reconcile

星标0

分支0

更新时间2026年7月22日 07:16

Fold a closed LCD work-item's acceptance criteria into the project's living current-state spec (SPEC.md) at closeout. Runs only when `living-spec: on` in lcd-conventions. Use at a work-item's finish line — after audit PASS (Deep) or when the JOURNAL flips to done (Standard) — to compact the frozen delta into the one document that says "what the system does now". Not for mid-work edits (that's lcd:refine).

安装

用 Codex 或 Claude 帮你安装复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。

来源

stid

stid/lcd

打开 GitHub 仓库查看创作者相关仓库

下载

LCD Reconcile (the compaction step)

The one-directional fold delta → living spec at closeout. A work-item spec.md (or a Standard JOURNAL's inline ACs) is the frozen delta — what one change added. SPEC.md is the living current-state index — what the system does now. Reconcile compacts the former into the latter so the next person scoping work reads one lean document instead of replaying every closed work-item. The append-only logs (DECISIONS.md, triage-log.md) are never rewritten by this — it only adds a supersede entry when a capability is replaced.

Distinct from lcd:refine: refine fires autonomously mid-work for small drift corrections within a work-item and explicitly avoids bulk rewrites. Reconcile fires once, at a defined lifecycle gate, and processes a completed delta. Different trigger, scope, and size.

Step 0 — gate on the flag (no-op when off)

Read .claude/rules/lcd-conventions.md's  block. If living-spec is not on, or there is no spec: key, do nothing and say so ("living-spec is off — skipping compaction"). This is the common case; most projects never turn it on, and Quick-lane work never reaches here. Resolve artifact-root and the spec: path (default <artifact-root>/SPEC.md). If SPEC.md doesn't exist yet, create it from ${CLAUDE_PLUGIN_ROOT}/templates/living-spec.md (fill <PROJECT_NAME>), then continue.

Step 1 — extract the delta's capabilities

Pick the source for slug <slug>:

Deep lane: <artifact-root>/work/<slug>/spec.md.
Standard lane: the work-item JOURNAL.md (its inline ACs use the identical **AC-N** (surfaces: <CSV>): <body> format, so the same parser reads it).

Run parse-acs.sh <source> (ships in the plugin bin/, on PATH). Three outcomes:

Rows returned → each <id>\t<surfaces>\t<body> row is a capability to fold.
Non-zero with "no AC declarations found" → the surface-less Standard case (a JOURNAL with no inline ACs). Fall back to one capability: surface none, body = the JOURNAL Goal line, no AC id. Do not treat this as an error.
Non-zero with a "malformed AC" / "invalid surface token" message → a real defect in the delta. Stop and report it; don't fold a malformed spec.

Step 2 — fold each capability into SPEC.md (new vs replace)

For each capability, read the matching surface section of SPEC.md and judge — this is the judgement that makes reconcile a skill, not a script:

NEW capability (the system couldn't do this before): add one terse row under its surface section — | <one-line current behaviour> | <AC ids> | D-NNN (<slug>) |. Keep it to a single line; if it wants a paragraph, the detail stays in the delta / the D-NNN, not here.
REPLACES an existing capability (this change altered behaviour already indexed): rewrite that row in place to the new behaviour, and append a supersede entry to DECISIONS.md (## D-NNN · <date> · <capability> now <new behaviour> with Status: active, and mark the prior decision superseded by D-NNN). SPEC.md stays current-only — never keep a graveyard of old rows there; the history lives in the append-only DECISIONS.md.

Stale markers. A · stale: <date> suffix on a row's Provenance means Quick-lane work changed that behaviour after the last fold (a marker, not a fold — the Quick lane's one allowed SPEC touch). When your fold covers that capability, the rewrite clears the marker (verify the new row text reflects the Quick change too, not just this work-item's delta). Leave markers on rows this fold doesn't touch.

Decay of backfill rows into rigor. A row whose provenance reads D-NNN (backfill) was seeded by onboard from code, not from audited ACs (a lower-confidence tier). When the work-item you're folding touches that capability, treat it as a REPLACES: rewrite the row with the now AC-pinned behaviour and set provenance to the work-item's real D-NNN (dropping the (backfill) marker) — this is how the provisional seed becomes verified over time. Don't leave a backfilled row in place when you have an AC-pinned truth for it.

Step 3 — provenance, stamp, log

Provenance points at the durable D-NNN, with the slug in parens as a breadcrumb (it may dangle after /lcd:tidy prunes the work-item — that's expected). If the work-item never earned a project-wide D-NNN, create one now recording the capability ("D-NNN · · shipped: ", scope project-wide) so SPEC has a stable anchor to cite.
Stamp Last compacted: <date> at the top of SPEC.md.
Log one line in the work-item JOURNAL.md LOG: <date> — reconciled: folded AC-1..N → SPEC.md/<surface>. No new telemetry shape — the closeout line in triage-log.md already records the finish.

The protocol (bounded by refinement-protocol)

Per the global refinement-protocol rule, even though closeout invoked you:

Restate the fold: "folding stats-surfaces: AC-1 (CLI) + AC-2 (HTTP) → 2 new SPEC rows".
List what stays unchanged — the SPEC sections / DECISIONS entries you are NOT touching.
Make the edits, then log.

Autonomy boundary — when to STOP and ask

A capability's behaviour contradicts an existing SPEC row in a way that isn't a clean supersede (two work-items disagree on current behaviour) → surface it; don't pick a winner silently.
The fold would need to delete a capability (a feature was removed). Confirm the removal is intended, then rewrite/remove the row and record a superseding D-NNN — never drop behaviour from the index without a recorded why.

