用 Codex 或 Claude 帮你安装 复制这段 Prompt,粘贴到 Codex、Claude 或其他助手里,让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/notque/vexjoy-agent --skill do命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| name | do |
| description | Classify user requests and route to the correct agent + skill. Primary entry point for all delegated work. |
| user-invocable | true |
| argument-hint | <request> |
| allowed-tools | ["Read","Bash","Grep","Glob","Skill","Task"] |
| routing | {"triggers":["route task","classify request","which agent","delegate to skill","smart router"],"category":"meta-tooling"} |
ROUTER, not worker. Classify → agent+skill → dispatch. All execution goes to agents. Catching yourself reading/writing code or analyzing — pause and route to an agent. Main: Classify→Select→Dispatch→Evaluate→Re-route→Report.
Do the whole thing (tests+docs). Product, not plan. Permanent solve over workaround. Search before building; test before shipping. Decompose into agent-sized tasks. The result reads as "that's done," not "that's a start." Partial → follow-up. Inject Simple+. Confidence in handling directly is a signal to route.
Dense-Complete Writing (build-dispatch.py injects; skills/shared-patterns/dense-complete-writing.md). User: banners+summary. Internal: JSON/reasoning/stacking (Verbose overrides).
Google Developer Documentation Style (build-dispatch.py injects; skills/shared-patterns/google-devdocs-style.md), alongside Dense-Complete. Precedence: completeness floor (never drop a required point) > Google construction (active voice, second person, context-before-instruction, formatting) > Dense-Complete length.
Every phase: /do > Phase N: PHASE_NAME — description...
After Phase 2: === routing banner. Both required.
Read CLAUDE.md first.
| Complexity | Agent | Skill | Direct |
|---|---|---|---|
| Trivial | No | No | ONLY user-named file by path |
| Simple | Yes | Yes | Route |
| Medium | Required | Required | Route |
| Complex | 2+ | 2+ | Route |
Beyond user-named file = Simple+, MUST route. Uncertain → UP. Depth: references/progressive-depth.md. NOT Trivial: repos/URLs, opinions, git, codebase Qs, retro, comparisons.
Parallel FIRST: 2+ failures / 3+ subtasks → multiple Agent tools. Research→research-coordinator-engineer; coord→project-coordinator-engineer; plan+exec→subagent-driven-development; feature→feature-lifecycle (.feature/→feature-state.py status). Force Direct: OFF.
Creation Detection (MANDATORY): create/scaffold/build/"add new"/"new [component]" targeting agent/skill/pipeline/hook/feature/plugin/workflow/voice. ANY + Simple+ → is_creation=true, Phase 4 Step 0. Not: debug/review/fix/refactor/explain/audit.
Gate: Complexity set. Creation → [CREATION REQUEST DETECTED]. Trivial: direct. Simple+: Phase 2.
Goal: fill every slot the request earns — agent(s), skill, pipeline. The semantic self-route is PRIMARY and runs FIRST — the orchestrator reads the manifest in-session and decides for itself, with no routing sub-dispatch (self-route beat the Haiku hop by +8.1 accuracy points, zero new safety misses: scripts/routing-ab-results/self-route-v1/VERDICT.md). pre-route.py is a guardrail that runs AFTER the semantic decision and never short-circuits it.
Contract: read for INTENT. Route on what the user MEANS. Trigger keywords are hints, never gates. Plain or non-native phrasing routes as well as jargon: "send my commits to the server" routes like "git push". Cost: one manifest read per request — measured and accepted.
Step 0: Semantic self-route (PRIMARY — runs first)
The routing manifest (scripts/routing-manifest.py) is the runtime form; docs/routing-map.md is the human-readable committed form of the same data. Both are generated from frontmatter, so frontmatter is the single source of truth. CI checks staleness via scripts/generate-routing-map.py --check.
Resolve SDIR (this probe also identifies the harness for Phase 4 Model Selection), then read the manifest (hash-gated cache or regenerate):
SDIR="${HOME}/.claude/scripts"; [ -d "$SDIR" ] || SDIR="${HOME}/.hermes/scripts"; [ -d "$SDIR" ] || SDIR="${HOME}/.factory/scripts"; [ -d "$SDIR" ] || SDIR="${HOME}/.codex/scripts"; [ -d "$SDIR" ] || SDIR="${HOME}/.reasonix/scripts"
bash "$SDIR/get-routing-manifest.sh"
Use bash explicitly so routing does not depend on the script's executable bit.
