tailrocks-improve-reconcile

Reconciles the standalone plans/ backlog against current repository truth, optionally re-verifying every row, and updates only plans/README.md. Use this skill only when the user explicitly requests it

SKILL.md

GitHub

Improve reconcile

Use this skill

This skill synchronizes standalone plan status with repository truth. The sole writable output is plans/README.md. The index uses the shared status terms TODO, IN_PROGRESS, BLOCKED, STALE, and RETIRED, the same terms as the plan skill. Execution verdicts stay separate from index status.

Use this skill when the user passes only the optional repository path and modifiers. A retained sweep selector is invalid. Do not use this skill to edit plans, source, or roadmap items. Roadmap items belong to tailrocks-reconcile. Invoke this owner directly. No routing skill dispatches it.

Before you start

Use this skill only when the user explicitly requests it. A model must not select it from task similarity.

Obey the active user request first. If the request conflicts with a safety rule in this skill, stop. Report the conflict.

Before any action, read references/runtime-trust.md. Read references/reconciliation.md in steps 2 through 4. Resolve each relative link against the directory that holds this SKILL.md file.

--deep re-verifies every indexed row without sampling. --batch makes status selection deterministic and non-interactive. It preserves every evidence, command, compare-and-swap, and no-change rule. Neither modifier grants plan-body, source, roadmap, Git, or remote authority.

Procedure

  1. Bind the backlog. Record the root, the revision, the dirty state, index bytes, and every indexed plan. Refuse a roadmap item. tailrocks-reconcile owns those.

  2. Verify each row. Verify plan existence, stamped SHA, current evidence, changed paths, dependencies, and status claim. Default coverage can inspect rows whose state or evidence changed. --deep re-verifies every row without sampling. Run target commands only with explicit authority and every control below:

    • A read-only target
    • Frozen inputs
    • Disabled network
    • Scrubbed secrets
    • External cache
    • Bounded processes
    • TERM-then-KILL cleanup
    • A post hash.

    Otherwise mark proof unavailable.

  3. Mark truth only. Mark TODO, IN_PROGRESS, BLOCKED, STALE, or RETIRED. Missing evidence is not completion. A finding fixed independently becomes RETIRED with its fixing SHA. A defective plan becomes BLOCKED. It routes to a new tailrocks-improve-plan invocation. Never rewrite it here.

  4. Publish the index. Reject symlinked and escaping paths. Compare-and-swap one updated plans/README.md atomically. Roll back only owned bytes. On drift, leave bytes unchanged. Report the conflict.

Result

The terminal shows exactly one index update or a no-change receipt. No plan body, source, roadmap, branch, commit, remote, issue, or comment changed. Every row has current evidence.

Completion checks

Before the result is complete, make sure that each item below is true:

  • Only plans/README.md changed.
  • Every row uses a shared status term.
  • Every status change has current evidence.
  • No plan body or source byte changed.
  • Drift left bytes unchanged with a reported conflict.

References

Read these references at the stated times:

  • Read references/runtime-trust.md before any action for the trust rules.
  • Read references/reconciliation.md in steps 2 through 4 for row verification and index rules.