Code Forge Agent Skill

Code Forge

Required Skills

Read, apply:

• framework:knowledge-priming -- Load proj context (stack, arch, conventions) so impl matches real proj (always)
• framework:context-anchoring -- Load/find context anchor doc; enrich as impl decisions made (always)
• framework:learning-harvest -- Load prior operational learnings inform impl; harvest new patterns at session end (always)
• framework:collaborative-judgment -- Surface real judgment calls w/ structured opts vs silent assume (always)
• framework:architecture -- Layer place, dep direction, struct valid (always)
• framework:clean-code -- Craft rails: SRP, naming, complexity, err handle (always)
• framework:domain-driven-design -- Aggregates, entities, VOs, domain svcs (conditional: only when touch domain folder)
• framework:secure-coding -- Trust bounds, injection prevent, secrets mgmt (conditional: only boundary-cross code)
• framework:test-quality -- AAA struct, isolation, assert quality, naming (always when write tests)

Workflow

Step 1: Establish Implementation Context

Use framework:learning-harvest Load behavior. Focus hint: "implementation session — focus: implementation craft, quality signals, reliability".

Use framework:context-anchoring Doc Discovery check existing context anchor doc for feature impl.

• If found → Load (context-anchor Load behavior). Present struct ack -- feature name, status, decision count, open Qs, constraints. Honor all logged decisions/constraints as active commits.
• If not found → Nudge user: "Have design doc/blueprint for feature? Or work from discussed?" Accept either graceful.
• User provides doc → load, follow.
• Proceed without → all atom rails still apply; just no approved design doc to reference. Work from verbal reqs in convo.

Design completeness check — STOP gates before Step 2:

Check 1 — status: Read frontmatter status.

• approved → pass.
• Anything else → STOP: "Context doc not approved (status: [value]). Run design-blueprint first. Proceed anyway?" Confirm → log in Decisions Log, continue as "Without approved design."

Check 2 — levels present: Scan body for ## Design: Level 3 and ## Design: Level 4.

• Both present → pass.
• Either missing → STOP: "Missing [Level 3 / Level 4 / both]. Proceed anyway?" Confirm → log absent levels in Decisions Log, treat as gaps to fill during implementation.

Both pass → proceed as "With approved design".

Step 2: Plan Implementation Order

With approved design: Extract component list, layer assigns from context anchor doc. Use L2 (Components) decisions for layer place, L3 (Interactions) for dep flow.

Without approved design: Classify req components→arch layers using layer defs from framework:architecture. Each component, determine:

• Primary responsibility? (biz rules, data access, coord, external I/O)
• Which layer in loaded arch doc matches responsibility?
• Dep constraints for that layer?

If framework:architecture no loaded layer defs (neither defaults nor custom doc resolved), warn: "No arch rules avail. Run /architecture-refiner define arch standards. Proceed w/o arch guidance." Continue w/ only remaining atom rails.

Present proposed layer assigns→user for approval before proceed.

Both cases, plan inside-out impl order following dep direction from loaded arch doc — start innermost layer (no outward deps), work outward. Each layer's deps should exist when built.

Classify each op per flow patterns in loaded arch doc (e.g., cmd vs query flows, or equiv distinction your arch style).

Present impl plan -- ordered component list, layer assigns, flow classifs -- confirm w/ user before write code.

After plan approved, ask user choose review mode:

> "How review impl?"

> 1. Layer-by-layer (rec) -- Impl each layer fully, pause for review before next. One review pt/layer.

> 2. Full autonomy -- Impl everything end-to-end, present complete result. One review pt at end. (If blueprint exists, still pause any deviation from approved design.)

> 3. Component-by-component -- Pause after each individual component for feedback. Max review pts.

Default layer-by-layer if user no preference.

Step 3: Implement Per Component

Each component in planned order, gen code+tests together -- tests not afterthought.

