Forge

Forge

Preferred (script-orchestrated):

SKILL_ROOT="<absolute-path-to-folder-containing-this-SKILL.md>"
SCRUTINY_BIN="$(bash "${SKILL_ROOT}/scripts/ensure-bin.sh")"
"$SCRUTINY_BIN" forge [--cwd <repo-root>] [--client <client>] <URL|KEY|#N>
# or: forge --inline --input "<desc>"

That requires the source CLI (acli / gh / glab) with install URLs on miss, mirrors ticket under .scrutiny/forge-<id>/, exports Figma via fcli when links exist, asks spawn (default single)|team, playwright (skip if missing), TDD, coverage, e2e → scaffolding (guess+confirm prefix, optional branch/worktree) → optional test-plan confirm → optional LOC estimate gate (max_loc) → implement agent → verify gate → ship step.

Scaffolding (host-owned, before implement): host guesses a conventional prefix from ticket type/labels/title → confirm via Select (guess first). If enable_branch, detect git state and offer create branch / +worktree / use current (default depends on whether you're on a base branch); worktree switches the working dir for implement+commit. Non-TTY follows branch_headless (auto = create only when on a base branch; never = current).

Implement agent must write .scrutiny/forge-<id>/pr.json (pr_title, pr_body citing the ticket URL only, commit_subject starting with the chosen prefix, commit_body), delete non-implementation junk (e.g. playwright temp media), and must not create branches, commit, push, or open a PR. After the agent exits, scrutiny forge confirms the commit subject (Input, default = AI value or guess), commits, then on a TTY asks whether to create a draft PR (PR-title Input + base branch prompt). --yes / non-TTY skips prompts and uses the defaults.

Sibling of /scrutiny and /parley (same binary, ~/.scrutiny/config.toml).

Usage

• /forge <Jira-URL|KEY-123> — Jira (acli)
• /forge <GitHub-issue-URL|#N|N> — GitHub (gh)
• /forge <GitLab-issue-URL> — GitLab (glab)
• /forge --inline <description> — no remote ticket
• /forge — ask for URL or description
• Branch name with PROJ-123 → Jira fetch when no arg

Bulk mode

"$SCRUTINY_BIN" forge bulk [--dry] [--concurrency <N>] [--yes]

Many tickets in one run, each on its own branch + worktree, run concurrently.

• Collect: menu (Paste ticket URL/key / Done), one at a time. Done ends; Done on first prompt exits doing nothing.
• Validate all (fetch + complexity), then ask same settings for all or per-item (same questions as single forge).
• Per item: new branch + git worktree <type>-<projectkey>-<number> (e.g. feat-nero-8729).
• Concurrency cap forge.bulk_concurrency (default 3); override --concurrency <N>.
• Non-headless (claude/cursor + tmux/zellij/iTerm2/Terminal.app): each item = own terminal container named by ticket key; panes named by role (PO/developer/tester/reviewer/evangelist/tdd-plan/implement) in that item's worktree; TDD plan validated in that item's pane.
• Ship step (confirm commit subject, PR title/body, draft PR) serialized on main terminal, one item at a time.
• Requires a git repo. tmux most reliable; Terminal.app best-effort.

Flags:

• --yes — headless: newline-separated keys/URLs from stdin, auto-answer from complexity, auto-commit + auto draft PR, no prompts.
• --dry — full flow, no agents, no real PR. Branches + worktrees still created; panes shown but never auto-closed; offered to delete created branches + worktrees at end.
• --concurrency <N> — override cap.

Binary

Skill root = folder containing this SKILL.md.

SKILL_ROOT="<absolute-path-to-folder-containing-this-SKILL.md>"
SCRUTINY_BIN="$(bash "${SKILL_ROOT}/scripts/ensure-bin.sh")"

Config: ~/.scrutiny/config.toml → [forge] (see shipped config/default.toml).

When [forge] max_loc is set, forge spawns an m-tier (or [agent_models].forge_loc_estimate) read-only agent after brief/TDD plan and before implement. Agent estimates PR add+del LOC under loc_exclude_* rules, writes loc-estimate.json with confidence, and if over budget prompts proceed/stop (non-interactive aborts).

Hard token rules

• Agents read ticket JSON + session JSON + brief markdown (and pack after impl). No re-run acli / gh / glab / fcli.
• No dump full config into prompts — use suggested_forge / session fields.
• Caveman I/O. Partition workstreams. No full-repo fish when context paths exist.

Agent workflow (IDE chaining — only if not using scrutiny forge)

0. Parse args

• --inline / remaining text → inline description
• URL / key / number → remote
• Empty → ask user for URL or description; STOP until provided
• Detect client: Cursor → cursor, Claude Code → claude, Codex → codex; else config default_client

Prefer jumping to $SCRUTINY_BIN forge instead of step 1+.

