Skip to content
Back to skills

Plans

ASecurity

> **Revision:** 2026-07-30 third final fail-closed correction after live > Rosary-contract review. **Goal:** Deliver a portable per-session `park-work` skill that returns `safe_to_close=true` only after strict durable evidence folds to completed or parked and the exact receipt is written and read back. **Architecture:** `SKILL.md` owns orchestration and human-facing boundaries. `scripts/fold.py` owns typed protocol validation, the deterministic fold, caller-stable retry planning, repository b...

  • 2 stars
  • 0 votes
  • 0 copies
  • 3 views
  • Added September 19, 2026
ai-agentspythongobashgitbackend

Works with

  • terminal
  • cli
  • mcp

Security analysis

A100/100

Pro scans all 3 files and shows the line behind each finding

Scanned September 19, 2026

npx -y skills add jamestexas/agents --skill plans --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Plans?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Plans
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/jamestexas-plans/badge)](https://www.skillsdirectory.com/skills/jamestexas-plans)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
# Park-Work Skill Implementation Plan

> **Revision:** 2026-07-30 third final fail-closed correction after live
> Rosary-contract review.

**Goal:** Deliver a portable per-session `park-work` skill that returns
`safe_to_close=true` only after strict durable evidence folds to completed or
parked and the exact receipt is written and read back.

**Architecture:** `SKILL.md` owns orchestration and human-facing boundaries.
`scripts/fold.py` owns typed protocol validation, the deterministic fold,
caller-stable retry planning, repository binding, receipt validation, and the
atomic resume gate. `tests/test_fold.py` exercises those behaviors directly.
Rosary remains the durable authority; no other repository is modified.

**Tech stack:** Markdown, Python 3 standard library, `unittest`, bound Git/jj
commands, Rosary MCP, and existing Bash repository gates.

## Global constraints

- Track work episodes independently from provider sessions.
- Require explicit caller-stable episode/intent IDs before park evaluation.
- Mint a fresh attempt ID for every evaluation retry.
- Require exact unique named checks with fixed categories and structured
  evidence.
- Require an explicit protocol phase: checkpointable preflight or durable.
- Never validate a preflight as receipt-safe.
- Require current-client child evidence independent from Rosary dispatches.
- Derive PR backing only from structured bead PR URL metadata.
- Keep completion unavailable until `rosary-a6166d` supplies command-bound,
  authoritatively ordered verify history; never synthesize missing history
  fields.
- Permit parking from a mechanically nonterminal Rosary status with unknown
  close history, exact known PR applicability, and a passing resume resolver.
- Bind both resume selectors to one registered repository/origin and bind
  every VCS command to its root.
- Correlate successful receipt anchor, repository binding, backend, root,
  workspace, and immutable head.
- Inspect all same-intent successes, reject semantic conflicts independent of
  order—including exact acceptance/PR contracts—and preserve exact duplicate
  source bytes/comment IDs.
- Fail active resume closed until `rosary-04faf5` supplies an atomic episode
  claim/lease.
- Never close a bead, terminate a provider, delete a worktree, or rewrite user
  changes.
- Preserve unrelated `.beads/` state and do not hand-edit bead projection
  data.
- Keep v1 per-session; bulk discovery, provider termination, Canonical Hours
  projection, and Rosary mechanism implementation remain out of scope.

## Task 1: Capture final-review regressions first

**Files:**

- Modify `skills/park-work/tests/test_fold.py`

Add executable fixtures whose expected values are hand-derived from the design:

- terminal completed-gate blocking and durable parked;
- checkpoint-required parked and blocked terminal preflight;
- active dispatch, active/unavailable current-client child, conflict, and
  unpreserved work;
- exact check-name/category/cardinality schema;
- PR-backed conditional checks and malformed exact provider evidence;
- malformed IDs, RFC3339 timestamps, repository, sessions, references, resume
  shapes, target correlation, and outcome correlation;
- process-separated unsafe/success retry identity and duplicate prior success;
- common repository binding for both selector types and contradiction cases;
- missing/drifted resume targets; and
- atomic resume gate order.

Run before production edits:

```bash
PYTHONDONTWRITEBYTECODE=1 python3 skills/park-work/tests/test_fold.py
```

Expected RED: the existing helper lacks the new protocol entry points and
cannot import `bind_repository`, `prepare_attempt`, or `resume_gate`.

## Task 2: Implement the typed helper

**Files:**

- Modify `skills/park-work/scripts/fold.py`

Implement standard-library-only functions and matching CLI commands:

```text
evaluate
validate-receipt
confirm-readback
prepare-attempt
bind-repository
resume-gate
```

### Observation validation

Require exact top-level phase, PR-derived bead metadata, and exact check
objects. Reject missing/duplicate/unknown/mis-categorized/inapplicable checks,
invalid outcomes, malformed evidence, and non-RFC3339 timestamps with
`InputError`/CLI exit 2.

Encode check-specific authoritative evidence:

- Rosary active/dispatch records;
- Codex/Claude current-client child arrays;
- bound Git/jj conflict and operation arrays;
- preflight checkpointable vs. durable preservation references;
- typed unavailable close-condition evidence without synthetic command/kind/
  sequence/latest/authoritative/complete facts;
- structured Rosary terminal status;
- exact `gh pr view <url> --json state,mergedAt,url` response; and
- exact resume resolver target.

