Skip to content

adamayoung/TMDb / .claude/skills/deliver

deliver

Verified

FileKindSize
SKILL.mdskill md47.7 KB
references/auto-and-async.mdmarkdown9.5 KB
references/next-mode.mdmarkdown40.2 KB
references/reflexive.mdmarkdown3.0 KB
references/review-loops.mdmarkdown4.1 KB
references/worktree-lifecycle.mdmarkdown27.5 KB
references/wrap-up.mdmarkdown7.8 KB

references/review-loops.md

View on GitHub
# /deliver — review loops (reference)

Read on demand from `/deliver` **Phase 4** (code review) and **Phase 5**
(security review). The rules and caps live in `SKILL.md`; this file holds the
procedure detail and the incidents behind the rules.

## Why review once, after implementation settles (Phase 4)

Code review is an independent, adversarial lens — apply it once the design has
settled, not after every Canon TDD item. The test list is deliberately
emergent; early items are reshaped by later ones, so per-item review just
flags churn. The per-item quality gate already lives in the implement phase —
the passing test, the inline docs, the lint hook, and refactor-on-green.

## The 4a reference-unit gate (template→replicate deliveries)

For one pattern applied across **N≥3 cohesive units** (e.g. 25 sibling mocks,
"validate every public String input", one method mirrored across services):

1. Add a `Phase 4a — reference-unit review @ <sha>` task to the ledger.
2. Have the implement phase commit the **first unit on its own**.
3. Run `/review-changes` scoped to that commit
   (`git diff origin/main...<sha>`) and converge its Critical/High findings
   per the standard loop.
4. Only then replicate the pattern across the remaining units. **Phase 9 must
   not start while the 4a task is open.**

A wrong foundational pattern caught here is one *not* baked into all N units —
the #359 "reference-first" win: a cross-module DocC break caught in
`MockGenreService` before it replicated into 25 mocks. For **full-otherwise**
deliveries (multi-unit but not template-replicate), do **not** interleave
per-unit review — the units don't build on each other, so per-unit review only
adds churn; the single end-diff fan-out's per-dimension adversarial pass
already covers the whole diff.

## Why review fixes must be committed

`/review-changes` diffs **committed** history (`git diff origin/main...HEAD`),
so an uncommitted fix re-reviews as still-broken and the loop never converges.
Commit each fix before re-invoking.

## Enumerate-all-sites: the incident behind the rule

PR #364 ("encode every path interpolation, validate every public String
input"): the plan grep and `/security-review` each missed a *different* subset
of sites; Phase 4 review found a 4th encode site (`ReviewRequest`); the
`claude-review` bot then flagged three more unvalidated `String`-ID service
methods. Each pass found a subset; none enumerated the class. Sweep by
**type**, not by eyeballing — e.g. `grep 'path = "/.*\('` for String-into-path
interpolations *and* scan public service signatures for `String`/`*.ID`
params (`Int` IDs are injection-safe and need nothing) — confirm the full
list, then implement against it.

## Specialist skills: the incident behind the checkpoint

Delivery #359 hand-rolled an `NSLock`/`@unchecked Sendable` mock design and
only consulted `swift-concurrency` when the user prompted — at which point the
skill both validated the choice *and* caught a missing `@unchecked Sendable`
removal plan the ADR had omitted. The checkpoint is topic-triggered (the
moment the domain appears), not stuck-triggered, and applies to fanned-out
subagents/Workflows too — give them the same instruction.

## Security review (Phase 5) — what actually bites in this library

- The `HTTPClient` / networking layer.
- `api_key` and session / credential handling.
- URL and query construction.
- `Decodable` paths over untrusted API data.
- Anything that logs.
- Dependency / permission changes (`Package.swift` / `Package.resolved`,
  `.github/workflows/`, `.claude/settings*`).

`/security-review` grades **High / Medium / Low** (no "Critical" tier),
filters out findings below confidence 8, and excludes documentation files —
a docs-only diff returns nothing. It produces findings; it does **not** fix.
Fix test-first via `canon-tdd` where the defect is reproducible (input
validation at a public boundary, a decoding guard), or by the minimal direct
change where it is not (removing a secret-leaking log line, tightening a
permission). Phase 9's `make ci` is a **correctness** gate with no SAST step,
so Phase 5 is the pipeline's **only** security gate.
slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 194kUpdated Claude CodeCodexVerifiedInstall