Task Code

task-code workflow (v3)

Overview

2-phase sub-workflow mapping the user's CLAUDE.md Stage ③.b 子任务编码 discipline onto the harnessed runtime, fully harnessed.workflow.v3 schema (Phase v3.0-3.4 W0 T3.4.W0.7 — D-09 L0 Discipline Substrate + D-05 conditional invokes_tools + D-15 planning-with-files plugin).

Per-phase config loads from workflows/task/code/workflow.yaml; engine.runRouting spawns each phase as a sub-agent via @anthropic-ai/claude-agent-sdk 0.3.142+.

Karpathy 4 心法 (L0 Discipline Substrate always-on)

Phase 01-code 的 upstream 是 karpathy — runtime engine 加载 workflows/disciplines/ karpathy.yaml discipline rules 应用 cross-cutting (Think Before Coding / Simplicity First / Surgical Changes / Goal-Driven Execution + ≤200L hard limit + no-feature-creep

• trust-internal-code + no-comments-default). 不 invoke slash cmd, 通过 hook 强制 behavioral rule per D-09 L0 Discipline Substrate.

mattpocock conditional route (D-05 invokes_tools)

Phase 01-code 按 phase fact context 条件性 fire 2 mattpocock 招式:

• improve-codebase-architecture — 周期架构健康审查 (when phase.architecture_health_audit == true)
• diagnose — bug 系统化排错 (when subtask.bug_root_cause_unknown == true)

2 触发条件 OR-chain, 任 1 fire 即 invoke 对应招式 — 不互斥 (sister CLAUDE.md "mattpocock 招式按需召唤" pattern, NOT exclusive). 无触发 = pure karpathy 心法 only.

CodeGraph navigation (opt-in, v4.17.0)

If the project has a .codegraph/ index (opt-in CodeGraph semantic index — sister capabilities.yaml codegraph entry), prefer the codegraph MCP tools (codegraph_explore) for symbol lookup / call paths / impact analysis instead of grep/glob/file-by-file Read crawling. No .codegraph/ → skip this (no install nag).

Phase 02-progress planning-with-files plugin 真接 (Q-AUDIT-5a LOCKED Option A)

02-progress invokes the Claude Code plugin slash command /plan to update progress.md in .planning/phases/<NN>-<slug>/ — 跟踪 subtask 完成 / blocked / next step per bundled "跨 session 恢复" 模式 + R20.6 Manus-style 持久化. Requires the planning-with-files Claude Code plugin (install via Claude Code plugin marketplace).

How to invoke

!harnessed checkpoint intent task-code

The banner above (when present) means this invocation is REGISTERED with the engine (an intent marker) — not yet compliant: the steps below (prompt → spawn → checkpoint complete) resolve it, and a per-turn <workflow-intent> reminder persists until they run.

The numbered sequence below is the state machine — execute it with Bash. Do NOT improvise an equivalent flow from the Overview above: freelancing bypasses the engine (no ledger, no evidence guard). harnessed gives you the spawn-ready prompt; YOU spawn the subagent with a CC-native Task / Agent tool (keeps the session responsive + lets clarification round-trips reach the user).

Do NOT pipe to harnessed run task-code — that is the CI/headless path (in-process SDK spawn that blocks the session inside Claude Code).

1. Bash: harnessed prompt task-code --task "$ARGUMENTS" --json → parse {prompt, max_iterations, model}.
2. Spawn a CC-native subagent (Task / Agent tool) with that prompt and model, then drive delivery with harnessed's own completion gate:
   • on return, write the subagent's final output to a file and run harnessed checkpoint complete task-code --result-file <path> — it is fail-closed on the declared artifacts, the TDD boundary, and the verbatim <promise>COMPLETE</promise>.
   • if it blocks, run harnessed checkpoint fail task-code --failing-tests <n> to record the attempt; it prints BUDGET-EXHAUSTED / NO-PROGRESS / BREAK-LOOP when a stop condition is reached.
   • respawn ONLY while none of those three has fired. Any one of them means stop: re-scope the subtask, fix the blocker, or escalate to the user. Never respawn past a stop directive.
3. If the output contains STATUS: NEEDS_CLARIFICATION + a question list: STOP, relay them verbatim via AskUserQuestion, append the answers to the spec, then re-spawn the same sub.
4. On <promise>COMPLETE</promise>: write the subagent’s final output to a file, then Bash harnessed checkpoint complete task-code --result-file <path> --summary "<one-line>". Fail-CLOSED — it blocks unless every declared artifacts_expected file exists, the TDD boundary passes (non-empty evidence / both the red and green sides present / the test file was not deleted), and the result carries a verbatim <promise>COMPLETE</promise> (or a structured COMPLETE status). --result <text> is the inline variant; --result-file wins and is quoting-safe on Windows. --force records an audited override (evidence_status=overridden) — it does not silently pass.
5. If the complete gate blocked: Bash harnessed checkpoint fail task-code --failing-tests <n> to record the attempt. It prints BUDGET-EXHAUSTED / NO-PROGRESS / BREAK-LOOP once a stop condition is reached. Respawn ONLY while none of those three has fired; any one of them means STOP — re-scope the subtask, fix the blocker, or escalate to the user.

References

• D-09 — L0 Discipline Substrate always-on (karpathy 心法 4 条 cross-cutting)
• D-05 — phase-level invokes_tools conditional tool fire
• D-15 + Q-AUDIT-5a — planning-with-files = Claude Code plugin slash cmd /plan
• D-02 — SKILL.md name: bare slash cmd (task-code NOT task/code) per ADR 0030
• workflows/disciplines/karpathy.yaml — 4 心法 + ≤200L hard limit 等 rules (L0 substrate)
• workflows/capabilities.yaml — improve-codebase-architecture / diagnose / planning-with-files entries
• workflows/defaults.yaml — ralph_max_iterations.task-code.* values (T3.4.W2.2 followup)
• docs/WORKFLOW.md — 4-stage workflow mermaid + Stage ③ Execute 章节