Hold the decision internally as JSON. It stays unprinted; the [do-route] marker is its sole external trace:
{
"agent": "primary agent-name or null",
"agents": ["extra agent names for parallel fan-out; [] when one agent covers it"],
"skill": "skill-name or null",
"pipeline": "pipeline-name or null",
"reasoning": "one sentence why",
"confidence": "high/medium/low"
}
Routing rules (ALL apply):
SECTION-INTEGRITY RULE (HARD CONSTRAINT — never violate):
- `agent` MUST be a name listed in the manifest's AGENTS: section, or null. Never put a skill name in `agent`.
- `skill` MUST be a name listed in the SKILLS: section, or null. Never put an agent name in `skill`.
- `pipeline` MUST be a name listed in the PIPELINES: section, or null.
- If no agent fits, return `"agent": null` — DO NOT promote a skill into the `agent` slot. The router falls back to a default agent (e.g. `general-purpose`) and pairs it with your chosen skill.
- Skills marked FORCE are still skills, not agents. They fill the `skill` slot only. Example: `shell-config` is a SKILL — on a match set `"skill": "shell-config"` and pick a separate agent (or null) for `agent`.
- Pipelines marked FORCE are still pipelines. They fill the `pipeline` slot only, and the run still needs its own `agent` and `skill`.
- Every name in `agents` MUST also be an AGENTS: name, and distinct from `agent`.
FORCE-ROUTE RULE: manifest entries marked FORCE — in SKILLS: or in PIPELINES: — MUST be selected when their domain clearly matches the user's intent. FORCE matching is SEMANTIC, not keyword-based — match what the user MEANS, not individual words:
- "push my changes" → pr-workflow ✓ (git push) | "push back on this design" → NOT pr-workflow (means resist)
- "configure my fish shell" → shell-config ✓ (the Fish shell) | "fish for bugs" → NOT shell-config (means search)
- "quick fix to the login page" → quick ✓ (small edit) | "quick overview of the architecture" → NOT quick (means explore)
A FORCE pipeline (5 of the 29) binds the `pipeline` slot exactly as a FORCE skill binds `skill`. `pre-route.py` reads FORCE pipelines and applies their semantic guard policy; the semantic route still owns intent and must apply the same MEANS-not-words test above.
PIPELINE-SELECTION RULE: pick a pipeline whenever the work has REAL PHASES. The PIPELINES: section ships in every manifest and 29 pipelines are available; reach for one on ANY of:
(1) the intent semantically matches a pipeline's description or its `t:` triggers, OR
(2) the shape is multi-phase — 3+ distinct steps, gather-then-synthesize, mixed script+LLM work, or intermediate artifacts worth keeping, OR
(3) Phase 1 classified the request Complex.
The user saying the word "pipeline" is one signal among these, never the gate. Examples:
- "write an article in vexjoy voice about X" → voice-writer ✓ | "research X with artifacts and sources" → research-pipeline ✓
- "comprehensive review of these 8 files" → comprehensive-review ✓ (outranked by `right-size-review.py` when a real diff exists)
- "add caching to the API and update the docs" → feature-pipeline ✓ (design → implement → document; nobody said "pipeline")
- "help me understand how auth works across this repo" → explore-pipeline ✓ (parallel exploration; a plain pipeline earns the pick on shape, no FORCE flag needed)
Return null when the whole job is one step for one agent: "fix the typo on line 42 of foo.py", "debug this failing test", "review this 10-line function".
MULTI-AGENT RULE: `agents` holds EXTRA agents beyond `agent`; `[]` when one agent covers the work. Fan out when the parts run at once against separate files: 2+ independent failures, 3+ independent subtasks, per-package or per-language review, gather from several domains. Keep a single agent when the parts touch the same files or each step consumes the previous step's output. Complex (Phase 1) starts from fan-out and justifies staying single.
