Dev Workflow

Dev Workflow — Adaptive Feature Implementation

One engine, two interactive entry modes. The caller (slash command / prompt) sets the mode; everything else lives here.

Inputs: task description or issue reference, plus the mode. Output: implemented + tested + gated code, ready to /ship, with a gate file as proof.

Core principles

1. The orchestrator keeps all context — planning is never delegated.
2. One human stop (interactive mode): everything the user must judge is on one screen at Phase 2.
3. Rigor scales with the level — never run level-2 ceremony on a typo, never run a typo circuit on a migration.
4. Escalate, never restart — if scope grows, move up a level reusing all acquired work, and say so.
5. Verification comes from the manifest (.agents/verification.yaml) — never guess commands.
6. Quality is proven by the gate file, not by the user re-reading the diff.

Phase 0 — PROBE (silent)

Run the project-probe skill: read .agents/verification.yaml, create or refresh it if absent/stale. Read repo conventions (CLAUDE.md / AGENTS.md, local project memory pointer if present).

Phase 1 — EXPLORE

1. If the task is an issue reference (#42, owner/repo#42): gh issue view 42 — extract title, description, acceptance criteria, labels, Figma URLs.
2. Explore the codebase: architecture, impacted files, patterns to follow, dependencies, risks.
3. Surface detection (feeds gate lenses and level scoring):
   • Frontend: impacted files .tsx/.jsx/.vue/.svelte/.css, components/CLAUDE.md or components.json present, Figma URL in issue, UI keywords. If detected: read the components inventory; fetch Figma design context when available.
   • SEO/GEO: public indexable surfaces — homepage/landing/blog, .mdx, metadata, schema/JSON-LD, robots.txt, sitemap.*, llms.txt, title/meta/H1/canonical/FAQ, SEO keywords in issue.
4. Assess the level:

State the detected level and why. In interactive mode the user can override it at the Phase 2 stop.

Escalation rule (applies in every phase): the moment reality exceeds the level (4th file touched at level 0, schema change discovered at level 2, 10+ plan steps...) — announce it, re-classify the task against the level grid (a level-0 fix that turns out multi-file may jump straight to level 2), keep everything already produced (explore synthesis, partial plan, code). A quick-fix that grows becomes a level-1/2 run mid-flight; it never aborts.

Phase 2 — PLAN

Skip at level 0 (go straight to Phase 4 with a one-line intent statement).

Build the plan: numbered atomic steps — each with what / where (absolute paths) / how (pattern to follow) / constraints (what NOT to touch) / acceptance criteria (testable) / test strategy (P0-P3). If frontend: components to reuse, tokens (never hardcoded values), Figma mapping. If SEO/GEO: impacted public files, facts that stay Non vérifié without preview/GSC.

• interactive → ⛔ THE STOP: present ONE screen = explore synthesis + detected level + full plan + acceptance criteria + test strategy. The user validates, adjusts, or changes the level. This is the only stop before Phase 6.
• quick-fix (escalated to level ≥ 1) → present the (mini-)plan as the stop, exactly like interactive at that level.

Phase 3 — RED (conditional)

Run only if level ≥ 2 AND testability.harness ≠ none in the manifest.

Write the acceptance tests derived from the plan's acceptance criteria — and run them to prove they FAIL for the right reason before any implementation. These tests encode what the user validated; the gate will prove they turned green.

If skipped (no harness): acceptance criteria will be verified at runtime in the gate (manifest testability.runtime_verify, or the verify skill when the runtime offers it) and the gate file records absents: ["no test harness — acceptance verified at runtime"].

Phase 4 — IMPLEMENT

Execute plan steps sequentially. For each step: code + its unit tests together → run the manifest's commands (lint, typecheck, test) → green before the next step. Follow repo conventions; no unrequested refactoring; no out-of-scope edits.

Parallelizing two steps is allowed ONLY if their file sets are disjoint — and with worktree isolation when the runtime provides it. Sequential is the default, not the fallback.

Phase 5 — GATE

Level 0: run the manifest commands, present the fix (no gate file). Done — Phase 6. In autonomous mode the minimum level is 1: a gate file is always produced, because the PR needs proof.

Levels 1-4: commit the implementation first — the gate evaluates the committed diff (<base>...HEAD); uncommitted work is invisible to it. Then run the quality-gate skill on the default-branch diff with the validated plan and the manifest. Gate level = task level (1 → 1 round with a quick structural-smell check; 2 → ≤3 + final thermo-nuclear-code-quality-review; 3-4 → ≤4 + design-audit / seo-geo-audit / a11y-enforcer lenses for the surfaces detected in Phase 1, then final thermo-nuclear-code-quality-review). Do not run a separate maintainability review outside the gate; the final thermo-nuclear lens is part of the gate and is de-duplicated against earlier findings. Commit the gate file with the branch.

Phase 6 — HANDOFF

Present the final report: verdict + rounds + findings summary, decisions_prises_en_ton_nom (every deviation from the validated plan), absents, diff stats, and the RED→GREEN evidence when Phase 3 ran.

• Level 0 (quick-fix): present the fix, the manifest verification results, and a suggested commit (include .agents/verification.yaml if the probe just created it); never auto-commit.
• Levels 0-2 (interactive): propose [S] /ship | [C] commit only | [R] re-run the gate (levels 1-2). This menu is the handoff decision, not a second checkpoint — do not re-open the plan here.
• Levels 3-4 (interactive): require the user to read decisions_prises_en_ton_nom (quote it in full) before proposing the same options. This is the only careful read left to the human.
• Frontend + new components: propose /ds-doc --update before ship.

Runtime capabilities

• Claude Code: dispatch Phase 1 exploration to an Explore subagent; use TaskCreate for 2+ plan steps; quality-gate uses native /code-review as primary lens with parallel lens subagents; parallel implementation only for disjoint-file steps, preferably in worktrees.
• Sequential runtimes (Codex CLI, OpenCode): run every phase inline in order; quality-gate lenses run one at a time with an explicit reset between lenses. Same phases, same rules, same gate file.

Anti-patterns

• ❌ Running the full ceremony on a level-0 task, or a level-0 circuit on risky scope
• ❌ Restarting from scratch when scope grows — escalate and reuse
• ❌ More than one human stop in interactive mode (the plan is THE stop; Phase 6 levels 3-4 is a read, not a re-plan)
• ❌ Hardcoding verification commands instead of reading the manifest
• ❌ Writing RED tests that can't fail, or skipping the failing run
• ❌ Shipping levels 3-4 without surfacing decisions_prises_en_ton_nom

Referenced knowledge & skills

• Verification manifest: project-probe skill (.agents/verification.yaml)
• Quality loop & gate file: quality-gate skill (docs/quality/GATE-*.yaml)
• Final structural maintainability lens: thermo-nuclear-code-quality-review skill (inside quality-gate, not as a separate post-gate review)
• Testing strategy: .claude/knowledge/testing/test-levels-framework.md, test-priorities-matrix.md
• Design/SEO gates: design-audit, seo-geo-audit, a11y-enforcer skills
• Design system docs: ds-doc skill