workflow-coding

name	workflow-coding
description	Implement the smallest Smelter source change needed to turn RED tests GREEN under workflow gates.
version	0.55
type	workflow
consumes	*.test.* (RED) OR active_feedback
produces	src files changed
default_pattern	C
default_agent	executor
supports_patterns	["A","C"]
team_template	{"C":{"parallel_split_by":"module","agents":["executor"],"aggregator":"architect","conflict_resolver":"conflict-resolver","sync_point":"api-contract"}}
watchdog	E
can_delegate_to	["ui-ux-pro-max","copywriting","vercel-react-best-practices","designer"]
gate	{"pre":[{"tdd_cycle":"events[].type=test_red or legacy failing test_cycles[] exists"},{"allowed_in_mode":true}],"post":[{"src_changed":true},{"typecheck_clean":true},{"scoped_tests_pass":true}]}

workflow-coding

Implement the smallest source change that turns RED tests GREEN. This is not a terminal stage.

Announce: I'm using workflow-coding to turn RED tests GREEN. Next skill on pass: workflow-agent-review.

Required Inputs

• RED evidence from workflow-write-test: real failing command exit, not ; echo "exit=$?" masking.
• Or unresolved active_feedback targeting workflow-coding.
• Current tasks.md or implementation-plan artifact relevant to the change.

Execution Checklist

1. Read the failing test and related source before editing.
2. Change only source files needed for the RED case or active feedback.
3. Do not edit tests in this stage. If tests are wrong or missing, route back to workflow-write-test.
4. Keep the patch minimal: no hardcoded secrets, no silent error swallowing, validate external input at boundaries, avoid deep nesting.
5. Run scoped tests for changed surface.
6. Run typecheck (tsc --noEmit or repo equivalent) when applicable.
7. If typecheck/scoped tests fail, fix and rerun; this is not a retry because code changed.

Gates

• Pre-gate: RED evidence exists; otherwise route to workflow-write-test.
• Post-gate: source changed, typecheck clean, scoped tests pass.
• Full regression is opt-in at workflow-human-check; do not speculate it would pass.

Forbidden

• Claiming done/fixed/complete before workflow-human-check returns complete.
• Skipping workflow-agent-review because the change is small.
• Skipping workflow-e2e for UI/CLI/API/hook/script behavior changes.
• Asking the user whether to continue; the chain is autonomous until workflow-human-check.
• Using TODOs or broad fallbacks to hide the defect.

Subagents

For independent /implement modules, dispatch executor subagents only when parallelization is safe. Give exact task text, file map, acceptance criteria, and excerpts from implementation-plan.md; require self-review and test output. Subagent self-review never replaces Smelter review/E2E/human-check.

Fail Routing

• Missing RED evidence -> workflow-write-test
• Scope mismatch in /implement -> workflow-implementation-plan
• Scope mismatch in /fix -> workflow-investigate or mode upgrade
• Persistent verification failure after code changes -> stay in workflow-coding with evidence

Terminal State

After post-gate passes, invoke workflow-agent-review. Do not stop, summarize as complete, or jump directly to E2E.