1. ensure-bin → forge-fetch

TICKET="$("$SCRUTINY_BIN" forge-fetch --cwd <repo-root> --client <client> --input "<arg>")"
# or: forge-fetch --inline --input "<desc>"

Show ticket path. Read suggested_forge from ticket JSON.

2. Confirm session (prompt only unset knobs)

If prompt_e2e false and e2e is null in ticket suggested_forge, treat as false.

SESSION="$("$SCRUTINY_BIN" forge-plan-write \
  --ticket "$TICKET" --client <client> --model <model> \
  --approach <tdd|heads_down|plan> --e2e <true|false> \
  --agents <n> --testers <n> --reviewers <n> --evangelists <n> \
  --cwd <repo-root>)"

Show session path. Read flags: enable_figma, enable_lore, enable_po, enable_ticket_writeback, skip_ai_review, counts, approach.

3. Context + brief

CONTEXT="$("$SCRUTINY_BIN" forge-context --ticket "$TICKET" --cwd <repo-root>)"
BRIEF="$("$SCRUTINY_BIN" forge-brief --ticket "$TICKET" --session "$SESSION" --context "$CONTEXT")"

Read markdown_path from brief JSON. Pass brief path (+ ticket/session paths) to every later agent.

4. Approach branch

plan

1. Spawn planner agent (model from session). Input: brief + ticket paths only.
2. Planner writes implementation plan (files, steps, risks) to a temp markdown path; report path to user.
3. Ask user confirm / request changes (loop). Exception: never skip this confirm in plan mode.
4. Continue to Phase A with approved plan attached.

tdd

1. After PO/requirements (Phase A), spawn test planner → structured test-case list (unit + e2e if session.e2e).
2. Ask user: confirm / change / reject. Loop until confirm.
3. Testers write failing tests (red). Agents implement to green. See Phase C.

heads_down

• No human confirm after session plan.
• PO + test plan use reviewer subagent approve (max 2 revision rounds) instead of user.
• Implement until done; still run tests.

Phase A — Gather + PO (full port)

A1. Conventional commit prefix

scrutiny forge guesses the prefix (feat|fix|docs|refactor|perf|test|chore) from ticket type/labels/title and confirms it with the user before implement. The agent must honor the provided prefix (in commit_subject), not re-derive it. IDE chain (no scrutiny forge): derive as before, ask if ambiguous.

A2. Figma (session.enable_figma)

If false → skip silently.

If true:

1. Use figma_urls from ticket JSON + scan description/comments (already in ticket).
2. Jira only: associated designs via Atlassian GraphQL + basic auth (ATLASSIAN_EMAIL / ATLASSIAN_TOKEN) as in address-ticket (best-effort; continue on fail).
3. If URLs exist and fcli available: fcli file info/inspect/export → /tmp/<id>-figma/. Pass paths to PO/designers.
4. No fcli → show URLs; ask for manual assets (skip ask in heads_down).

A3. Lore (session.enable_lore)

If false → skip silently.

If true and parent epic exists (Jira): maintain lore/<EPIC-KEY>.md (create/append ticket entry). Pass lore path to PO.

A4. Split evaluation

If ticket too large for one PR (many areas / mixed types / sequential deps), propose split. User chooses. Jira: optional sub-tasks via acli. Proceed with first chunk only. Skip silently if not warranted. In heads_down: keep as one unless clearly unsafe — then note in report.

A5. Product Owner (session.enable_po)

If false → synthesize minimal requirements from ticket title/description/brief; skip approve.

If true — spawn PO with: brief path, ticket path, attachments_dir, figma paths, lore, commit prefix, context related_paths / test_harness.

PO output structure (same substance as address-ticket):

## Ticket: <id>
**title** | Type | Priority | Commit prefix

### Understanding
### Requirements (numbered, testable)
### Acceptance Criteria (checklist)
### Design Needs (Yes/No + what)
### Implementation Plan
### Files to Modify / New Files
### Unit Tests / E2E Tests (honor session.e2e — if false, E2E = None)
### Risks / Open Questions

Approve requirements

• Not heads_down: show PO doc → Approve / Request changes / Reject.
• heads_down: spawn reviewer to challenge gaps; revise ≤2 rounds; then proceed.

A6. Designers (conditional)

From PO Design Needs:

• No → skip
• Yes + Figma assets → 1 designer
• Yes + no Figma → 2 designers (research + propose) in parallel

Pass design guide to test planner + developers.

Phase B — Test plan

Spawn test planner (not code). Input: approved requirements, design guide, context test_harness, session.e2e.

Output: unit (+ e2e if e2e) case list with file paths, AC coverage, edge cases.

