spectre: Skill for Claude Code

Codex SPECTRE Implementation

Trigger: codex, spectre, codex install, sessionstart, agents.override, registry, spectre-learn, spectre-recall, hooks.json, config.toml, doctor Confidence: high Created: 2026-03-30 Updated: 2026-07-19 Version: 5

Current Design

Codex SPECTRE installs the workflow as Codex-native skills, subagent TOML configs, and generated SessionStart hooks. Project knowledge and session continuity are written into managed AGENTS.override.md blocks so hook output stays short.

Reusable project knowledge is both configured as normal skills and injected as a compact trigger registry:

• spectre-learn writes project skills under .agents/skills/{category}-{slug}/SKILL.md.
• The recall registry lives at .agents/skills/spectre-recall/references/registry.toon.
• spectre-recall is generated as an explicit search/load skill.
• Project installs sync .agents/skills/*/SKILL.md into Codex [[skills.config]].
• spectre-apply contains a {{REGISTRY}} placeholder that both the hook and src/lib/knowledge.js substitute before writing the managed knowledge block.
• bootstrap → handoff-resume → load-knowledge is the required SessionStart order.

Workflow task execution now uses a two-artifact contract:

• spectre-create_tasks writes {OUT_DIR}/specs/execute.md plus {OUT_DIR}/specs/tasks.json.
• execute.md is the compact primary-agent index (document manifest, task detail source, execution summary, wave plan, parent-task index, slicing rules).
• tasks.json is the full mutable detail/status source (meta + phases[]); primary execution/review/validation consumers should slice it by parent task id instead of reading the whole file.
• Do not reintroduce the old specs/tasks.md task-list flow or a Markdown fallback/converter.

Review gates use one cross-runtime contract:

• spectre-plan_review prefers the opposing CLI with an explicit high-effort model: Codex launches Claude Code with --model opus --effort high; Claude Code launches Codex with -m gpt-5.6-sol -c 'model_reasoning_effort="high"'. The launcher allows each attempt up to 20 minutes, but passes no duration guidance to the reviewer.
• spectre-task_review uses a focused medium-effort opposing-runtime contract: Codex launches Claude Code using --model opus --effort medium; Claude Code launches Codex using -m gpt-5.6-sol -c 'model_reasoning_effort="medium"'. The primary agent runs consumer-safety preflight, explicitly launches and monitors the reviewer for up to 20 minutes, validates the report, allows one repair, owns native fallback and write-back, and passes no duration guidance to the reviewer.
• spectre-code_review uses a high-effort opposing-runtime contract: Codex launches Claude Code using --model opus --effort high; Claude Code launches Codex using -m gpt-5.6-sol -c 'model_reasoning_effort="high"'.
• If the opposing runtime is unavailable or fails validation after one repair attempt, the gate dispatches one native reviewer with the same manifest, adversarial lenses, severity/evidence rules, exclusions, and report schema. This fallback does not block completion and must record its reason plus runtime/model metadata.
• spectre-code_review is an adversarial, evidence-gated review for correctness, regressions/integration, security, performance/reliability, overengineering, and test adequacy. It does not use subjective numeric scores.
• spectre-execute delegates its final cumulative review to spectre-code_review --orchestrated; do not reintroduce a separate final-review prompt inside execute.
• Canonical workflow skills live under plugins/spectre/skills/; regenerate plugins/spectre-codex/ and keep regression assertions in scripts/test_sync-codex.cjs aligned with these invariants.