main merged

.agent/evidence/code-review-findings.md

@@ -0,0 +1,31 @@
+# Code review record — cumulative diff `main...chore/skill-setup-and-pstack-gates`
+
+Two independent fresh-context reviewers: Standards/Correctness axis + Spec/Validator axis.
+Spec validator independently re-ran AC1–AC6 commands; all reproduced green
+(44-skill portability exit 0 · validate-plugins exit 0 · references exit 0 ·
+bun test 52/52 · audit spot-check matched · evidence file real).
+
+## Findings and dispositions
+
+| # | Sev | Finding | Disposition |
+|---|-----|---------|-------------|
+| 1 | MAJOR | Four triage labels missing from tracker vs docs claims | **Resolved by verification** — `gh label list` confirms needs-triage/needs-info/ready-for-agent/ready-for-human/wontfix all exist; earlier "missing" was a truncated listing on both sides |
+| 2 | MINOR | CI installed ajv unpinned via `npm install --no-save`, ignoring lockfile | **Fixed** — workflow now `bun install --frozen-lockfile` |
+| 3 | MINOR | New references gate unwired (no alias, no CI, absent from CLAUDE.md gates) | **Fixed** — alias added; CI runs all three gates; path filter includes `pstack/**`; CLAUDE.md updated |
+| 4 | MINOR | `decodeURIComponent` throws on malformed escapes, crashing whole run | **Fixed** — try/catch, reported as unresolved reference |
+| 5 | NIT | Commit message says "pinning"; ranges are caret | **Accepted as-is** — bun.lock is the pin; not rewriting pushed-stack history for prose |
+| 6 | NIT | Checker edge cases (Windows seps, reference-style links, fences, hyphen placeholders) | **Partially fixed** — hyphenated placeholders exempted; rest documented as known limits in header comment |
+| 7 | NIT | `skillRootOf` bogus probe when md sits directly at skills root | **Won't fix** — second probe covers it; no such file exists (44 skill dirs) |
+| 8 | NIT | `.scratch` ticket location; two doc commits could be one | **Won't fix** — `.scratch/<feature>/` is the setup-matt-pocock canonical convention; history preserved |
+| D1 | Drift | Gate-5 audit's re-audit grep read like a failing gate | **Fixed** — reworded to classify expected absences |
+
+## Unrelated-changes check
+
+CLAUDE.md + docs/agents/* (~160 lines) sit outside the SPEC Objective text but are
+covered by recorded Assumption 1 (user-mandated skill-setup scope, branch name).
+Retained with this note as reviewer sign-off context.
+
+## Parked
+
+PORTABILITY gate 4 remainder (Cursor GUI + second non-Cursor host clean-session smoke):
+requires human-launched sessions. Unblock recorded in SPEC Assumption 2 and issue #6.