Guidelines (keep):

• Every AC ≥1 test
• Bug → regression test
• No locale string asserts — use i18n keys
• No intermediate-state asserts that sibling epic tickets will break
• Follow project naming/layout from context
• Case titles: affirmative; start with a bare verb (renders…, returns…, opens…, does not…); no should / should not; no TC-* / numbered / ticket-id prefixes; nested describe = SUT/area only. Prefer matching nearby it()/test() style when present.

Approve test plan

• tdd / plan (interactive): user confirm/change/reject
• heads_down: reviewer subagent; ≤2 revisions

Phase C — Implement

Team sizes from session: testers, agents (developers).

C1. Red (testers)

If testers > 0 and approach is tdd or plan (after test approve): spawn tester agents — tests only, no production code. Partition unit vs e2e or by area. Verify fail for missing behavior (not syntax).

If heads_down and testers > 0: same red phase after auto-approved test plan.

If testers = 0: developers own tests while implementing.

C2. Green (agents)

Spawn agents developers — production code to pass tests; do not weaken tests. Partition by independent files/areas. Resolve same-file conflicts. Lint/compile sanity.

C3. Verify — host-owned gate (not the agent)

scrutiny forge runs the verify gate itself after implement; the agent no longer decides "done". Do NOT re-implement this in the IDE chain.

• Commands: [forge].verify_commands if set, else auto-derived from sniffed harness + project files: tests (vitest→npx vitest run --reporter=json, cargo-test→cargo test, …; e2e only when session.e2e) plus lint/build (cargo clippy; <pm> run lint|typecheck|build from package.json; tsc --noEmit).
• Coverage: gated when [forge].verify_coverage and measurable (per-framework json summary); unmeasurable → warn, never block.
• Loop: on red, host spawns forge-verify-fix with a surgical payload — failing test file:line + message, uncovered line ranges — up to verify_max_loops (default 2).
• Commit gate: green → commit; still red + TTY → ask "commit anyway?"; still red + non-TTY → no commit, exit non-zero.

Phase D — Post-impl probe (reuse scrutiny probe)

If session.skip_ai_review → skip AI; optional note.

Else:

EVAL="$("$SCRUTINY_BIN" eval --cwd <repo-root> --client <client>)"
MAP="$("$SCRUTINY_BIN" map --cwd <repo-root> --eval "$EVAL")"
PACK="$("$SCRUTINY_BIN" pack --cwd <repo-root> --map "$MAP")"
SCAN="$("$SCRUTINY_BIN" scan --cwd <repo-root> --map "$MAP" --pack "$PACK" --eval "$EVAL")"
PLAN="$("$SCRUTINY_BIN" plan-write \
  --eval "$EVAL" --map "$MAP" --pack "$PACK" --scan "$SCAN" \
  --client <client> --model <session.model> \
  --security true --performance false --error-handling true \
  --reviewers <session.reviewers> --evangelists <session.evangelists>)"

Spawn reviewers / evangelists per plan (pack-only). Evangelists = clean-code / architecture. Merge scan + AI → caveman list. Interactive triage like /scrutiny unless heads_down (then auto-keep high-severity only, list rest).

Model: pass session model to Task model= (Claude: Anthropic aliases/ids only, not Cursor slugs).

Phase E — Ticket writeback (session.enable_ticket_writeback)

If false → skip.

Compare original description vs approved requirements; propose improved description / comment.

• Jira: ask (or heads_down skip write unless clearly valuable comment). Update via acli jira workitem edit --from-json with ADF if user approves.
• GitHub: gh issue comment with summary + AC.
• GitLab: glab issue note.
• Inline: skip remote writeback.

Phase F — Final report

## Forge complete: <id>
**<title>**
Source: … | Approach: … | Model: …

### Team
PO / designers / testers / agents / reviewers / evangelists counts

### Implementation
files +/− lines

### Tests
unit / e2e status

### Acceptance Criteria
[x] …

### Writeback
yes/no

### Next
PR / manual steps

Edge cases

• Empty description → PO from title + comments
• Already done in codebase → stop and tell user before PO
• No test framework → skip red/green ceremony; agents implement + note
• Missing acli/gh/glab → forge-fetch errors; offer --inline paste
• Conflicting designer advice → ask user (heads_down: pick simpler)

Token-saving checklist (mandatory)

1. Artifact-first paths only
2. Brief not full ticket dump in every spawn
3. Context related_paths before Grep sprawl
4. Post-review = scrutiny pack only
5. Config force knobs → zero prompts
6. enable_figma / enable_lore false when unused
7. Partition agent prompts by workstream
8. Caveman findings
9. reviewers=evangelists=0 → skip AI review