SPECIFICITY RULES:
- Pick the most specific match. "Go tests" → golang-general-engineer + go-patterns, not general-purpose.
- Agent handles the domain. Skill handles the methodology. Pick both when possible.
- Prefer entries whose description semantically matches the request, not just keyword overlap.
- A task verb in the request (review, debug, refactor, test) prefers the skill matching that verb.
- GENUINE git / version-control operations — actually pushing code, committing files, opening or merging a pull request — ALWAYS select pr-workflow. Metaphorical uses ("commit to a decision", "merge ideas in your head", "push back on a proposal") NEVER route to pr-workflow.
- Return a single skill name as a string, not an array. Multiple candidates → pick the primary one.
COMBINATION DOCTRINE. Four surfaces compose; they do not compete. Fill every slot the request earns.
| Surface | Answers | Slot |
|---|---|---|
| Agent | who owns the domain | agent, plus agents for fan-out |
| Skill | which methodology runs | skill |
| Pipeline | which phase structure holds the work | pipeline |
| Stack | which extra rigor rides along | Phase 3 stack |
One surface filled is the FLOOR, not the ceiling — and 29.2% of dispatches sit at or below it (evidence_route_decisions 2026-08-15; a fallback agent or fallback skill counts as unfilled). Combine upward: Simple = agent+skill. Medium = agent+skill+stack. Complex = 2+ agents (agent plus agents), 2+ skills (skill plus Phase 3 stack), and a pipeline unless one phase truly covers the work — which is how Phase 1's "2+/2+" row is satisfied with one skill string.
Composition rules:
quick with any pipeline (quick means one step — drop the pipeline); comprehensive-review with right-size-review.py (a real diff wins); objective-loop as fallback beside a real domain skill (the fallback yields); two pipelines (pick the outer one, nest the other through Step 1c).Step 0b: Apply the routing decision
Use the agent and skill fields directly. Low confidence → verify against the INDEX files.
Skill-greediness gate (HARD — non-negotiable for Simple+). Null skill → pick: review→systematic-code-review, debug→workflow (systematic-debugging), refactor→workflow (systematic-refactoring), audit→systematic-code-review (whole-repo→full-repo-review), explain→codebase-overview, compare→decision-helper (agent A/Bs→agent-comparison), plan→planning, loop→objective-loop. Fallback: objective-loop.
Agent-greediness gate (HARD — non-negotiable for Simple+). general-purpose is the last resort, not the default. Measured share of dispatches: 42.5% (128/301, evidence_route_decisions 2026-08-15). Target band: unmeasured -- see the learning-db.py route health report for the current band and its provenance. A null agent works this table before general-purpose is permitted:
| Domain signal in the request | Agent |
|---|---|
| Go | golang-general-engineer |
| Python | python-general-engineer |
| TypeScript UI, React, bundling, state | typescript-frontend-engineer |
| TypeScript runtime bug, async race, type error | typescript-debugging-engineer |
| Node backend, REST, auth, webhooks | nodejs-api-engineer |
| Swift, Kotlin, PHP | swift-general-engineer, kotlin-general-engineer, php-general-engineer |
| SQL schema, query plans, migrations | database-engineer (SQLite + Peewee → sqlite-peewee-engineer) |
| ETL, warehouse, stream processing | data-engineer |
| Kubernetes, Helm, Ansible | kubernetes-helm-engineer, ansible-automation-engineer |
| Metrics and dashboards, search clusters, message queues | prometheus-grafana-engineer, opensearch-elasticsearch-engineer, rabbitmq-messaging-engineer |
| This toolkit: Python hooks | hook-development-engineer |
| This toolkit: skills, agents, routing tables, ADRs, INDEX files | toolkit-governance-engineer |
| Harness or toolkit upgrade sweep | system-upgrade-engineer |
| Tests, coverage, E2E | testing-automation-engineer |
| Web performance; design system and accessibility | performance-optimization-engineer, ui-design-engineer |
| React Native, Expo | react-native-engineer |
| API docs and runbooks; explainers and articles | technical-documentation-engineer, technical-journalist-writer |
| Review: quality / system + security / ADR + business logic / perspectives | reviewer-code, reviewer-system, reviewer-domain, reviewer-perspectives |
| Broad investigation; agent coordination; pipeline scaffolding | research-coordinator-engineer, project-coordinator-engineer, pipeline-orchestrator-engineer |
| MCP servers | mcp-local-docs-engineer |
No row fits → read the manifest's AGENTS: section again and match on description before falling back.