.agent/evidence/gate5-fallback-audit.md

@@ -0,0 +1,64 @@
+# Gate 5 audit — bounded fallback for unavailable capabilities
+
+Scope: all 44 skills under `pstack/skills/` (SKILL.md plus references/playbooks).
+Contract: `pstack/PORTABILITY.md` portable core — "degrade explicitly when a
+capability is unavailable"; acceptance gate 5.
+
+> **Revision 2 (2026-08-22, post-merge review of PR #8).** Revision 1 claimed
+> "15/44 direct hits" and "44/44 audited after fix; zero open gaps." Both claims
+> were wrong or unreproducible: the stated grep yields **13** direct hits on
+> SKILL.md files; three skills (`bro`, `tdd`, `unslop`) were absent from the
+> verdict table; one non-hitting `principle-*` skill was miscounted (19 claimed,
+> 20 actual); and `create-verification-skill` was labeled COVERED despite having
+> no capability-degradation clause for its core drive dependency. This revision
+> replaces those claims with reproducible counts and fixes both gaps found.
+
+## Method (reproducible)
+
+1. Direct-hit grep over SKILL.md files only:
+
+   ```sh
+   grep -rilE 'unavailable|fallback|degrade|not available' pstack/skills/*/SKILL.md
+   ```
+
+   → **13 skills**: arena, automate-me, figure-it-out, how, interrogate,
+   no-comments, principle-guard-the-context-window, recall, reflect,
+   setup-pstack, show-me-your-work, swarm, why.
+
+   (An all-markdown variant returns 34 files; per-skill classification below is
+   authoritative because coverage lives in whichever file binds the capability.)
+
+2. For the remaining **31** skills: sweep capability terms
+   (`subagent|delegate|spawn|host|tracker|browser|playwright|network`) per skill
+   directory, then read the cited lines to separate real bindings from prose.
+
+## Verdicts — 44/44 classified (13 direct + 31 swept)
+
+| Skill / group | n | Verdict | Evidence |
+|---|---|---|---|
+| Direct hits (listed above) | 13 | COVERED | explicit degradation language in SKILL.md |
+| architect | 1 | COVERED | SKILL.md L16 "If no tracker is available, keep the same short checklist"; L36 "Otherwise run independent candidates on the parent model … Never invent a model identifier" |
+| blast-radius | 1 | COVERED | SKILL.md step 6: "Use distinct confirmed models when available; otherwise inherit the parent model and do not claim cross-model evidence." |
+| poteto-mode | 1 | COVERED | references/plan.md L5 tracker fallback; L26–27 role/model fallbacks ("Otherwise launch a general delegate …", "Otherwise inherit the parent model and never invent an identifier") |
+| maintain-verification-skill | 1 | GAP → FIXED (`badf6ac`) | serial self-execution fallback added; bounded with batching discipline in revision 2 |
+| create-verification-skill | 1 | GAP → FIXED (this branch) | revision 1 missed it entirely: no degradation existed for an undriveable app; added explicit stop-and-report clause in §1 Drive |
+| teach | 1 | COMPLIANT (vacuous) | zero capability-term hits |
+| technical-writing | 1 | COMPLIANT (benign) | L93 "the client and the host" is prose-style guidance, not a capability binding |
+| typescript-best-practices | 1 | COMPLIANT (benign) | "model variants" = TypeScript union modeling |
+| tdd | 1 | COMPLIANT (benign) | L28 lists browser automation etc. as optional regression-check examples ("Examples include"), no hard binding |
+| bro | 1 | COMPLIANT (vacuous) | restatement task; zero capability-term hits (absent from revision 1's table) |
+| unslop | 1 | COMPLIANT (vacuous) | prose-style edits; zero capability-term hits (absent from revision 1's table) |
+| principle-* without direct hits | 20 | COMPLIANT (vacuous) | only conditional mentions remain: build-the-lever L3/L20 ("When you fan work out to subagents…" — optional), make-operations-idempotent L20 ("respawns cleanly" metaphor); guard-the-context-window is among the 13 direct hits |
+
+Count check: 13 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + 1 + 20 = 44 ✓
+
+## Residual
+
+After both fixes, every skill carries an explicit verdict above and the counts
+reconcile to 44. Two gaps were found and fixed in total across revisions:
+`maintain-verification-skill` (badf6ac, bounded here) and
+`create-verification-skill` (this branch). No skill remains unclassified.
+Revision 1's "15/44 · zero open gaps" statement was false as written — it was
+not reproducible from the stated method, omitted three skills, miscounted the
+principle group by one, and mislabeled create-verification-skill COVERED. It is
+superseded by this document, not amended in place.

.agent/evidence/pr9-rerun-post-merge-2026-08-22/pr-body.md

@@ -0,0 +1,36 @@
+## Summary
+
+Re-raise of the post-merge review remediation that was closed unmerged as #245 (opened against the wrong default base — upstream instead of this fork; branch content was correct). Supersedes #245. Closes #11 (T7), closes #12 (T8); refs #13.
+
+The branch carries three remediation commits plus a clean merge of current main:
+
+- `a4eaaf1` fix(pstack): gate-5 degradation fixes and spec carve-out
+- `d6b4292` fix(scripts): floor guard and CRLF tolerance in reference checker
+- `12935ba` docs(evidence): reconcile gate-5 audit; fix dead evidence citations
+- `215edb0` Merge upstream tip (4612556) — absorbs docs(pstack) workflow/boundary port (#238); `git merge-tree` dry-run showed zero conflicts
+
+## What changed, honestly
+
+1. **Reference gate hardening** (`scripts/check-pstack-references.mjs`): floor guard and CRLF tolerance so the gate degrades predictably instead of mis-reporting on Windows-style line endings or empty match sets.
+2. **Gate-5 degradation fixes** under `pstack/`: bounded fallback behavior when subagents are unavailable, with spec carve-out documented rather than silently absorbed.
+3. **Evidence truthfulness**: dead evidence citations repaired; gate-5 audit wording now matches what the code actually does.
+4. **Reconciliation**: main absorbed; all gates proven against current upstream content, not a stale base.
+
+## QA evidence (fresh, at HEAD `215edb0`, 2026-08-22)
+
+| Gate | Command | Result |
+|---|---|---|
+| Portability (gate 1) | `npm run check-pstack-portability` | exit 0 |
+| References (gate 3) | `npm run check-pstack-references` | exit 0 |
+| Marketplace validation | `npm run validate-plugins` | exit 0 |
+| Runtime tests | `bun test pstack` | 52 pass / 0 fail, exit 0 |
+
+Raw logs: `.agent/evidence/pr9-rerun-post-merge-2026-08-22/` (`.log` files are gitignored by repo policy; substance mirrored above and in `.scratch/pstack-generic-gates/T7`/`T8`).
+
+**Known unrelated failure (not introduced here):** bare `bun test` at repo root discovers 139 tests across 32 files including `orchestrate/*`, which errors on missing optional deps (`@slack/web-api`, `zod/v3`). `orchestrate/` is byte-identical to `origin/main` (verified via diff), so this is pre-existing upstream state — filed separately as #13 with repro and suggested directions.
+
+## Process notes
+
+- Tickets: T7 → #11, T8 → #12 (markdown mirrors in `.scratch/pstack-generic-gates/`).
+- Fresh-context code review ran over `origin/main..HEAD` before this PR was raised; findings and verdict recorded in `.agent/evidence/code-review-findings.md`.
+- Review-only PR: raised by the authoring agent, intentionally **not merged** — merge approval stays human.

