speq-skill / Docs / Model Routing

---

Model Routing

speq-skill routes work between a main session, workflow skills, and specialist sub-agents. Routing is hardcoded in the generated Claude and Codex plugin artifacts.

---

Principle

Orchestration is cheap. Reasoning is expensive.

Workflow skills coordinate the work — gather context, ask clarifying questions, verify preconditions, dispatch specialist roles. Planning, expert implementation, and review use the heavier reasoning tier because mistakes there compound downstream.

---

Claude routing

Skill or agent	Model	Effort	Notes
/speq:plan	sonnet	inherited	Thin orchestration
/speq:implement	sonnet	inherited	Thin orchestration
/speq:record	sonnet	inherited	Thin orchestration
/speq:mission	inherited	inherited	Interactive bootstrap
/speq:plan-pr	sonnet	inherited	Thin orchestration — headless
/speq:implement-pr	sonnet	inherited	Thin orchestration — headless
/speq:audit	sonnet	inherited	Thin orchestration — health check
Utility skills	inherited	inherited	Reference material for the caller
planner-agent	opus	xhigh	Spec deltas, ADRs, task decomposition
plan-reviewer	opus	xhigh	Adversarial plan review
implementer-agent	sonnet	high	Standard implementation tasks
implementer-expert-agent	opus	xhigh	Tasks tagged [expert]
code-reviewer	opus	xhigh	Adversarial implementation review
audit-agent	opus	high	Mission ↔ spec-library sync
recorder-agent	sonnet	medium	Deterministic spec merge and archive
git-agent	sonnet	medium	Generic git/GitHub operations — no reasoning premium

---

Codex routing

Skill or agent	Model	Effort	Notes
/speq:plan	gpt-5.4	medium	Thin orchestration
/speq:implement	gpt-5.4	medium	Thin orchestration
/speq:record	gpt-5.4	medium	Thin orchestration
/speq:mission	inherited	inherited	Interactive bootstrap
/speq:plan-pr	gpt-5.4	medium	Thin orchestration — headless
/speq:implement-pr	gpt-5.4	medium	Thin orchestration — headless
/speq:audit	gpt-5.4	medium	Thin orchestration — health check
Utility skills	inherited	inherited	Reference material for the caller
planner-agent	gpt-5.5	xhigh	Spec deltas, ADRs, task decomposition
plan-reviewer	gpt-5.5	xhigh	Adversarial plan review
implementer-agent	gpt-5.4	high	Standard implementation tasks
implementer-expert-agent	gpt-5.5	xhigh	Tasks tagged [expert]
code-reviewer	gpt-5.5	xhigh	Adversarial implementation review
audit-agent	gpt-5.5	high	Mission ↔ spec-library sync
recorder-agent	gpt-5.4	medium	Deterministic spec merge and archive
git-agent	gpt-5.4	medium	Generic git/GitHub operations — no reasoning premium

---

Expert task tagging

planner-agent marks tasks requiring deep reasoning with [expert]:

- [ ] 2.1 Add CLI flag parsing
- [ ] 2.2 Implement lock-free queue for concurrent spec writes [expert]

When /speq:implement processes a task group:

• Untagged tasks route to implementer-agent
• [expert] tasks route to implementer-expert-agent
• Expert tasks run first when they establish invariants that standard tasks depend on

Use [expert] for concurrency, subtle correctness, cross-file refactors, novel algorithms, and security-sensitive work. Avoid it for routine CLI plumbing, fixtures, docs, or straightforward use of existing patterns.

---

Packaging

The checked-in source currently lives in .claude/skills and .claude/agents. scripts/plugin/build.sh treats that as the shared source and generates platform-specific outputs:

• Claude artifacts under dist/marketplace/plugins/speq-skill
• Codex artifacts under dist/marketplace/codex/plugins/speq-skill
• Codex marketplace manifest under dist/marketplace/codex/.agents/plugins/marketplace.json

Both generated outputs expose /speq:* skills. Claude-specific workflow syntax is translated out of the Codex output during generation.