### Fold

- Preflight with checkpointable preservation may return only
  `eligible=false, action=checkpoint`.
- Terminal `done|closed` remains unsafe because completion cannot be proven
  without `rosary-a6166d`.
- Durable `open|in_progress|blocked` plus known applicable PR evidence and a
  resolved target returns parked/write-receipt despite unknown close history.
- Every other well-formed observation is unsafe.

### Receipt

Require canonical prefixed UUIDs, anchor, provider sessions, absolute
normalized path, VCS enum, branch string, immutable head, references,
successful outcome/safety correlation, and for parked exactly one string resume
target and resolved immutable head matching passing resolver evidence and
`repository.head`. Correlate the confirmed anchor, complete validated
repository binding, VCS source/backend, both preservation workspaces, and
parked resolver with the receipt repository and head. Accept only durable
phase.
Parse exact fenced receipt comments in the helper and require matching
validated bytes; a missing readback or conflicting same-intent receipt remains
unsafe.

### Retry and binding

- Missing stable IDs produce a retry command and no attempt ID.
- Inspect every matching prior durable success; reject semantic conflicts
  independent of input order, including the exact bead object and derived
  `pr_backed` contract, and return only semantically identical duplicates with
  their exact source bytes/comment IDs.
- Otherwise mint a fresh attempt.
- Normalize registered origins without collapsing nondefault ports, require
  one match, correlate full receipt/selector/path/backend identity, validate
  exact root/Git-dir/common-dir/remote command observations, and return bound
  argv-style commands.

### Resume

Return `authorized=false`, allow only read-only inspection, name
`rosary-04faf5`, and block workspace creation, resumed observation, and work in
that order. Truthy comment-dedupe input must not authorize.

Run:

```bash
PYTHONDONTWRITEBYTECODE=1 python3 skills/park-work/tests/test_fold.py
python3 -m py_compile \
  skills/park-work/scripts/fold.py \
  skills/park-work/tests/test_fold.py
```

Expected GREEN: 79 tests, 0 failures; compilation exits 0.

## Task 3: Align the skill workflow

**Files:**

- Modify `skills/park-work/SKILL.md`

Document:

- explicit stable invocation IDs and the stop/retry contract;
- current-client Codex and Claude child adapters;
- exact typed check schema/evidence mapping;
- checkpoint-before-any-receipt for an eligible nonterminal parked candidate;
- durable-phase-only receipt construction/readback;
- unavailable Rosary completion gating, nonterminal parking, and exact provider
  PR evidence;
- common registered-repository binding and explicit VCS roots;
- read-only resume inspection; and
- the missing atomic claim gate.

Preserve every absolute destructive boundary and v1 scope boundary.

## Task 4: Align design and plan

**Files:**

- Modify
  `docs/superpowers/specs/2026-07-30-work-episode-park-resume-design.md`
- Modify this plan

Replace prior behavior that:

- treated checkpointability as preservation;
- minted unrecoverable retry identity;
- inferred child quiescence from Rosary;
- inferred completion evidence;
- used different resume selector binding; or
- treated comments as single-holder resume safety.

Update fixture totals to the executable 79-test suite. Prose/token tests remain
supplemental and do not substitute for executable behavior.

## Task 5: Required verification

Run fresh:

```bash
PYTHONDONTWRITEBYTECODE=1 python3 skills/park-work/tests/test_fold.py
python3 -m py_compile \
  skills/park-work/scripts/fold.py \
  skills/park-work/tests/test_fold.py
scripts/build.sh readme
scripts/build.sh check
git diff --check
```

Run targeted CLI probes:

1. terminal checkpointable preflight returns `eligible=false` and
   `action=checkpoint`;
2. uncorrelated receipt anchor/repository/backend/workspace evidence exits 2;
3. conflicting prior successes fail closed in both input orders while exact
   duplicate sources are preserved;
4. different remote ports, selector mismatch, and false Git common-dir
   relationships exit 2; and
5. no synthetic Rosary history fields are accepted; `open|in_progress|blocked`
   park, while `done|closed` remain unsafe without `rosary-a6166d`.

Verify status contains only the intended skill/helper/tests/design/plan,
README only if generated content changed, the final report, and the
pre-existing unrelated `.beads/beads.jsonl`.

## Task 6: Report and commit

Write
`.superpowers/sdd/2026-07-30-park-work-skill/final-fix-3-report.md` with:

- exact RED and GREEN commands/output;
- every final-review finding disposition;
- protocol/schema decisions;
- files changed;
- required gates and targeted probes;
- residual concerns; and
- commit SHA.

Stage no `.beads/` data. Commit all final corrections once:

```bash
git -c commit.gpgsign=false commit \
  -m "[agents-dba741] fix(park-work): gate completion on authoritative history"
```

Files in this skill

  • 2026-07-30-park-work-skill.md9.5 KB
  • 2026-07-30-workerd-work-board.md67.7 KB

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.