Imported from ForgePlan/marketplace (plugins/fpl-skills/skills/wrap/SKILL.md). Install upstream with npx skills add ForgePlan/marketplace --skill wrap. Copyright stays with the author.

wrap — the wrap stage

The last stage. Answers is this cycle actually closed, or does it only look closed.

ADR-020 decided the shape of this in full — REFRESH as a depth-gated cycle-close node — and the decision reached the authoring guide and the routing map but never reached /forge-cycle: the stage sat at primary: "inline" and the word refresh appeared nowhere in the executor (marketplace#233). A decision the runner does not implement is a document.

Depth decides whether an artifact is created

Per ADR-020: at Standard and Tactical the reconciliation is already covered by /decay-watch, /forge-heal and the journal, so a REFRESH there would duplicate rather than add. Depth comes from the artifact — /forge-cycle Step 4.65 records it; if it is the untouched standard default, say so and treat it as Standard.

Через /forge-cycle эта стадия — только Deep+. wrap.depth_filter в матрице стоит deep+, а отфильтрованная стадия пропускается молча, поэтому строка Standard/Tactical выше относится к прямому вызову /wrap, а не к прогону цикла. Через цикл на Standard сверку делают /decay-watch и /forge-heal.

Process

Step 1 — Run the reconciliation, all four

forgeplan drift --json              # decisions whose affected files changed after they were made
forgeplan stale --json              # artifacts whose valid_until has expired
forgeplan health --json             # verdict, drafts, anomalies

plus /forge-heal for the findings ledger, and a link audit: for each artifact touched this cycle, does every link still resolve and point the right way.

drift and stale take no artifact ID — both are workspace-wide (verified against the CLI and the MCP schema; neither accepts a parameter). Run them whole and filter to this cycle's artifacts yourself; report the rest as context, not as this cycle's findings. Writing forgeplan drift <ID> gets you the whole workspace with an argument the tool ignored.

A clean drift is weak evidence, not proof. Its parser misses affected_files stored as a markdown table — measured on ADR-005, which returned changed_files: [] while git log showed three of its ten files had changed (Anomaly #18, forgeplan#293). When the cycle touched code, say what git log --since=<artifact created> shows alongside what drift returned.

Keep the output. Step 3 needs the text, not your summary of it.

Step 1b — Coverage, when the cycle touched code (marketplace#263)

forgeplan coverage        # decision coverage per code module; no --json form

Workspace-wide and honest about it: measured 2026-09-04 it reported "1% (2/185 modules)" because it scans everything under the workspace root, including directories no artifact governs. Filter to the modules this cycle touched and report whether they gained a governing decision; the workspace total is context for the ## Gaps closed section, never a finding against this cycle. A cycle that shipped code into a module with no decision coverage is worth one sentence in the REFRESH — it is tomorrow's drift.

Step 1c — Optional: capture the cycle's loose decisions

forgeplan_capture(decision="...") (MCP; needs a configured LLM provider) turns a decision made in conversation into a NOTE or ADR — use it for calls made during the cycle that never rose to an artifact. One capture per real decision; do not capture the cycle's summary, that is what the REFRESH is for. If the provider is absent the tool fails loudly — record that as "capture unavailable", not as "no loose decisions".

Step 2 — Below Deep, stop here

Report what the four returned and finish. No artifact. Say which depth was applied and why there is no REFRESH, so the absence reads as a decision rather than an omission.

Step 3 — At Deep and Critical, write the REFRESH

r = mcp__forgeplan__forgeplan_new(kind="refresh", title="Cycle close: <what the cycle did>")
mcp__forgeplan__forgeplan_update(id=r["id"], body=...)
mcp__forgeplan__forgeplan_link(source=r["id"], target="<EVID of this cycle>", relation="based_on")

Four sections, all required. guardian returns CONCERNS on a REFRESH missing any of them, or on a readiness assertion with no reconciliation output behind it (guardian Step 5, ADR-020 / CLAUDE.md G9) — not BLOCKER, so a rubber-stamp is caught by the reviewer chain rather than by the gate:

## Synced
<pasted output of forgeplan drift and forgeplan stale>

## Gaps closed
<pasted output of /forge-heal — what was found, what was fixed, what was left and why>

## Links updated
<the link audit: what was checked, what moved, what is still broken>

## Ready verdict
YES | NO — <one sentence>. On NO, name the blocker and who owns it.

YES or NO, those words. ADR-020 writes the verdict as "yes/no + blocker"; the requirement that the literal uppercase tokens appear is guardian's, at plugins/agents-pro/agents/guardian.md (the REFRESH structural-completeness row). A verdict written any other way (READY, ✅, "looks good") reads to that gate as a missing verdict, and the REFRESH comes back CONCERNS for a wording choice rather than a real gap.

Created via generic artifact-author. There is no kind-specialist for REFRESH and that is deliberate (ADR-020) — the reconciliation is done by the existing skills; this records their real output and the verdict.

Step 4 — The anti-false-green rule

Paste the actual tool output. Never a self-reported "all good."

ADR-020 calls this load-bearing, and it is the whole value of the node: a REFRESH that rubber-stamps manufactures a "cycle closed" signal that nothing else in the graph will contradict. The same discipline as completed meaning you re-ran the verification yourself, not an agent said it passed.

If a reconciliation step could not run — tool missing, command failed — write that, with the error. An unrun check recorded as clean is the failure this section exists to prevent.

CLAUDE.md records the limit honestly (G9): guardian can check that the four sections exist and that a verdict is present. It cannot tell pasted output from convincing prose. That half is on the author, whose identity is on the artifact.

Report

wrap PRD-081 — depth deep → REFRESH-004 created

  drift        2 of 9 affected files changed since creation — src/gate.ts, src/thresholds.ts
  stale        none
  health       healthy · 0 drafts · 2 medium anomalies (unrelated to this cycle)
  forge-heal   1 finding, AUTO-resolved (phase_mismatch on PRD-081)
  links        6 checked, 6 resolve, 1 direction corrected (EVID-233 informs → based_on)

  verdict      YES — the two drifted files are this cycle's own changes

At Standard the same block, ending no REFRESH — Standard depth, per ADR-020.

What this does not do

• It does not activate anything. The cycle's artifacts are activated at activate; this records that the dust settled.
• It does not fix what it finds. /forge-heal fixes; this runs it and records the result. A finding left open belongs in the REFRESH body with a reason, not silently repaired here.
• It cannot verify its own honesty. See Step 4.

Neighbours

/wrap runs the others for one cycle and records the result. It does not replace them: they sweep the workspace, this closes a cycle.

Reference: ADR-020 (REFRESH as depth-gated cycle-close), PRD-024 (wrap stage), CLAUDE.md G9, marketplace#233.