Pairing rule (the measured defect). 72% of those 128 (92) carried a named domain skill, fallbacks excluded: the router read the domain and had no slot to say so. A specific domain skill therefore obliges a matching domain agent — or a stated reason no agent covers it.
Fallback reason (MANDATORY). Every general-purpose pick carries a written reason, one line, in both places the dispatch records it: the Step 3 banner's -> Agent: why field, and task_spec.constraints handed to build-dispatch.py. Shape: general-purpose: <why no listed agent covers this domain>. A pick with no reason is a routing bug — return to the table.
Section validator (MANDATORY before dispatch):
agents = tokens(manifest, "AGENTS:", "SKILLS:")
skills = tokens(manifest, "SKILLS:", "PIPELINES:")
if route.agent not in agents:
if route.agent in skills: route.skill ||= route.agent
route.agent = None; record_misroute(...)
route.agent ||= "general-purpose"
No pair→general-purpose+objective-loop. [cross-repo]→.claude/agents/. Code→domain agents.
Step 1: Deterministic safety-net (pre-route.py — runs AFTER the semantic decision, never short-circuits it)
Use its result ONLY as a guardrail. Run once per /do; Phase 3 reads its stack:
REQUEST_FILE=$(mktemp); printf '%s' "{user_request}" > "$REQUEST_FILE"
python3 "$SDIR/pre-route.py" --request-file "$REQUEST_FILE" --json-compact
rm -f "$REQUEST_FILE"
→ PRE_ROUTE_RESULT.
"confidence": "high" with a force_route match for pr-workflow or a security skill overrides a disagreeing semantic pick: genuine push, commit, create-PR, and merge work, and security work, MUST hit the quality gates (lint, tests, CI). Record match_type. The agent stays the Step 0 pick, or the Agent-greediness table result when Step 0 returned null.pre-route.py already suppress idiom false positives ("fish out", metaphorical commit/merge), so a guarded or non-matching result leaves the semantic pick standing. Matching only force-routes is by design — the semantic route owns the long tail.Step 2: Apply skill override — "review"→systematic-code-review, "debug"→workflow (systematic-debugging pipeline), "refactor"→workflow (systematic-refactoring pipeline), "TDD"→test-driven-development. Full table in INDEX.
Step 3: Routing banner (MANDATORY — first visible output)
===================================================================
ROUTING: [brief summary]
===================================================================
Selected:
-> Agent: [name] - [why]
-> Skill: [name] - [why]
-> Pipeline: PHASE1 → PHASE2 → ... (if pipeline; phases from skills/workflow/references/pipeline-index.json)
-> Extra Rigor: [verification patterns for code/security/testing when needed]
Invoking...
===================================================================
Trivial: Classification: Trivial - [reason], Handling directly.
Gate: Agent+skill set, banner shown. Phase 3.
Stack on signals.
| Signal | Enhancement |
|---|---|
| Substantive | Retro knowledge when material |
| "with tests"/"production ready" | test-driven-development+verification-before-completion |
| "research needed"/"investigate first" | research-coordinator-engineer |
| Comprehensive/thorough/full review or 5+ files, no diff | parallel-code-review (Security, BizLogic, Arch) |
| Multi-file review, real diff | right-size-review.py; T1→3,T2→12,T3→17,T4→27. CRITICAL+1. Outranks comprehensive-review. |
| Complex implementation | Offer subagent-driven-development |
| "local only"/"no push"/"keep it local"/"stay local" | Inject shared-patterns/local-only.md |
| Voice profile (e.g. voice-example-profile) | Stack voice-writer; voice-*=profile |
| Material unresolved choices | planning — ambiguity-triage.md; ask only when question value is high |
| Needed knowledge or approval exists with another person | planning — human-source-elicitation.md |
| Observation can settle a high impact uncertainty | planning — empirical-prototype.md |
| Phase, worker, or session transition | planning — context-boundary.md |
| Objective with done-criteria / "loop until done" | Stack objective-loop |
| Protected PR/security intent with a Go source operand | Keep pr-workflow/security-review primary and stack go-patterns from PRE_ROUTE_RESULT.stack or router pairs_with |
Review overlap: real-diff row wins; fallback only without diff.