.agent/evidence/t10-port-ledger.md

@@ -0,0 +1,50 @@
+# T10 port ledger — upstream pstack 0.14.2 → pstack-generic
+
+Source: upstream/main 4612556 · Target: fix/port-upstream-pstack-0142
+Rule: every semantic upstream addition is either ported in host-neutral phrasing
+or omitted with a recorded reason. Identity files stay ours (SPEC non-goals).
+
+## autopilot-full.md
+
+| Upstream addition | Decision |
+|---|---|
+| `/goal` arming on operator go, persists across turns | PORTED — "persistent program goal … host's native goal or plan mechanism when it has one, otherwise a recorded objective every audit tick re-reads" |
+| Terminal `/loop` audit cadence + monitored-shell sleep | PARTIAL PORT — cadence stays on the existing host-neutral sentence ("event, wait, or recurring-work feature … bounded polling"); Cursor's terminal-/loop and cloud-sleeper mechanics are host-specific and omitted |
+| Output-notification sentinel per tick | PORTED verbatim-in-meaning |
+| Never leave cadence to memory / lossy completion notifications | PORTED |
+| Re-read playbook from trunk via `git show origin/main:<path>` + re-read armed goal | PORTED (path adjusted to this file) |
+| Fix drift during tick, treat as urgent | PORTED |
+| Liveness/status probe per owner | PORTED (replaces vaguer "probes delegated-work status") |
+| Count only side effects as progress (commits/pushes/PR-check deltas/store reports) | PORTED |
+| Stuck-lane stand-down + immediate replacement, "do not wait for a polite return" | PORTED |
+| Bugbot pass counters / `cursor-team-kit` deslop / control-cli-control-ui names | OMITTED — already generalized in ours ("skeptical automated-review triage", "installed deslop equivalent", "installed CLI or UI control skill"); naming the Cursor plugin would re-couple the generic variant |
+| Cloud-sleeper wake chain | OMITTED — Cursor cloud concept; ours' bounded-polling fallback covers it |
+
+## autopilot-stack.md
+
+| Upstream addition | Decision |
+|---|---|
+| `/goal` arming on explicit go | PORTED (same neutral phrasing as full) |
+| Audit-tick hardening block (sentinel, trunk re-read, side-effects progress, stuck stand-down) | PORTED (path adjusted) |
+| `gt` restack/submit mechanics, Graphite UI delivery | OMITTED — ours keeps source-control-host adapter phrasing ("whatever stacking tool the team uses", "active source-control adapter"); semantics preserved, tool coupling dropped |
+| Cloud-agent division-of-labor framing | OMITTED — host-specific |
+
+## opening-a-pr.md
+
+| Upstream addition | Decision |
+|---|---|
+| Per-invocation worktree isolation for sequential writers (`fetch && reset --hard origin/<branch>`) | PORTED |
+| `/technical-writing` layers except Diátaxis + one-word-per-action / keep articles / avoid `-ing` | PORTED (skill referenced generically) |
+| Conventional Commits title spec (types, scope examples, imperative subject, real symbol, no trailing period) | PORTED |
+| Sectioned description template Why/Scope/Tradeoffs/Blast Radius/Verification + drop-when-empty + media after sections | PORTED |
+| Boilerplate ban extended from "small PRs" to all PRs | PORTED (stronger form adopted; duplicate clause removed) |
+| Open ready never draft + ready-command fallback + verify via PR viewer | PORTED (host-neutral wording; absorbs ours' viewer sentence) |
+| Babysit reversal: opening ≠ babysit; separate pass only on request post-stack; stalls-build rationale | PORTED — replaces ours' auto-babysit line (upstream semantics win; recorded here) |
+| `gt` stacking specifics | OMITTED — team-tool neutrality retained |
+
+## plugin.json
+
+| Upstream change | Decision |
+|---|---|
+| name/displayName/description/publisher/homepage/repository/keywords/tags | OMITTED — identity is human-owned and intentionally generic (`pstack-generic`, praxstack) |
+| version 0.14.2 | PORTED AS `0.14.2-generic.1` — signals upstream-content parity of carried guidance while keeping the generic identity scheme |