coalesce-labs/implement-plan

Implement Plan

You are tasked with implementing an approved technical plan from thoughts/shared/plans/. These plans contain phases with specific changes and success criteria.

Prerequisites

# Check project setup (thoughts, CLAUDE.md snippet, config)
if [[ -f "${CLAUDE_PLUGIN_ROOT}/scripts/check-project-setup.sh" ]]; then
  "${CLAUDE_PLUGIN_ROOT}/scripts/check-project-setup.sh" || exit 1
fi

# Auto-discover most recent plan (workflow context + filesystem fallback)
RECENT_PLAN=""
if [[ -f "${CLAUDE_PLUGIN_ROOT}/scripts/workflow-context.sh" ]]; then
  RECENT_PLAN=$("${CLAUDE_PLUGIN_ROOT}/scripts/workflow-context.sh" recent plans)
fi
if [[ -n "$RECENT_PLAN" ]]; then
  echo "📋 Auto-discovered recent plan: $RECENT_PLAN"
else
  echo "⚠️ No recent plan found in workflow context or filesystem"
fi

Session Tracking

SESSION_SCRIPT="${CLAUDE_PLUGIN_ROOT}/scripts/catalyst-session.sh"
if [[ -x "$SESSION_SCRIPT" ]]; then
  CATALYST_SESSION_ID=$("$SESSION_SCRIPT" start --skill "implement-plan" \
    --ticket "${TICKET_ID:-}" \
    --workflow "${CATALYST_SESSION_ID:-}")
  export CATALYST_SESSION_ID
  "$SESSION_SCRIPT" phase "$CATALYST_SESSION_ID" "implementing" --phase 1
fi

Initial Response

Auto-discovery has already run in Prerequisites above. Check its output and follow this priority:

1. If user provided a plan path as parameter: Use the provided path (user override). Skip to Step 3.

2. If no parameter provided AND Prerequisites output shows a discovered plan (📋):

   • Show user the discovered plan path
   • Ask: "Proceed with this plan? [Y/n]"
   • If yes: use it and skip to Step 3
   • If no: proceed to option 3

3. If no parameter AND Prerequisites shows no plan found (⚠️):

   • List available plans from thoughts/shared/plans/
   • Show most recent 5 plans with dates and ticket numbers
   • Ask user which plan to implement
   • Wait for user input with plan path

STEP 3: Read and prepare

Once you have a plan path:

• Read the plan completely (no limit/offset)
• Check for any existing checkmarks (- [x]) to see what's done
• Read the original ticket and all files mentioned in the plan
• Extract ticket from plan frontmatter (source_ticket field) and update Linear state to stateMap.inProgress from config using Linearis CLI (run linearis issues usage for syntax). If Linearis CLI is not available, skip silently and continue implementation. Skip the status transition when CATALYST_PHASE is set — under a phase agent (e.g. phase-implement, or this skill invoked as a sub-task from phase-pr / phase-monitor-merge during PR resolution / CI fix-up loops) the deterministic coordinator (CTL-558) owns the Linear status write-back. A direct write here would regress the ticket from PR back to Implement, producing operator-visible state flicker (CTL-601). Mirrors the gate in create-pr/SKILL.md:227-232.
• Think deeply about how the pieces fit together
• Create a todo list to track your progress
• Start implementing if you understand what needs to be done

Implementation Philosophy

Plans are carefully designed, but reality can be messy. Your job is to:

• Follow TDD: write tests before implementation code in each phase
• Follow the plan's intent while adapting to what you find
• Implement each phase fully before moving to the next
• Verify your work makes sense in the broader codebase context
• Update checkboxes in the plan as you complete sections

TDD Rhythm Per Phase

For each phase, follow Red → Green → Refactor:

1. Red — Write the tests specified in the plan's "Tests First" section. Run them to confirm they fail.
2. Green — Implement the minimum code from the plan's "Implementation" section to make tests pass.
3. Refactor — Clean up while keeping tests green. Apply any refactoring notes from the plan.

This order is non-negotiable. If a phase doesn't have a "Tests First" section, write tests for the phase's expected behavior before implementing. The tests serve as executable acceptance criteria.

When things don't match the plan exactly, think about why and communicate clearly. The plan is your guide, but your judgment matters too.

If you encounter a mismatch:

• STOP and think deeply about why the plan can't be followed

• Present the issue clearly:

  Issue in Phase [N]:
  Expected: [what the plan says]
  Found: [actual situation]
  Why this matters: [explanation]
  
  How should I proceed?
  

Verification Approach

Within each phase (TDD cycle):

• Write tests first → run them → confirm they fail (Red)
• Write implementation → run tests → confirm they pass (Green)
• Refactor if needed → run tests → confirm they still pass (Refactor)

After completing a phase:

• Run the full success criteria checks (usually make check test covers everything)
• Fix any issues before proceeding
• Update your progress in both the plan and your todos
• Check off completed items in the plan file itself using Edit
• Check context usage - monitor token consumption
• Push + ensure the draft PR (phase-agent mode) — In phase-agent mode the draft PR opens at the first phase commit (see the implement-plan-draft-pr-early fence); interactive /catalyst-dev:implement-plan runs skip this automatically via the CATALYST_PHASE gate.