Question-value policy. Complexity alone never forces questions. Inspect the request, repository, supplied material, and prior decisions first. Load planning/references/ambiguity-triage.md, then choose:
| Unresolved state | Action |
|---|---|
| Low impact, reversible, or covered by a safe convention | State the assumption and execute. |
| One high impact decision owned by the current user | Ask one question with a recommendation, then execute. |
| Two or more high impact decisions owned by the current user | Use depth-first-interview.md: batch the independent frontier, then traverse genuine dependencies; interviews initiated by the router cap at five questions and three decision rounds. |
| Facts, constraints, preferences, or approval exist with another person | Use human-source-elicitation.md; draft the artifact and never send without authorization. |
| Evidence can settle the choice faster than discussion | Use empirical-prototype.md: Question → Evidence → Verdict → Next action. |
Explicit "interview me" or "grill me" opts into exhaustive material coverage until shared understanding or a user stop; it has no arbitrary question or round cap. An implicit interview must earn its interruption cost through likely avoided rework and remains bounded. "Just build it," "skip questions," and equivalents use recommended defaults and continue.
An interview is not terminal when it suspends an active delivery objective. Compile the decisions and automatically resume the originating build, fix, install, validation, or other execution flow; never report the decision artifact as completion of the original objective. Stop after compilation only when the user explicitly requested only an interview artifact or excluded implementation.
Ask the current independent frontier in one logical round. In Markdown, include the full frontier with a recommendation for each and wait once. A harness native structured question UI may chunk the frontier only at its capacity per call; do not recompute between chunks unless an answer invalidates a pending question. Single question turns are reserved for true dependency branches. Interviews initiated by the router cap at five questions and three decision rounds, plus at most one concise confirmation response; explicit grills construct and exhaust the material decision tree, with the kinds and number of questions determined by what shared understanding requires. Do not add ceremonial questions. Answering the last frontier does not authorize execution by itself: ask one concise shared understanding confirmation. Skip that extra confirmation only when the same response explicitly says "proceed", "build it", "looks right, continue", or equivalent. Resume nested execution only after explicit proceed or confirmation; a request for only an interview stops at the artifact.
At each pipeline or worker boundary, apply planning/references/context-boundary.md: continue while live context is evidence; use a fresh worker with a complete Task Spec when durable artifacts are sufficient; use pause.md only for plan or session lifecycle; use session-handoff plus the Task Spec for inline worker or agent transfer; compact only under context pressure.
Check pairs_with before stacking. Skills with built-in verification gates may suffice.
anti-rationalization-core always + verification-checklist (code/debug) + anti-rationalization-review + anti-rationalization-security + anti-rationalization-testing; external: untrusted-content-handling. Max: load verification-before-completion references/anti-rationalization-enforcement.md.
Gate: Enhancements applied. Phase 4.
Step 0: Creation — ADR at adr/{name}.md, adr-query.py register, plan.
Step 1: Plan (Simple+) — task_plan.md; skip Trivial.
Step 1b: Quality-loop (Medium+ code mod) — references/quality-loop.md 14 phases. P2 agent=implementation. Force-route in loop. Skip non-code/Trivial/Simple.
Step 1c: Workflow — Pipeline pick or Complex no pick or explicit → ${CLAUDE_SKILL_DIR}/references/workflow-dispatch.md. Both 1b+1c → quality-loop OUTER, workflow in IMPLEMENT.