Every component:

• Prefer the simpler path first. Before writing custom code: does a stdlib function, platform built-in, or existing dependency already cover this? If yes — use it. Can it be expressed in shorter code? Use that. Write custom code only when simpler options genuinely fall short.
• Place correct arch layer per framework:architecture. Valid dep direction follows loaded arch rules.
• Apply framework:clean-code self-valid during gen. Run inline checks: SRP comply, meaningful naming, low cyclomatic complexity, proper err handle, no magic vals, clean func sigs, no dead code, appropriate abstract level, clear control flow, minimal comments (code self-doc).
• Write tests using framework:test-quality self-valid.

Conditional checks per component:

• If domain layer → Apply framework:domain-driven-design self-valid.
• If trust boundary (HTTP handler, external API call, user input process, file I/O) → Apply framework:secure-coding self-valid.
• If blueprint exists AND Level 4 was confirmed present in Step 1 → Verify component fulfills L4 (Contracts) spec. Flag any deviation from agreed contract. If user proceeded without L4 (Step 1 Check 2 failed), skip this check — there are no contracts to verify against.

Post-Gen Verification (applies every component, all review modes):

After gen each component, before present→user:

• Run Self-Valid Checklist from each applicable atom against every func/class this component. Atoms use imperative STOP-verify lang -- follow literally.
• Run Active Anti-Pattern Scan from each applicable atom. Check every box scan list.
• Violations found → fix before present. Don't present code you know violates atom checklist.
• Judgment calls flagged (see each atom's Ambiguity Signals) → collect. Present using framework:collaborative-judgment protocol before show code. Don't silent resolve.
• All checks pass, no flagged judgment calls → present w/ brief comply note (e.g., "All clean-code, DDD checks pass"). Keep one line when clean -- only verbose when report violations, fixes.

Pacing -- follow user's chosen review mode:

• Layer-by-layer: Impl all components within layer, present full layer (code+tests) for review before next layer.
• Full autonomy: Impl all layers continuous. Present complete impl (all code+tests) at end. Skip→Step 4 (Cross-Component Verif) after all components done.
• Component-by-component: Present each component w/ tests individually. Wait approval before next.
• Exception (all modes): Component needs significant deviation from plan (new dep, changed contract, unexpected complexity), pause immediately, discuss before continue -- regardless chosen review mode.

Step 4: Cross-Component Verification

Step checks arch coherence -- not code quality (verified per-component Step 3). After all components impl:

• With blueprint: Verify interaction flows match L3 (Interactions) design. Every designed interaction traceable in code.
• Dep direction: Apply framework:architecture verif across all components — verify inter-component dep direction follows loaded arch rules. No layer import from layer not permitted depend.
• Zero Impl Rule: Check no new components, interactions, contracts intro beyond planned Step 2. Something added, flag -- may be necessary, but should be conscious decision, not scope creep.
• Final security scan: Apply framework:secure-coding across component boundaries. Check data flowing between components crosses trust bounds safely.
• Learnings check: If operational learnings loaded Step 1, verify previously-flagged patterns not recur this impl.

Step 5: Enrich Context

Throughout Steps 3-4, use framework:context-anchoring Enrich behavior keep living doc current:

• Add key files as created -- path, purpose, layer assign.
• Capture impl decisions -- lib choices, pattern selects, deviations from blueprint, tradeoffs made.
• Resolve open Qs -- Qs from design phase answered during impl, log resolution.
• If no context doc exists, significant impl decisions made → suggest create. Decisions worth preserve future sessions.

Harvest learnings. Use framework:learning-harvest Harvest behavior. Session context: "implementation session — code generation from design contracts". Synthesize and propose cross-cutting patterns from this session — implementation gotchas, design-to-reality gaps, library/framework lessons. User confirms what enters the document. STOP: run this before closing the feature lifecycle below.

Close feature lifecycle: Write status: complete to context doc frontmatter. STOP: required.

STOP: do not write status to requirement_doc. The feature file's status is owned by whoever manages the requirement — a human, or an external system it may live in. This molecule manages its own context doc only.

After enrich context doc, recommend review:

> "Impl complete. Recommend run /review on gen code before consider feature done -- provides independent quality assess against same atom standards, catches issues generator may blind to, captures learnings future sessions."