tunneleven/c4flow:code

/c4flow:code — Strict Serial Task Loop

Phase: 3: Implementation Agent type: Main agent (coordinator), dispatches sub-agents per task State: CODE_LOOP in .state.json

One task at a time. Each task is a complete cycle: pickup → branch → TDD → verify → review → PR/merge → close. No task starts until the previous one is merged.

┌──────────────────────────────────────────────────────┐
│  while bd ready --assignee $ACTOR has tasks:          │
│                                                       │
│  1. PICKUP   bd ready → claim → dolt push             │
│  2. BRANCH   git checkout -b feat/<id>-<slug>         │
│  3. TDD      RED gate pause → GREEN → REFACTOR        │
│  4. VERIFY   tests + coverage + bd preflight          │
│  5. REVIEW   c4flow:review (CRITICAL/HIGH blocks)     │
│  6. PR/MERGE c4flow:pr → merge to main                │
│  7. CLOSE    bd close --reason → dolt push            │
│                                                       │
│  no tasks left → advance currentState to INFRA        │
└──────────────────────────────────────────────────────┘

---

Step 0: Read State and Resolve Actor

Read .state.json

STATE=$(cat docs/c4flow/.state.json 2>/dev/null)
CURRENT_STATE=$(echo "$STATE" | jq -r '.currentState // empty')
EPIC_ID=$(echo "$STATE" | jq -r '.beadsEpic // empty')
FEATURE_SLUG=$(echo "$STATE" | jq -r '.feature.slug // empty')
TASK_LOOP=$(echo "$STATE" | jq '.taskLoop // null')

Legacy migration: if currentState is "CODE", update it to "CODE_LOOP":

if [ "$CURRENT_STATE" = "CODE" ]; then
  jq '.currentState = "CODE_LOOP" | .taskLoop = null' \
    docs/c4flow/.state.json > docs/c4flow/.state.json.tmp \
    && mv -f docs/c4flow/.state.json.tmp docs/c4flow/.state.json
  CURRENT_STATE="CODE_LOOP"
fi

Abort if state is not CODE_LOOP:

[ "$CURRENT_STATE" != "CODE_LOOP" ] && echo "ERROR: currentState is $CURRENT_STATE, expected CODE_LOOP" && exit 1

Resolve Actor

Actor is used for both bd ready --assignee filtering and bd update --claim.

Priority order:

1. Explicit arg — parse skill invocation args for --assignee <name> or from <name> / pickup task from <name>
2. BD_ACTOR env var
3. git config user.name — always available, used as final fallback

# 1. Parse explicit assignee from args (passed in from orchestrator)
ACTOR=$(echo "$SKILL_ARGS" | grep -oP '(?<=--assignee |from )["\x27]?[A-Za-z0-9 _-]+["\x27]?' | tr -d "\"'" | head -1)

# 2. BD_ACTOR fallback
ACTOR=${ACTOR:-$BD_ACTOR}

# 3. git config fallback
ACTOR=${ACTOR:-$(git config user.name)}

echo "Actor: $ACTOR"

Check for Resume

If taskLoop is non-null, a task was in progress when the session ended. Resume from saved subState:

SUBTASK_ID=$(echo "$TASK_LOOP" | jq -r '.currentTaskId // empty')
SUBSTATE=$(echo "$TASK_LOOP" | jq -r '.subState // empty')

if [ -n "$SUBTASK_ID" ] && [ -n "$SUBSTATE" ]; then
  echo "Resuming task $SUBTASK_ID from subState $SUBSTATE"
  # Run Step 0.5 (Pre-flight Sync) first, unless subState == "CLOSING"
  # Then jump to the right sub-state — see: Resume Logic section below
fi

---

Step 0.5: Pre-flight Sync

Skip this step if resuming from taskLoop.subState == "CLOSING" — go directly to Step 7.

Invoke c4flow:sync. This syncs both Dolt beads from DoltHub and git from GitHub origin.

