Without acknowledgement,
   isolated branch/fork work is allowed under the provisional rules in the protocol.
4. Use one branch/worktree per task. Do not overwrite another agent's branch,
   force-push shared history, modify shared experiment inputs or silently broaden scope.
5. Preserve query-side read-only, zero-LLM, zero-embedding behavior, immutable Bundle
   identity, source provenance and human-curated edits. Record intentional contract changes
   in the relevant existing design document and task decision.
6. Report evidence at recoverable checkpoints and before stopping. Link exact commits,
   commands with exit status, tests, artifacts and remaining work in the Issue and PR.
   An unpushed local result or chat claim is not a handoff.
7. Validate changed behavior with affected tests and builds; normal CI remains enabled.
   Documentation changes need link/template validation. A local full race/verify suite,
   archive audit or recovery/scale rehearsal is not a prerequisite for a small experiment.
   Report blocked checks explicitly; preserve completed paid experiments as evidence.
8. Public reports must be understandable without private artifacts. Restricted data stays
   in its authorized destination; missing access is `unknown`, not `pass` or `absent`.

The collaboration system records work; it does not start agents, grant credentials,
approve model spending, or automatically accept/merge their work. Existing explicit
authorization continues to apply within its scope. No approval is needed merely to read,
research, reproduce free local checks, or prepare an isolated fix and PR.

Current direction: [v0.1 planning #107](https://github.com/xcosmosbox/Cairn/issues/107),
[version roadmap](ROADMAP.md) and [finite construction plan](docs/coordination/reports/2026-10-26-v01-build-completeness-plan.md).
The plan is not an implementation result. v0.1 requires meaningful source knowledge and
provenance to survive construction; finite failed-step retry/semantic revision precedes
partial completion for retryable/repairable failures. Raw retention, warning labels or
partial-result robustness cannot certify semantic completeness. This supersedes #105's
one-shot/no-semantic-revision direction while preserving its actual history and deletion
of unnecessary cost/time/archive/full609/recovery/repeated-approval gates. Use affected
checks and normal CI; missing usage is unknown and all intended outcomes remain visible.