Plan and Execute

Mission

Create bite-sized implementation plans and proactively dispatch execution to the cheapest capable backend.

Hard Rules

1. Route first. Do not self-implement Codex-eligible tasks before trying Codex.
2. Route repository exploration for code tasks to Codex (read-only) before using Claude-native Explore/Task subagents.
3. Route analysis/reasoning research to Codex (gpt-5.4). Use Claude WebSearch/WebFetch for live web research.
4. Keep architecture and product decisions in Claude.
5. Codex is fail-closed by default. On Codex failure, ask the user for explicit approval before Claude/Sonnet fallback.
6. Verify outputs before handoff.

1) Backend Readiness Check

Before execution, check availability:

command -v codex >/dev/null && echo "codex:ok" || echo "codex:missing"

If Codex is missing for code tasks, stop and ask the user whether to proceed on Claude/Sonnet fallback.

2) Planning Requirements

• Save plan as docs/plans/YYYY-MM-DD-feature.md.
• Keep tasks 2-5 min each.
• Every task must include:
  • objective
  • concrete file paths
  • command(s)
  • routed backend
  • done criteria
• Add a routing summary table to the plan.

3) Routing Policy

Work Type	Backend	Model	Invocation
Repository exploration / static codebase analysis	Codex CLI	gpt-5.3-codex	CODEX_MODEL=gpt-5.3-codex ${CLAUDE_PLUGIN_ROOT}/skills/plan-and-execute/codex-runner.sh "[prompt]" read-only /path/to/project
Implementation/refactor/tests	Codex CLI	gpt-5.3-codex	CODEX_MODEL=gpt-5.3-codex ${CLAUDE_PLUGIN_ROOT}/skills/plan-and-execute/codex-runner.sh "[prompt]" workspace-write /path/to/project
Code review/spec review	Codex CLI	gpt-5.3-codex	CODEX_MODEL=gpt-5.3-codex ${CLAUDE_PLUGIN_ROOT}/skills/plan-and-execute/codex-runner.sh "[prompt]" read-only /path/to/project
Logic review / analysis / non-code reasoning	Codex CLI	gpt-5.4	${CLAUDE_PLUGIN_ROOT}/skills/plan-and-execute/codex-runner.sh "[prompt]" read-only /path/to/project
Live web / docs / external research	Claude native	—	Use WebSearch / WebFetch tools; stay on Claude
Independent option gathering (compare model outputs)	Codex × 2 in parallel	both	Run two codex-runner.sh invocations concurrently — one with CODEX_MODEL=gpt-5.3-codex, one with default gpt-5.4
Orchestration/synthesis	Claude	—	Stay on Claude
Fallback	Sonnet 4.6	—	Task(prompt, model:sonnet) only after explicit user confirmation when Codex fails

4) Prompt Contract for Routed Tasks

For every routed task prompt, include:

1. Goal and acceptance criteria
2. File paths and constraints
3. Required commands/tests
4. Expected output format (patch, summary, citations, etc.)

5) Execution Loop

1. Load the plan.
2. Execute in batches of up to 3 independent tasks.
3. For independent option-gathering tasks, run two Codex invocations concurrently (one per model).
4. For all other tasks, dispatch to mapped backend with the correct CODEX_MODEL first.
5. Handle exit codes per fallback policy.
6. Review and verify outputs between batches.
7. Continue until all tasks are complete.

6) Codex CLI Usage

Item	Value
Model (code tasks)	gpt-5.3-codex — set via CODEX_MODEL=gpt-5.3-codex
Model (logic/analysis/research tasks)	gpt-5.4 (default)
Sandbox modes	workspace-write, read-only
Flags	--full-auto --skip-git-repo-check

Accepted invocation formats:

• Standard: codex-runner.sh "<prompt>" [sandbox-mode] [working-dir]
• Legacy-compatible: codex-runner.sh "<prompt>" <working-dir> [sandbox-mode]
• Legacy default: if only <working-dir> is provided, sandbox defaults to workspace-write

"${CLAUDE_PLUGIN_ROOT}/skills/plan-and-execute/codex-runner.sh" \
  "$PROMPT" \
  workspace-write \
  /path/to/project

"${CLAUDE_PLUGIN_ROOT}/skills/plan-and-execute/codex-runner.sh" \
  "$PROMPT" \
  read-only \
  /path/to/project

"${CLAUDE_PLUGIN_ROOT}/skills/plan-and-execute/codex-runner.sh" \
  "$PROMPT" \
  /path/to/project

"${CLAUDE_PLUGIN_ROOT}/skills/plan-and-execute/codex-runner.sh" \
  "$PROMPT" \
  /path/to/project \
  read-only

7) Parallel Dual-Model Option Mode

Use this when you want diverse analysis from both Codex models for the same objective. Run two codex-runner.sh invocations concurrently in the background.

# Run gpt-5.3-codex (code perspective) in background
CODEX_MODEL=gpt-5.3-codex /bin/bash "${CLAUDE_PLUGIN_ROOT}/skills/plan-and-execute/codex-runner.sh" \
  "$PROMPT" read-only /path/to/project &
CODEX_PID_1=$!

# Run gpt-5.4 (logic/analysis perspective) in background
/bin/bash "${CLAUDE_PLUGIN_ROOT}/skills/plan-and-execute/codex-runner.sh" \
  "$PROMPT" read-only /path/to/project &
CODEX_PID_2=$!

wait $CODEX_PID_1; wait $CODEX_PID_2

9) Fallback and Error Handling

Exit Code	Action
0	Success
20	Fail-closed signal. Ask user for explicit approval before Claude/Sonnet fallback.
10, 11, 12, 13, 1	Only if CODEX_FAIL_CLOSED=0: allow fallback/report per policy.

10) Environment Defaults

Variable	Default	Purpose
CODEX_MODEL	gpt-5.4	Override per-invocation; use gpt-5.3-codex for code tasks
CODEX_MODEL_CODE	gpt-5.3-codex	Reference value for code construction/review tasks
CODEX_MODEL_LOGIC	gpt-5.4	Reference value for logic review/analysis/research tasks
CODEX_EFFORT	xhigh	Reasoning effort
CODEX_TIMEOUT	600	Timeout in seconds
CODEX_FAIL_CLOSED	1	Stop and ask user before Claude fallback on failure

Telemetry note:

• Router metrics track token offload plus backend attempt/success/failure health.
• Statusline format: Offload C:<tokens> Σ:<tokens> | S/F C:<s>/<f> | RL C:<remaining>%@<reset-time>.

11) Anti-Patterns

• Creating a plan without a backend route per task
• Using Claude-native Explore/Task subagents for Codex-eligible work before attempting routed runners
• Implementing code directly in Claude when Codex is available
• Using Claude for logic/analysis tasks when Codex (gpt-5.4) is available
• Skipping parallel dual-model invocation for independent option-gathering tasks
• Skipping verification between batches
• Falling back without first attempting the mapped backend
• Proceeding on Claude/Sonnet after Codex failure without explicit user approval