On success: → Proceed to Step 1: PICKUP

On failure (any sync error): → Show the error output verbatim → Ask the user:

Sync failed. How do you want to proceed?
[continue] Use local data (may be stale — tasks you see may not reflect latest team activity)
[stop]     Exit now to fix the sync issue

→ If user chooses [continue]: note the sync failure in the session, proceed to Step 1 → If user chooses [stop]: exit the skill immediately, leave .state.json unchanged

---

Step 1: PICKUP — Find and Claim Task

Discover unblocked tasks

READY_JSON=$(bd ready --assignee "$ACTOR" --json 2>/dev/null)
READY_COUNT=$(echo "$READY_JSON" | jq 'length')

If READY_COUNT is 0:

No unblocked tasks assigned to <actor>.
All tasks may be blocked, in progress, or complete.
Check status with: bd dep tree <epic-id>

→ If all tasks closed, advance to INFRA (see Completion Gate). → Otherwise, show blocked tasks and wait for user.

If READY_COUNT > 1 — show list and ask user to confirm or select:

Ready tasks for <actor>:
  [1] bd-a1b2  [P1] "Add login endpoint"
  [2] bd-c3d4  [P2] "Add JWT validation"
  [3] bd-e5f6  [P2] "Add rate limiting"

Pick a task (default: 1, the highest priority unblocked task):

Claim and set in-progress

TASK_ID=<selected-id>

# 1. Claim the task (assigns to current actor)
CLAIM_RESULT=$(bd update "$TASK_ID" --claim --json 2>&1)
CLAIM_OK=$(echo "$CLAIM_RESULT" | jq -r '.status // empty' 2>/dev/null)

if [ "$CLAIM_OK" != "ok" ]; then
  echo "Task $TASK_ID is already claimed. Re-running bd ready..."
  # Loop back to discovery — do NOT proceed
fi

# 2. Move from "active" → "in-progress" immediately
bd update "$TASK_ID" --status "in-progress" --json 2>&1

Critical: Status must be set to in-progress before any coding starts — not at close time. This makes task progress visible to the team immediately.

Sync to DoltHub immediately

bd dolt commit -m "claim: task $TASK_ID in_progress" 2>/dev/null
DOLT_RESULT=$(bd dolt push 2>&1)
if [ $? -eq 0 ]; then
  echo "Synced: task $TASK_ID is in-progress, visible to team on DoltHub"
else
  echo "WARNING: Dolt push failed. Task is claimed locally but team may not see it yet."
  echo "Run 'bd dolt push' manually when network is available."
fi

Write taskLoop to .state.json

TASK_TITLE=$(echo "$READY_JSON" | jq -r --arg id "$TASK_ID" '.[] | select(.id == $id) | .title')
TASK_SLUG=$(echo "$TASK_TITLE" | tr '[:upper:]' '[:lower:]' | sed 's/[^a-z0-9]/-/g' | sed 's/--*/-/g' | sed 's/^-\|-$//g')
BRANCH_NAME="feat/${TASK_ID}-${TASK_SLUG}"

jq --arg id "$TASK_ID" \
   --arg slug "$TASK_SLUG" \
   --arg branch "$BRANCH_NAME" \
   '.taskLoop = {
     currentTaskId: $id,
     currentTaskSlug: $slug,
     currentBranch: $branch,
     subState: "CODING",
     completedTasks: (.taskLoop.completedTasks // []),
     failedTasks: (.taskLoop.failedTasks // [])
   }' docs/c4flow/.state.json > docs/c4flow/.state.json.tmp \
  && mv -f docs/c4flow/.state.json.tmp docs/c4flow/.state.json

---

Step 2: BRANCH — Create Task Branch

Always cut from latest main

git checkout main 2>&1
PULL_RESULT=$(git pull 2>&1)

if echo "$PULL_RESULT" | grep -qi "error\|conflict\|fatal"; then
  echo "BLOCKED: git pull failed — $PULL_RESULT"
  echo "Fix network/conflict issue, then resume."
  # Update subState to reflect BLOCKED
  jq '.taskLoop.subState = "BLOCKED"' \
    docs/c4flow/.state.json > docs/c4flow/.state.json.tmp \
    && mv -f docs/c4flow/.state.json.tmp docs/c4flow/.state.json
  exit 1
fi

Check for existing branch

if git show-ref --verify --quiet "refs/heads/$BRANCH_NAME"; then
  echo "Branch $BRANCH_NAME already exists."
  echo "Options: [resume] continue from existing branch | [recreate] delete and start fresh"
  # Wait for user input
fi

Create branch

git checkout -b "$BRANCH_NAME"
echo "Branch: $BRANCH_NAME"

---

Step 3: TDD — Red Gate Sub-agent

Dispatch a TDD sub-agent. Provide full task context — do NOT make the sub-agent read files.