# CTL-783: make the PR the durable off-disk work record from the FIRST commit.
# Run after EVERY plan-phase commit: first run opens the draft PR, later runs
# just push (draft_pr_ensure is idempotent). Interactive runs (no
# CATALYST_PHASE) skip — no surprise pushes. Fail-open: never blocks the phase.
if [[ -n "${CATALYST_PHASE:-}" && -r "${CLAUDE_PLUGIN_ROOT}/scripts/lib/draft-pr.sh" ]]; then
  # shellcheck source=/dev/null
  source "${CLAUDE_PLUGIN_ROOT}/scripts/lib/draft-pr.sh"
  if [[ "$(draft_pr_enabled)" == "true" ]]; then
    draft_pr_push || true
    draft_pr_ensure "main" "${TICKET_ID:-${CATALYST_TICKET:-}}" >/dev/null 2>&1 || true
  fi
fi

Don't let verification interrupt your flow - batch full suite runs at natural stopping points. But always run the specific tests you wrote during each Red → Green cycle.

Context Management During Implementation

Monitor context proactively throughout implementation:

After Each Phase:

✅ Phase {N} complete!

## 📊 Context Status
Current usage: {X}% ({Y}K/{Z}K tokens)

{If >60%}:
⚠️ **Context Alert**: We're at {X}% usage.

**Recommendation**: Create a handoff before continuing to Phase {N+1}.

**Why?** Implementation accumulates context:
- File reads
- Code changes
- Test outputs
- Error messages
- Context clears ensure continued high performance

**Options**:
1. ✅ Create handoff and clear context (recommended)
   - Use `/create-handoff` to generate properly formatted handoff
   - Format: `thoughts/shared/handoffs/{ticket}/YYYY-MM-DD_HH-MM-SS_description.md`
   - Includes timestamp for lexical sorting by recency
2. Continue to next phase (if close to completion)

**To resume**: Start fresh session, run `/implement-plan {plan-path}`
(The plan file tracks progress with checkboxes - you'll resume automatically)

{If <60%}:
✅ Context healthy. Ready for Phase {N+1}.

When to Warn:

• After any phase if context >60%
• If context >70%, strongly recommend handoff
• If context >80%, STOP and require handoff
• If user is spinning on errors (3+ attempts), suggest context clear

Educate About Phase-Based Context:

• Explain that implementation is designed to work in chunks
• Each phase completion is a natural handoff point
• Plan file preserves progress across sessions
• Fresh context = fresh perspective on next phase

Creating a Handoff:

When recommending a handoff, guide the user:

1. Offer to create the handoff using /create-handoff
2. Or create a manual handoff following the timestamp convention
3. Handoff filename format: thoughts/shared/handoffs/{ticket}/YYYY-MM-DD_HH-MM-SS_description.md
4. Include: completed phases, next steps, key learnings, file references
5. Update plan file with checkboxes for completed work

Quality Gates (After All Phases Complete)

After all implementation phases pass, run quality gates before marking work as done. These gates catch issues that per-phase testing might miss.

Gate execution order:

Quality Gates:
├── 1. /validate-type-safety  → tsc + reward hacking scan + tsconfig check + tests + lint
├── 2. /security-review       → scan for security vulnerabilities (built-in Claude Code skill)
├── 3. code-reviewer agent    → style/guideline adherence check
└── 4. pr-test-analyzer agent → test coverage verification

Running the Gates

Gate 1: Type Safety Validation

Invoke /validate-type-safety. This runs the full 5-step gate (type check, reward hacking scan, test inclusion, tests, lint). If it fails, fix issues and re-run before proceeding.

Gate 2: Security Review

Invoke the built-in /security-review skill. Review findings and fix any vulnerabilities before proceeding.

Gate 3: Code Review

Spawn the code-reviewer agent:

Agent(subagent_type="pr-review-toolkit:code-reviewer",
      prompt="Review the uncommitted changes for adherence to project guidelines and style.")

Address any findings that violate project conventions.

Gate 4: Test Coverage

Spawn the pr-test-analyzer agent:

Agent(subagent_type="pr-review-toolkit:pr-test-analyzer",
      prompt="Analyze test coverage for the uncommitted changes. Identify critical gaps.")

If critical gaps exist, write the missing tests.

