build

@CLAUDE.md

Build: $ARGUMENTS

If $ARGUMENTS is empty, check for recent /plan output in the conversation. If none found, ask the user what to build.

Cross-Harness Hook Routing

If the task touches Claude Code or GitHub Copilot CLI hook configuration, payloads, decisions, matchers, exit codes, timeouts, generated shims, or event translation:

1. Invoke Skill(skill="agent-harness-reference") before design or code.
2. Execute the change through Skill(skill="ai-agents-portability-campaign").
3. Use Skill(skill="ai-agents-generation-and-release") for generated mirrors.
4. Do not repeat vendor research unless the pinned source ledger is stale or the task explicitly requires a contract refresh.

Complexity Assessment

Before implementation, Task(subagent_type="analyst"): Read .claude/skills/analyze/references/engineering-complexity-tiers.md and the task description. Classify as Tier 1-5. Return: tier, rationale, and recommended oversight level. Use this to calibrate implementation approach:

• Tier 1-2: Implement directly. Async code review sufficient.
• Tier 3: Validate approach before coding. Active mentorship pattern (check in at milestones).
• Tier 4-5: Proof-of-concept first. Get design sign-off before full implementation.

Pre-Mortem (Risk Identification)

Before any code changes, invoke Skill(skill="pre-mortem") on the task as briefed. Capture the top 2-3 critical risks and their mitigations in the active plan or issue handoff. Risks surfaced by reviewers late in the cycle are usually knowable up front. A 5-minute pre-mortem is cheaper than a 10-round bot review.

Agent

Task(subagent_type="implementer"): You are a senior engineer. Discover the project's tech stack, coding patterns, and test conventions by reading the codebase. Build in thin vertical slices. Test-first when the project has tests. Commit atomically.

For each slice:

1. Read the spec AC for this slice. Every test must trace to an AC number from /spec output. Name the test test_<behavior> and include the AC identifier in the docstring or comment.
2. Understand the existing code patterns (read related files, check test conventions). If Serena is available, prefer it for symbolic search (canonical tool names: mcp__serena__find_symbol, mcp__serena__get_symbols_overview; some Claude harnesses surface the same tools under the plugin alias mcp__plugin_serena_serena__find_symbol / mcp__plugin_serena_serena__get_symbols_overview, so accept either when present). Otherwise fall back to Grep and Read for filesystem-level discovery. Serena is not guaranteed in every harness (fresh installs without MCP, copilot-cli runtime); the fallback keeps the slice executable across hosts.
3. Write the failing test first. This project has pytest 8+. TDD is unconditional. Never write code before a failing test exists. The test expresses the AC contract; code exists only to make it pass. Tests written after code confirm the code's behavior, not the spec's contract.
4. Write the minimum code to pass the test. Run the test, confirm it fails on the right assertion, then write code.
5. Refactor toward quality (cohesion, encapsulation, simplicity). Re-run the test.
6. Self-apply gate for detection tools. If this slice adds a guard, warning, or detector (hook, linter, threshold check), run it against the current branch NOW before committing. If it does not fire on conditions present in the branch, the threshold or detection logic is wrong. Fix the logic before step 7.
7. Commit with a conventional message. Each commit is one logical change. Test file and implementation file committed together.