Keep it light

Reconcile is a few SPEC rows + at most one supersede entry + a log line. If folding one work-item rewrites half of SPEC.md, the delta wasn't really one change — say so rather than forcing it.

同仓库更多 Skills

同仓库

onboard

stid/lcd

Make a project Lean Context Development aware. Use when a project has no .claude/rules/lcd-conventions.md yet (first-time LCD setup), when the user says "set up LCD here" / "make this project LCD-aware" / runs /lcd:onboard, or to refresh a stale MAP or conventions block. Generates a MAP.md of the actual code organization, seeds DECISIONS.md, records test-discovery/runner conventions, asks once where LCD artifacts live, and appends a triage pointer to CLAUDE.md. Works on new and existing projects. Retrofit-safe.

2026-07-220

recon

stid/lcd

Ground a LCD work-item in CURRENT best practice before deciding on an approach. Resolves up-to-date library/API/tooling guidance via Context7 and a targeted web search, in a subagent, and returns a few dated bullets. Use at the start of Standard/Deep work that touches a library, framework, SDK, API, or pattern whose current best practice you're not certain of — before writing the plan or choosing a dependency.

2026-07-220

redgreen-loop

stid/lcd

Autonomous build verb of LCD. Use when an LCD work-item (docs/lcd/work/<slug>/ or the conventions artifact-root) has failing tests to drive to green — after Deep-lane test-gen, or once a Standard-lane JOURNAL's STEPS reach failing tests. Iterates the test runner with bail, fixes the first failing test with minimum code, commits, repeats until green. Hard iteration cap and stuck escalation.

2026-07-220

refine

stid/lcd

Autonomously correct a LCD work-item's plan when reality drifts from it — a STEP/task that's now wrong, a task that needs splitting, a MAP zone/invariant that no longer matches the code, or stale knowledge. Updates the artifact AND logs the change without waiting to be asked. Use mid-work when you notice the recorded plan no longer matches what the work actually needs.

2026-07-220

triage

stid/lcd

Lean Context Development triage. Use when starting non-trivial new work in any LCD-onboarded project — a new feature, a multi-file change, a refactor that alters architecture, or anything you'd want a fresh session to pick up cold. Scores the work, picks a lane (Quick / Standard / Deep), states it in one line, and routes. Not for typo fixes, single-line edits, dependency bumps, formatting, or a bug fix whose cause is already known.

2026-07-220

name

reconcile

description

LCD Reconcile (the compaction step)

Step 0 — gate on the flag (no-op when off)

Step 1 — extract the delta's capabilities

Pick the source for slug <slug>:

Deep lane: <artifact-root>/work/<slug>/spec.md.
Standard lane: the work-item JOURNAL.md (its inline ACs use the identical **AC-N** (surfaces: <CSV>): <body> format, so the same parser reads it).

Run parse-acs.sh <source> (ships in the plugin bin/, on PATH). Three outcomes:

Rows returned → each <id>\t<surfaces>\t<body> row is a capability to fold.
Non-zero with "no AC declarations found" → the surface-less Standard case (a JOURNAL with no inline ACs). Fall back to one capability: surface none, body = the JOURNAL Goal line, no AC id. Do not treat this as an error.
Non-zero with a "malformed AC" / "invalid surface token" message → a real defect in the delta. Stop and report it; don't fold a malformed spec.

Step 2 — fold each capability into SPEC.md (new vs replace)

For each capability, read the matching surface section of SPEC.md and judge — this is the judgement that makes reconcile a skill, not a script:

NEW capability (the system couldn't do this before): add one terse row under its surface section — | <one-line current behaviour> | <AC ids> | D-NNN (<slug>) |. Keep it to a single line; if it wants a paragraph, the detail stays in the delta / the D-NNN, not here.
REPLACES an existing capability (this change altered behaviour already indexed): rewrite that row in place to the new behaviour, and append a supersede entry to DECISIONS.md (## D-NNN · <date> · <capability> now <new behaviour> with Status: active, and mark the prior decision superseded by D-NNN). SPEC.md stays current-only — never keep a graveyard of old rows there; the history lives in the append-only DECISIONS.md.

Step 3 — provenance, stamp, log

Provenance points at the durable D-NNN, with the slug in parens as a breadcrumb (it may dangle after /lcd:tidy prunes the work-item — that's expected). If the work-item never earned a project-wide D-NNN, create one now recording the capability ("D-NNN · · shipped: ", scope project-wide) so SPEC has a stable anchor to cite.
Stamp Last compacted: <date> at the top of SPEC.md.
Log one line in the work-item JOURNAL.md LOG: <date> — reconciled: folded AC-1..N → SPEC.md/<surface>. No new telemetry shape — the closeout line in triage-log.md already records the finish.

The protocol (bounded by refinement-protocol)

Per the global refinement-protocol rule, even though closeout invoked you:

Restate the fold: "folding stats-surfaces: AC-1 (CLI) + AC-2 (HTTP) → 2 new SPEC rows".
List what stays unchanged — the SPEC sections / DECISIONS entries you are NOT touching.
Make the edits, then log.

Autonomy boundary — when to STOP and ask

A capability's behaviour contradicts an existing SPEC row in a way that isn't a clean supersede (two work-items disagree on current behaviour) → surface it; don't pick a winner silently.
The fold would need to delete a capability (a feature was removed). Confirm the removal is intended, then rewrite/remove the row and record a superseding D-NNN — never drop behaviour from the index without a recorded why.

Keep it light

Reconcile is a few SPEC rows + at most one supersede entry + a log line. If folding one work-item rewrites half of SPEC.md, the delta wasn't really one change — say so rather than forcing it.