Read task context before dispatch

TASK_DETAIL=$(bd show "$TASK_ID" --json 2>/dev/null)
TASK_DESC=$(echo "$TASK_DETAIL" | jq -r '.description // empty')
TASK_CRITERIA=$(echo "$TASK_DETAIL" | jq -r '.acceptance // empty')
TASK_SPEC_REF=$(echo "$TASK_DETAIL" | jq -r '.notes // empty' | grep -i "spec ref" | head -1)

Sub-agent prompt template

Use skills/code/implementer-prompt.md as base, with TDD additions:

You are the TDD implementer for task: <task-title>

## Task
ID: <task-id>
Title: <task-title>
Description: <task-description>
Acceptance Criteria: <acceptance-criteria>
Spec Reference: <spec-ref>

## Feature Context
Feature: <feature-name>
Branch: <branch-name>
Spec: docs/specs/<feature-slug>/spec.md
Design: docs/specs/<feature-slug>/design.md
Tech Stack: docs/specs/<feature-slug>/tech-stack.md

## TDD Instructions

### Phase 1: RED (MANDATORY — do not skip)
1. Read the acceptance criteria carefully
2. Write the test file FIRST — no implementation code yet
3. Run the test suite: the new test MUST fail
4. If the test passes immediately → STOP and report TRIVIAL_TEST
   (this means implementation already exists, or test doesn't test anything real)
5. Report back with:
   - Test file path and content
   - Exact failure output
   - Confirmation: "RED state confirmed"

### Phase 2: Wait for RED gate approval
→ You will receive approval or adjustment instructions before proceeding.
→ Do NOT write implementation until approved.

### Phase 3: GREEN (after approval)
1. Write the minimum code needed to make the failing test pass
2. Run the full test suite — ALL tests (old + new) must pass
3. If existing tests break → fix the regression before proceeding
4. Report: tests passing, coverage delta

### Phase 4: REFACTOR
1. Review your implementation for: naming, duplication, complexity
2. Refactor if needed — keep it clean
3. Run tests again — must still pass
4. Report: DONE with summary of files changed, tests added

## Output Format
Use exactly these status codes:
- RED_CONFIRMED: <test file> <failure output>
- TRIVIAL_TEST: test passed immediately — something is wrong
- GREEN_DONE: all tests pass
- DONE: refactor complete, ready for verify
- NEEDS_CONTEXT: <question>
- BLOCKED: <reason>

RED gate pause

After sub-agent reports RED_CONFIRMED, pause and show the user:

── RED STATE CONFIRMED ─────────────────────────────────────
Task: <task-title> (<task-id>)
Branch: <branch-name>

Test: <test-file-path>
<test-content-snippet>

Failure output:
<failure-output>

This test fails because the implementation doesn't exist yet.
Does this test correctly capture the requirement?

[yes] Proceed to implementation
[adjust] Revise the test — describe what to change
────────────────────────────────────────────────────────────

If user says adjust: pass adjustment instructions back to the sub-agent, wait for new RED_CONFIRMED, pause again.

If sub-agent reports TRIVIAL_TEST:

WARNING: Test passed immediately without implementation.
Either the code already exists, or the test is too trivial.
Sub-agent will revise — please review the new test carefully.

Once user approves RED: send approval to sub-agent → it proceeds to GREEN → REFACTOR → reports DONE.

