Execute Unplanned Work (Fast Path)

Task Checkout (if assigned): Before starting work on an assigned task, checkout via Paperclip API:

curl -H "Authorization: Bearer $PAPERCLIP_API_KEY" \
     -H "X-Paperclip-Run-Id: $PAPERCLIP_RUN_ID" \
     -X POST "$PAPERCLIP_API_URL/api/companies/$PAPERCLIP_COMPANY_ID/issues/$PAPERCLIP_TASK_ID/checkout"

⛔ MANDATORY WORKFLOW — NEVER SKIP ANY STEP

You MUST execute steps 1 through 6 IN ORDER. Do NOT jump to implementation. Your FIRST action must be reading documentation (Step 1), NOT editing code. If you skip Step 1 or Step 5, the workflow is BROKEN.

Input

<work_description> #$ARGUMENTS </work_description>

If empty, ask: "What would you like me to work on?"

Step 1: Research (BLOCKING — must complete before any code changes)

Your first tool calls MUST be Read calls to load documentation. Do NOT start implementing.

1. If the description is vague, ask one or two clarifying questions first.

2. Classify scope first (trivial vs non-trivial):

   • Trivial: <=2 files, no entities/migrations/endpoints/auth model changes.
   • If trivial, use lightweight path:
     • read only docs/solutions/patterns/critical-patterns.md (if exists) and directly relevant doc,
     • skip mandatory research-agent spawning,
     • keep focused implementation + verification evidence.
   • If non-trivial, follow full workflow below.

3. Load docs baseline via skill (REQUIRED):

   • Apply docs-baseline-loading skill.
   • Read mandatory baseline docs before implementation.
   • Create any missing canonical baseline docs immediately using the skill's required minimum sections.
   • Then read scope-specific docs:
     • docs/solutions/patterns/critical-patterns.md (if exists)
     • docs/modules/<module>.md (backend/module scope)
     • docs/features/<feature>.md (frontend/UI scope)
     • docs/infrastructure/<component>.md (infrastructure scope)
     • docs/runbooks/README.md (operational/deploy scope)
   • Search docs/solutions/ by feature keywords for known gotchas.
   • If any other expected doc file is missing: create it immediately with a useful baseline (not placeholders). Minimum sections:
     • Purpose (business scope)
     • Source of Truth Files (exact paths)
     • Current Implementation Snapshot (what exists now)
     • Planned/Upcoming Contract (only if plan exists; clearly marked as planned)
     • Invariants and Gotchas
     • Safe Change Checklist for Future AI Work
     • Related Plan and Docs

4. Check existing context:

   • docs/brainstorms/ — recent brainstorm for this feature?
   • docs/plans/ — existing plan? If work belongs to a plan phase, suggest pwf-work-plan instead.

5. Execute research agents (REQUIRED for non-trivial scope):

   Execute the following agents by reading and applying their instructions:

Step 2: Task List

1. Derive a task list — concrete, dependency-ordered.

   • Each task: bold name + file path + sub-bullets with method names, fields, classes.
   • No vague summaries. Every task must specify what to change in which file.

2. Self-validate: Review every task. Does it have a file path? Does it have specific method names or field names? If not, rewrite it.

3. Debug route detection: If this work is a bug/failure/regression fix:

   • First validate reproducibility with bug-reproduction-validator (../../agents/workflow/bug-reproduction-validator.md) when the report is ambiguous.
   • Then apply systematic-debugging skill (root-cause -> pattern -> hypothesis -> minimal fix) before implementing broad changes.
   • deep stack/source uncertainty -> systematic-debugging/root-cause-tracing.md
   • async timing/flaky behavior -> systematic-debugging/condition-based-waiting.md
   • regression hardening after fix -> systematic-debugging/defense-in-depth.md

Built-in capabilities (use as needed during execution):

• Operational policy source: ../../references/rules/operational-guardrails.md
• Project overrides (optional): docs/workflow/operational-overrides.md (if present, it overrides defaults from guardrails)
• Database access: Load DB vars from project .env file for database queries when applicable. Never display credentials.
• Context7: Use the Context7 MCP (resolve-library-id then query-docs) before implementing with external libraries.

Step 3: Execute

For each task:

• Mark in progress.
• Read referenced files. Follow project rules and patterns from docs read in Step 1.
• Implement.
• Mark completed.

⚠️ CRITICAL: TypeORM Migration Atomic Chain (when applicable)

Follow the migration chain defined in ../../references/rules/operational-guardrails.md (generate -> drift-check -> local run). Treat this as blocking. Do not continue other tasks until the chain succeeds.

After all tasks:

• TypeScript Validation — Run npm run validate (or tsc --noEmit if no validate script). Fix ALL type/lint errors.
• Build Only When Explicit — Full build (npm run build) only runs when explicitly requested by user or during deployment workflows.

Step 4: Quality Review

Only run if 5+ files changed or multiple repos touched. Otherwise skip to Step 5.

If triggered, execute review agents by reading and applying their instructions:

1. Read ../../references/review-agent-selection-mapping.md to select applicable agents based on changed scope.
2. For each selected agent, read its agent file and execute the review instructions.
3. You can read multiple agent files in parallel.

Address critical findings only. Informational findings are noted but don't block.

Step 5: Documentation Maintenance (MANDATORY — never skip)

This step is MANDATORY even for small changes.

Apply docs-maintenance-after-work skill and execute its full flow:

• always run doc-shepherd,
• run plan-sync when plan context exists,
• run specialized doc writers conditionally,
• run pattern-extractor when applicable,
• pass the documentation quality gate before Step 6.

Step 6: Finish

Summarize: what was implemented, files changed, any caveats.

Verification Evidence (MANDATORY before completion claims)

Before any "done/fixed/passing" claim, apply verification-before-completion skill and use the evidence format from ../../references/rules/operational-guardrails.md.

Include a dedicated Documentation updates subsection listing:

• docs files updated/created
• what concrete knowledge was added (source-of-truth files, invariants, checklists, contracts)
• any remaining doc gaps explicitly marked for follow-up

Suggest:

• Commit with ticket number
• pwf-review for full PR review
• pwf-doc-capture if a non-trivial bug was fixed
• deploy-lambda reminder if Lambda repos were touched

Conventions

• Follow canonical policy in ../../references/rules/operational-guardrails.md.
• Follow commit policy in ../../references/rules/commits.md.
• Use optional project overrides in docs/workflow/operational-overrides.md when present.

Next Recommended Skills

• pwf-work-light for future trivial/local-only changes
• pwf-review for a full multi-agent review pass
• pwf-commit-changes after review approval
• pwf-doc-capture when a reusable fix/pattern emerged
• finishing-a-development-branch when branch/worktree is ready to close

• repo-research-analyst (../../agents/research/repo-research-analyst.md)

  • Purpose: Maps file paths, existing patterns, rules, existing enums, migration state
  • Input: Feature description, affected modules
  • Output: Concrete file paths, existing patterns

• learnings-researcher (../../agents/research/learnings-researcher.md)

  • Purpose: Surfaces relevant solutions from docs/solutions/
  • Input: Feature description, technical domain
  • Output: Applicable patterns, previous solutions

Read both agent files and execute their instructions. You can read multiple files in parallel.