Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add skills/blakemartz/delivery-loop/review-tasknpx skills add blakemartz/delivery-loop --skill review-taskgit clone --depth 1 https://github.com/blakemartz/delivery-loopWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00055 | $0.01105 |
| Opus 5 | $0.00028 | $0.00553 |
| Sonnet 5 | $0.00011 | $0.00221 |
| Haiku 4.5 | $0.00006 | $0.00111 |
Grade A, and why
review-task scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 2d ago.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 40 lines — stays where its author put it; the contents beside it link to each section on GitHub.
review-task — independent adversarial review
You are the reviewer role from the delivery spec (${CLAUDE_PLUGIN_ROOT}/docs/agentic_delivery_spec.md, §3.4, §11). Read .claude/delivery.conf for GATE_CMD, STYLEGUIDES_DIR, WORKTREE_SETUP_CMD, and NOTIFY_APPROVE_CMD.
Independence rules (these are the point)
- You must NOT read any implementer transcript, notes, or reasoning. If you were spawned with implementer context in your prompt, say so and refuse the review.
- Your only inputs: the linked issue (acceptance criteria + context),
gh pr diff <pr>, the conventions inSTYLEGUIDES_DIR, and a clean checkout of the PR branch. - Execute, don't trust. The PR body's checked boxes are claims, not evidence. You run every criterion yourself.
Protocol
gh pr view <pr> --json body,headRefName,number,statusCheckRollup— find the linked issue (Closes #N). No linked issue = request changes (protocol violation).- CI must be green or pending-then-green. A PR with a failing check is unreviewable: request changes citing the failing check, done.
gh issue view <N>— extract the acceptance criteria. These are the review contract.- Clean checkout:
git fetch origin && git worktree add .worktrees/review-pr-<pr> origin/<headRefName>(from the repo root; remove it when done). RunWORKTREE_SETUP_CMDin it if that value is set. - Execute every acceptance criterion in that checkout, plus the gate (
GATE_CMD). Record pass/fail with output. - Read the diff against the relevant styleguide and the spec sections the issue cites. Look for: criteria satisfied in letter but not intent, missing tests the criteria imply, scope creep beyond the issue, styleguide violations, changes to files outside the task's module without justification, and whether the solution follows best practice for the stack (spec §2.1) — an expedient or non-idiomatic approach that happens to satisfy the criteria is still a finding. Best practice is a review bar: flag a non-best-practice solution even when every criterion passes (as a review comment when no criterion is behind it — see Severity discipline; as request-changes when one is).
- Verdict:
- Every criterion passes and no substantive violation →
gh pr review <pr> --approve --body "<per-criterion results>"andgh issue edit <N> --add-label status:approved --remove-label status:in-review. Report "ready for human merge" — you never merge. - Any criterion fails, CI red, or substantive violation →
gh pr review <pr> --request-changes --body "<per-criterion findings, each actionable>"andgh issue edit <N> --add-label status:changes-requested --remove-label status:in-review. - Single-account fallback: if GitHub rejects
--approve/--request-changesbecause the PR is from the same account, submit the identical body asgh pr review --comment --body "<body>" <pr>(flags first) — thestatus:*label carries the verdict (spec §13.2). (If your harness allowlists specific command forms, this exact form is the one to allow.)
- Every criterion passes and no substantive violation →
- Verify the verdict landed.
gh pr view <pr> --json reviewsmust show your posted review. An unposted review did not happen: if posting was denied or failed, report FAILURE ("review executed but could not be posted: ") — never report approved/changes-requested as your outcome. Do not update issue labels for a verdict you could not post. On a confirmed approval only, ifNOTIFY_APPROVE_CMDis set in.claude/delivery.conf, run it so the human knows a PR is waiting to merge. Ring it whenever you land an approval in either form — a formal--approve, or the single-account approval--comment(which shows as a COMMENTED review, not APPROVED).status:approvedis the verdict of record, so do not condition the hook on the review showing APPROVED (in single-account mode it never will). Never run it on request-changes, and never for a verdict you could not post. - Cleanup:
git worktree remove .worktrees/review-pr-<pr> --force.
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- 2d ago First seen · 40 lines · 55 tokens per session scan A f88f2308d50c
review-task is a skill published in the GitHub repository blakemartz/delivery-loop (4 stars, last pushed 1mo ago), licensed MIT. It adds 55 tokens to every session and 1,105 once invoked, about $0.0003 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other skills, from other repositories
eliza-cloud-buy-domain
Use whenever a user wants to register or buy a custom domain for an Eliza Cloud app — including in the same request as building the app ("build me X and put it on Y.com"). Uses Cloudflare as registrar after explicit user confirmation, paid from the user's existing cloud credit balance. Pairs with build-monetized-app…
ordercli
Foodora CLI for checking past orders and active order status. Use when the user asks to check food delivery orders, track a Foodora delivery, view order history, reorder a meal, look up past food orders, check delivery status, or manage Foodora sessions and authentication.
asc
Drive App Store Connect from the terminal with the asc CLI — TestFlight builds, groups, testers and What to Test notes; App Store versions, metadata, keywords, screenshots and release notes; submissions and review health; signing, provisioning and notarization; crash and beta-feedback triage; pricing, subscriptions…
returns-policy
Answer return, refund, and warranty questions for electronics. Use when the user mentions returns, refunds, RMAs, warranty coverage, damaged items, or opened packaging.
pr
Open a pull request from the current worktree — verify green, commit scoped, push, and write the PR from the actual diff. Trigger on "/pr", "open a PR", "make a pull request".
payram-checkout-integration
Integrate PayRam checkout flow into web applications. Generate payment links, embed payment pages, handle redirects, and process payment confirmations. Supports Express, Next.js, FastAPI, Laravel, Gin, Spring Boot. Use when adding crypto checkout to e-commerce, building payment forms, implementing deposit flows, or…