File Improvement Findings

Recording findings during implementation. When a phase surfaces friction worth fixing — a bug noticed in adjacent code, a step that shouldn't need manual intervention, a gap in tooling — record it the moment it's observed:

"${CLAUDE_PLUGIN_ROOT}/scripts/add-finding.sh" \
  --title "Short imperative title" \
  --body "Reproduction + expected + observed + any links" \
  --skill implement-plan

Findings go to a shared queue (under orchestrate/oneshot, that skill's queue; direct invocations get a per-session queue). The block below files the queue at end-of-run. It's a safety net: when implement-plan runs under /orchestrate or /oneshot, the parent's filing step drains the same queue first and this block finds an empty file:

FEEDBACK="${CLAUDE_PLUGIN_ROOT}/scripts/file-feedback.sh"
CONSENT="${CLAUDE_PLUGIN_ROOT}/scripts/feedback-consent.sh"
FINDINGS_FILE="${CATALYST_FINDINGS_FILE:-.catalyst/findings/${CATALYST_SESSION_ID:-current}.jsonl}"

if [ -x "$FEEDBACK" ] && [ -f "$FINDINGS_FILE" ] && [ -s "$FINDINGS_FILE" ]; then
  COUNT=$(wc -l < "$FINDINGS_FILE" | tr -d ' ')
  if [ "$("$CONSENT" check)" != "granted" ] && [ -z "${CATALYST_AUTONOMOUS:-}" ] && [ -t 0 ]; then
    read -r -p "File $COUNT improvement tickets now? [Y/n] " yn
    case "$yn" in [Nn]*) : ;; *) "$CONSENT" grant >/dev/null ;; esac
  fi
  if [ "$("$CONSENT" check)" = "granted" ]; then
    FILED=0
    while IFS= read -r line; do
      TITLE=$(jq -r '.title' <<<"$line")
      BODY=$(jq -r '.body' <<<"$line")
      SKILL=$(jq -r '.skill // "implement-plan"' <<<"$line")
      RESULT=$("$FEEDBACK" --title "$TITLE" --body "$BODY" --skill "$SKILL" --json 2>/dev/null || true)
      STATUS=$(jq -r '.status // "failed"' <<<"$RESULT")
      if [ "$STATUS" = "filed" ]; then
        ID=$(jq -r '.identifier // .url // ""' <<<"$RESULT")
        echo "  filed: $ID  ($TITLE)"
        FILED=$((FILED + 1))
      fi
    done < "$FINDINGS_FILE"
    [ "$FILED" -eq "$COUNT" ] && rm -f "$FINDINGS_FILE"
  fi
fi

End Session Tracking

After all quality gates pass (or are skipped), end the session:

if [[ -n "${CATALYST_SESSION_ID:-}" && -x "$SESSION_SCRIPT" ]]; then
  "$SESSION_SCRIPT" end "$CATALYST_SESSION_ID" --status done
fi

Autofix Behavior

For gates 1 and 2, attempt to fix issues automatically and re-run the gate. For gates 3 and 4, address findings and verify. If a gate fails after 2 fix attempts, report the remaining issues to the user and ask how to proceed.

Skipping Quality Gates

If the plan or user specifies --skip-quality-gates, skip this section entirely. Report that quality gates were skipped in the completion summary.

If You Get Stuck

When something isn't working as expected:

• First, make sure you've read and understood all the relevant code
• Consider if the codebase has evolved since the plan was written
• Present the mismatch clearly and ask for guidance

Use sub-tasks sparingly - mainly for targeted debugging or exploring unfamiliar territory.

Resuming Work

If the plan has existing checkmarks:

• Trust that completed work is done
• Pick up from the first unchecked item
• Verify previous work only if something seems off

Remember: You're implementing a solution, not just checking boxes. Keep the end goal in mind and maintain forward momentum.

Agent Team Mode (Optional)

When invoked with --team flag or when the plan spans 3+ independent domains:

When to Use Team Mode

• Plan has phases that can be implemented in parallel
• Changes span distinct domains (frontend, backend, tests, infra)
• Each domain's changes don't overlap in files

Team Structure

Lead (Opus) — Coordinates implementation
├── Teammate 1 (Sonnet) — Frontend changes
│   └── Can spawn subagents for research
├── Teammate 2 (Sonnet) — Backend changes
│   └── Can spawn subagents for research
└── Teammate 3 (Sonnet) — Test changes
    └── Can spawn subagents for research

Process

1. Analyze plan phases — Identify which phases can be parallelized
2. Assign file ownership — Each teammate gets distinct files (no overlap)
3. Create task list — Use TaskCreate with dependencies between phases
4. Launch team — Spawn teammates with focused instructions
5. Review gates — Lead reviews teammate work via approvePlan/rejectPlan
6. Integration — Lead verifies all changes work together
7. Commit — Single atomic commit or per-phase commits

Important Constraints

• File ownership is strict — no two teammates edit the same file
• Sequential phases stay sequential — only parallelize truly independent work
• Lead reviews all code — use plan approval gates before proceeding
• Fallback gracefully — if agent teams unavailable, execute sequentially

Linear Integration

If a ticket is detected (from plan document's source_ticket frontmatter or from context):

• At implementation start (Step 3): Update ticket status to stateMap.inProgress from config using Linearis CLI (run linearis issues usage for syntax).
• Skip the status transition when CATALYST_PHASE is set — the deterministic coordinator (CTL-558) owns Linear write-back under phase agents. See the gate at Step 3 above for details. CTL-601 — without this gate, invoking this skill as a sub-task from another phase agent (typical in phase-pr / phase-monitor-merge resolution loops) regresses the ticket state to Implement and produces operator-visible flicker.
• If Linearis CLI not available, skip silently and continue implementation