agentsclimarketplace

Feature

Skill travelhawk/forgeshift/.claude/skills/feature

Build one feature through the full quality loop - plan, failing test, implement, fresh-context verify. Use for any feature or task from the spec/PROGRESS.md, or an ad-hoc feature request during a build.From its SKILL.md

Install
npx -y skills add travelhawk/forgeshift --skill feature

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

  • 2 stars2 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

5.5 KB, as published. Nobody here has run it

/forge:feature — One feature through the loop

Build "$ARGUMENTS" tests-first with independent verification. One feature per invocation; a whole backlog belongs in /forge:build, and a raw batch of (3+) independent items in the feature-pipeline workflow — point the user there when they list several.

1. Anchor

  • Establish the target product (normally the directory you launched claude in; ambiguous → ask, then cd).
  • Find the feature in docs/SPEC.md / PROGRESS.md and use its done-criteria. Ad-hoc → write 2–5 checkable ones now and get a nod.
  • Read its risk tier (T? on the feature's row). Untagged → classify by capability signal (FORGE_HOME="${CLAUDE_PLUGIN_ROOT:-$(forge-home)}", then $FORGE_HOME/docs/RISK-TIERS.md; ties up). /forge:feature F3 as tier 1 overrides for this run — state the tier you're using and why.
  • Run the suite. Starting red means fixing it first, or recording explicitly that the red is pre-existing and unrelated.
  • Note remote status; a GitHub remote with gh auth status passing enables step 5's PR flow.

2. Plan — size it honestly

  • Small (one file, diff describable in a sentence): plan inline, skip ceremony.
  • Medium: show the inline plan and proceed without waiting; the user can interrupt.
  • Large or judgment-heavy (new subsystem, data-model change, security-relevant): delegate to forge-blueprint, record it as docs/features/F<#>.md from $FORGE_HOME/templates/FEATURE.md, present it as an artifact/link, and block on approval.
  • Integrates an external HTTP API? Capture the verified request/response contract now — delegate to forge-prospector (it has web access) to confirm endpoint, auth header, body shape and the success/error signal against live vendor docs. forge-quench has no network; the recorded contract is the only spec it can verify the adapter against. Skip this and a faithful-looking-but-wrong adapter passes review (Postmark returns a non-zero ErrorCode inside an HTTP 200 — only the docs tell you that).

3. Build

  • Own branch feature/<F#-or-slug> — never commit a feature straight to main. Even a one-line fix gets a branch; the PR is the review record.
  • Implement yourself for small work, via forge-hammer for medium+ (give it the plan, done-criteria and paths). Effort follows the tier: T1 tests-first at high/xhigh, T2 tests-first at high, T3 fast on Sonnet at medium with a smoke test as its test.
  • The gate while building is typecheck, lint and the tests covering the diff — not the full suite on every cycle. Full suite and e2e run once, at step 4, before integration.
  • E2E: at most one spec for this feature, often none — only for what no other layer can reach. Written after the feature works, never e2e-first.

4. Verify (fresh context — depth follows the tier)

The verifier sees only the result, never the build reasoning. Run the full suite and the e2e suite here, before integrating. Fix CONFIRMED critical/high immediately; judge medium/low with the user if the fix isn't obvious.

Post-build tier re-check (T2/T3): the tier was seeded from the feature's description; the built diff is the truth. A cheap Haiku pass reads git diff --merge-base HEAD <branch> and escalates to the T1 security pass if the diff touched a tenant-scoped query, an external-input parser, token/secret handling, authn/authz, or raw SQL/shell. Raises depth only, never lowers it.

  • T1 — forge-quench on the diff and a parallel security pass (forge-warden, or a deep-review scoped to this diff). Done only when both pass; verify covers failure paths. Scope the review to the diff — an unscoped audit over the surrounding subsystem can cost more than the feature took to build.
  • T2 — one forge-quench pass on the diff. Core coverage, no edge-case grinding.
  • T3 — smoke check only: builds, renders, happy path. No forge-quench; the integrated deep-review before ship is the net that sweeps T3.

A T1 change never skips verify, however small. The cosmetic-skip shortcut is T3-only — say so when you take it.

5. Integrate — one PR per feature

  • git push -u origin feature/<...>, then gh pr create --head feature/<...> --base main --body-file <scratchpad> — body = summary, done-criteria checklist, step-4 evidence. --head is required so the session checkout stays on main.
  • Leave the PR open; this is the supervised lane, the merge call is the user's. Ask and it squash-merges (gh pr merge <n> --squash --delete-branch).
  • No remote: offer gh repo create --private --source . once; declined → git merge --no-ff (the merge commit is the audit trail). Never an unreviewed fast-forward onto main.

6. Close

  • Tick PROGRESS.md only with pasted evidence, and note the PR. Report anything the environment could not prove as open, never ticked.
  • Update docs/SPEC.md if the implementation legitimately deviated.
  • Report: what shipped, evidence, PR link, deviations, next feature.
  • Next → one command. Never /forge:ship while verification is red.

Escalation

Two failed attempts at the same problem → stop grinding: /forge:debug-hard for bugs, forge-blueprint or design-panel for design dead-ends. A third identical attempt is banned.

Keep looking

Skills are one crate of 627,801. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.