Update subState during TDD

# After RED confirmed (waiting for approval)
jq '.taskLoop.subState = "CODING"' \
  docs/c4flow/.state.json > docs/c4flow/.state.json.tmp \
  && mv -f docs/c4flow/.state.json.tmp docs/c4flow/.state.json

---

Step 4: VERIFY — Tests + Preflight

Update subState:

jq '.taskLoop.subState = "VERIFYING"' \
  docs/c4flow/.state.json > docs/c4flow/.state.json.tmp \
  && mv -f docs/c4flow/.state.json.tmp docs/c4flow/.state.json

Run test suite + coverage

Read coverage threshold from tech stack (default 80%):

THRESHOLD=$(grep -i "coverage" "docs/specs/$FEATURE_SLUG/tech-stack.md" 2>/dev/null \
  | grep -oP '\d+(?=%)' | head -1)
THRESHOLD=${THRESHOLD:-80}

Run tests and check coverage. If coverage < threshold:

Coverage is <X>%. Threshold is <THRESHOLD>%.
Options:
  [add-tests]  Add more tests to reach threshold (routes back to TDD sub-agent)
  [proceed]    Proceed anyway (recorded as known gap)

If any tests fail → set subState back to CODING, re-dispatch TDD sub-agent with failure context.

bd preflight check

PREFLIGHT=$(bd preflight --check --json 2>/dev/null)
PREFLIGHT_PASS=$(echo "$PREFLIGHT" | jq '.pass // false')

If preflight fails → route back to CODING:

echo "Preflight failed. Failing checks:"
echo "$PREFLIGHT" | jq -r '.issues[]?.message' 2>/dev/null
echo "Routing back to CODING sub-state to fix preflight issues."
jq '.taskLoop.subState = "CODING"' \
  docs/c4flow/.state.json > docs/c4flow/.state.json.tmp \
  && mv -f docs/c4flow/.state.json.tmp docs/c4flow/.state.json
# Re-dispatch TDD sub-agent with preflight issues as context

---

Step 5: REVIEW — Per-task Code Review

Update subState:

jq '.taskLoop.subState = "REVIEWING"' \
  docs/c4flow/.state.json > docs/c4flow/.state.json.tmp \
  && mv -f docs/c4flow/.state.json.tmp docs/c4flow/.state.json

Dispatch c4flow:review sub-agent for the current task branch diff vs main.

Severity routing

| Severity | Action | |----------|--------| | CRITICAL | Block — fix required. Route back to CODING | | HIGH | Block — fix required. Route back to CODING | | MEDIUM | Advisory — surface to user, proceed to PR | | LOW | Advisory — surface to user, proceed to PR | | Clean | Proceed directly to PR |

If CRITICAL/HIGH findings:

echo "Review blocked by CRITICAL/HIGH findings. Routing back to TDD sub-agent."
jq '.taskLoop.subState = "CODING"' \
  docs/c4flow/.state.json > docs/c4flow/.state.json.tmp \
  && mv -f docs/c4flow/.state.json.tmp docs/c4flow/.state.json
# Dispatch implementer sub-agent with review findings as context

---

Step 6: PR + MERGE

After review passes, invoke c4flow:pr for the task branch:

# c4flow:pr creates the PR for the current branch
# After PR is created and merged to main, capture PR number
PR_NUMBER=<returned-by-c4flow:pr>
MERGED_AT=$(date -u +"%Y-%m-%dT%H:%M:%SZ")

---

Step 7: CLOSE + SYNC

Resume entry point for subState == "CLOSING" — if arriving here via resume (Step 0.5 was skipped), start from the bd close command below.

Update subState to CLOSING:

jq '.taskLoop.subState = "CLOSING"' \
  docs/c4flow/.state.json > docs/c4flow/.state.json.tmp \
  && mv -f docs/c4flow/.state.json.tmp docs/c4flow/.state.json

Close the bead

bd close "$TASK_ID" --reason "Implemented: $TASK_TITLE. PR: #$PR_NUMBER. Branch: $BRANCH_NAME merged to main."

