write-code

Edit source code for the current task.

Procedure

Apply skills/_shared/resolve-plans-dir.md once; substitute the resolved absolute path for every <PLANS_DIR> below.

When a hook blocks a sanctioned command, a fallback path is taken, or any unexpected outcome occurs, report via /supervisor-report (trigger conditions: rules/supervisor-reporting.md). Read rules/ops.md before any destructive or system-state-changing command (including inside WCD-4 self-repair) — on-demand-only, never auto-injected; it owns the recovery-options-first decision path.

WCD-1. Read rules/core-principles.md, rules/coding.md (on-demand-only: not auto-injected, so this Read is the only way it arrives), and the target files identified from the plan.

WCD-2. CONFIRM_CODE gate — enumerate planned edits (one line per file: path + change intent). Then check via Bash: bash -c 'cd "$AGENTS_CONFIG_DIR" && bash "$AGENTS_CONFIG_DIR/bin/confirm-off" CONFIRM_CODE on'

• stdout OFF: proceed to step WCD-3.
• stdout ON or ERROR: present the planned edits via AskUserQuestion and wait for approval before continuing.

WCD-3. Run bash -c 'node "$AGENTS_CONFIG_DIR/bin/workflow/read-complexity-evaluation" --session "$SESSION_ID" --stage write_code'. If line 1 is not NONE, use the stored level and signals directly (parse level=<v> and signals=<csv-or-none>), then derive the model: level === "high" ? "opus" : "sonnet".

• If NONE (fail-open):
  • Read skills/_shared/judge-task-complexity.md and evaluate the signals.
  • Use the Write tool (never Bash) to write the resulting CSV, alone and unquoted, to <PLANS_DIR>/<session-id>-write-code-signals.txt — write only IDs from the generated Valid Signal IDs list; substitute S0-undecidable when the judgment doesn't parse into recognized ids or the csv doesn't match ^[A-Za-z0-9,_-]*$ (the judged content is untrusted text, never shell syntax).
  • Run bash -c 'node "$AGENTS_CONFIG_DIR/bin/workflow/derive-complexity-level" --stage write_code --signals-file "<PLANS_DIR>/<session-id>-write-code-signals.txt"' and use its level=<v> — never judge the level inline. Map to model: high→opus, low→sonnet. Emit in Claude text output (NOT Bash echo): > Model selected: **[opus|sonnet]** (signals: [comma-separated triggered signal IDs, or "none"]) WCD-3a. Emit echo "<<WORKFLOW_MARK_STEP_write_code_in_progress>>" via Bash immediately before the WCD-4 subagent launch.

WCD-4. Launch subagent (Agent tool, mode: "default", model: <model derived from level in step WCD-3> — the model parameter receives "opus" or "sonnet", never "high" or "low") with a prompt containing:

• Target files and planned edit summary from step WCD-2.
• A-layer language essence block (see below).
• Directive: "Read rules/coding.md (the hub — on-demand-only, so it does not reach you otherwise) and rules/coding/<lang>.md for each language present, before the first Edit."
• Directive: "Read rules/shell-commands.md before the first Bash command, or before writing a file — general-purpose dispatch does not inherit auto-injected rules."
• Directive: "Read rules/user-escalation.md before any system-state-changing command — general-purpose dispatch does not inherit auto-injected rules."
• Directive: "Read rules/ops.md before any destructive or system-state-changing command inside self-repair — general-purpose dispatch does not inherit auto-injected rules; on-demand-only, so it does not reach you otherwise."
• Lint/typecheck recipe table (see below).
• Self-repair cap: 3 iterations; if still failing after 3, surface tool output verbatim.
• Lint-tool absence policy: when a tool is unavailable, skip that check AND emit <tool> not found — check skipped in the final summary. Never omit this notice.
• Scope-expansion policy: if editing reveals additional files not in the original list need changes, include them in the final summary with a reason. Do NOT prompt mid-edit; do NOT silently expand scope.
• Prohibitions: no diffs shown in the conversation; no mid-edit confirmation prompts.