Sync to DoltHub

bd dolt commit -m "close: task $TASK_ID done" 2>/dev/null
bd dolt push 2>/dev/null && echo "Synced: task $TASK_ID closed"

Record in completedTasks + clear current

jq --arg id "$TASK_ID" \
   --argjson pr "$PR_NUMBER" \
   --arg merged_at "$MERGED_AT" \
   '.taskLoop.completedTasks += [{ id: $id, pr: $pr, mergedAt: $merged_at }]
    | .taskLoop.currentTaskId = null
    | .taskLoop.currentTaskSlug = null
    | .taskLoop.currentBranch = null
    | .taskLoop.subState = null' \
  docs/c4flow/.state.json > docs/c4flow/.state.json.tmp \
  && mv -f docs/c4flow/.state.json.tmp docs/c4flow/.state.json

---

Completion Gate → Advance to INFRA

After closing each task, check if more work remains:

REMAINING=$(bd ready --assignee "$ACTOR" --json 2>/dev/null | jq 'length')

If REMAINING > 0: run inter-iteration beads sync before the next pickup:

# Inter-iteration sync: pull latest bead states to prevent stale claims
bd dolt pull 2>/dev/null && echo "Beads synced before next pickup" \
  || echo "WARNING: Beads sync failed — using local data"

Then loop back to Step 1 (PICKUP). Step 0.5 (full Dolt+Git sync) runs once per session on entry; inter-iteration sync only refreshes bead state from DoltHub.

If REMAINING == 0, verify all epic tasks are closed:

OPEN_COUNT=$(bd list --json 2>/dev/null | \
  jq --arg epic "$EPIC_ID" '[.[] | select(.status != "closed")] | length')

If OPEN_COUNT == 0:

# All tasks done — advance to DEPLOY
jq '
  .currentState = "INFRA"
  | .completedStates += ["CODE_LOOP"]
  | .taskLoop = null
  | .failedAttempts = 0
  | .lastError = null
' docs/c4flow/.state.json > docs/c4flow/.state.json.tmp \
  && mv -f docs/c4flow/.state.json.tmp docs/c4flow/.state.json

echo "CODE_LOOP complete. All tasks closed. Advancing to INFRA."

If OPEN_COUNT > 0 but REMAINING == 0:

Some tasks exist but none are ready for <actor>.
Possible: tasks blocked on external dependencies, or assigned to other people.

Show blocked tasks and ask user how to proceed.

---

Resume Logic

On entry to CODE_LOOP skill, if taskLoop.subState is non-null:

Step 0.5 (Pre-flight Sync) runs for all resume states except CLOSING. Run Step 0.5 first, then jump to the resume action below.

| subState | Run Step 0.5? | Resume action | |----------|--------------|---------------| | CODING | ✅ Yes | Ask user: "Re-run TDD from RED, or continue from where you left off?" | | VERIFYING | ✅ Yes | Re-run test suite + bd preflight (skip TDD) | | REVIEWING | ✅ Yes | Re-dispatch c4flow:review (skip TDD + verify) | | CLOSING | ❌ Skip | Re-run bd close + bd dolt push (skip everything else) | | BLOCKED | ✅ Yes | Show block reason, wait for user to resolve |

Always confirm branch checkout before resuming:

git checkout "$CURRENT_BRANCH" 2>/dev/null || echo "WARNING: branch not found — may need to recreate"

---

Rules

Never:

• Start on main — always create task branch first
• Skip the RED gate pause — human approval is required before GREEN
• Run multiple task loops in parallel — this is a serial loop
• Proceed when preflight fails — fix and re-verify
• Proceed when CRITICAL/HIGH review findings exist — fix first
• Use bd list --status open for task discovery — use bd ready (blocker-aware)
• Forget --reason on bd close — audit trail is required

Always:

• Resolve actor before any bd command
• git pull before branching
• Sync (bd dolt push) at claim and at close
• Write subState to .state.json at every transition
• Provide full task text to sub-agents — don't make them read files

---

Prompt Templates

• skills/code/implementer-prompt.md — base TDD sub-agent template
• skills/code/spec-reviewer-prompt.md — spec compliance reviewer
• skills/code/code-quality-reviewer-prompt.md — quality reviewer

---

.state.json taskLoop Schema

{
  "currentState": "CODE_LOOP",
  "taskLoop": {
    "currentTaskId": "bd-a1b2",
    "currentTaskSlug": "add-login-endpoint",
    "currentBranch": "feat/bd-a1b2-add-login-endpoint",
    "subState": "CODING",
    "completedTasks": [
      { "id": "bd-0001", "pr": 12, "mergedAt": "2026-03-19T10:00:00Z" }
    ],
    "failedTasks": []
  }
}

subState values: CODING | VERIFYING | REVIEWING | CLOSING | BLOCKED