Step 2: Invoke agent
build-dispatch.py (MANDATORY) — source for [do-route], exact Skill-tool calls, thinking, budget, Task Spec, injections, worktree/local-only. Never hand-assemble.
python3 "$SDIR/build-dispatch.py" --json '{
"agent": "<agent>", "skill": "<skill; omit when agent-only>",
"pipeline": "<pipeline; omit when Phase 2 returned null>",
"complexity": "<trivial|simple|medium|complex>",
"model": "<sonnet|opus|codex|gpt-5.6-sol|gpt-5.6-terra|gpt-5.6-luna>",
"model_policy": "<low-risk|standard|high-risk|max-power>",
"model_effort": "<low|medium|high|xhigh|max>",
"provider": "<anthropic|openai|other>",
"manual_model_override": false,
"health": "-",
"fallback_reason": "<REQUIRED when agent=general-purpose; omit otherwise>",
"stack": ["s1","s2"],
"task_spec": {"intent": "...", "constraints": "...", "acceptance": "...",
"files": "...", "operator_context": "..."},
"flags": {"worktree": false, "local_only": false, "thinking_override": null},
"token_remaining": 480000
}'
agent/skill/complexity: Phase 2 (null→-). pipeline: the Phase 2 pick, passed so the marker carries it; omit when null. The builder validates each name against its index, then emits this exact action contract once per callable skill, primary first with ordered stack de-duplication: Call the Skill tool with \skill-name`.Shared-pattern stack entries remain prompt injections. Agents and pipelines stay out of Skill-tool calls. Fan-out: one call per agent, sameskill/pipeline. model: **required Medium+** (-trivial/simple). Usemodel_policyfor automatic selection — resolves via the harness-native provider lane.model_effortidentifies the benchmark point; advisory for Claude lanes (Agent tool has no per-call effort).provider: harness detection (anthropic|openai|other, default anthropic). A manual model change must set both manual_model_override=trueandmodel_effort; never inherit the policy effort silently. health: -(in-context weights read retired —docs/route-loop-validation.md). fallback_reason: **required when agent=general-purpose** — the one-line reason from the Agent-greediness gate, any prose; build-dispatch.pyslugifies it and appendsfallback=to the marker so every fallback is countable. Dispatch fails without it.stack: Phase 3. task_spec: mandatory Medium+; creation+"match ADR". thinking_override`: slow=security/arch/5+files; fast=lookups.
[do-route] = SOLE signal for routing-decision-recorder. Sub-agents excluded.
Fallback: [do-route] agent={a} skill={s|-} complexity={c}[ pipeline={p}] health=- model={m|-}, Task Spec inline, dispatch.
Model Selection (ADR model-selection-policy).
Harness-native routing. The SDIR probe (Phase 2 pre-route) identifies the harness: ~/.claude → provider anthropic, ~/.codex → provider openai, ~/.hermes/.factory/.reasonix → provider other. Default when absent: anthropic (Claude Code is primary). Each provider lane has its own automatic policy table; cross-provider dispatch is manual-only (explicit tool invocation, never a silent default).
Run deterministic work with scripts, not an LLM. Three decision axes: (1) the current session model — the harness runs Opus 5, and the owner directs Opus 5 as the Anthropic-lane default for every task class. (2) DeepSWE Pass@1 / cost / tokens / steps — agentic task completion rate, the quantitative source for models that have been measured. (3) Owner-observed felt quality — opus > gpt-5.5 (marginal). Benchmark ties or near-ties resolve in favor of felt quality. Cells: Pass@1 / cost / output tokens / steps; cost = avg USD per task, written as a plain number — slash-command templating substitutes dollar-digit positional parameters in this injected body, so a literal dollar sign before a digit corrupts on every argful invocation. Higher Pass@1 better, other three lower-is-better. Opus 5 has no DeepSWE run yet, so its cells read n/a — not yet benchmarked and its pts/USD cannot be computed until it is measured; it is selected on the session-model and owner-directive grounds above, not on a benchmark figure.
Start low, escalate on miss. Task-class tables are ceilings by risk class, not starting points. Default = lowest tier whose risk class matches; escalate one tier only when output misses the acceptance bar. High tiers cost 3-6x per Pass@1 point where measured (see the OpenAI lane's pts/$ column; the Anthropic lane's is pending an Opus 5 benchmark) — pre-paying for xhigh/max "to be safe" wastes the 200 USD/month plan budget. Fan-out rule: parallel readers use the lane's low-risk point; one synthesis agent may run one tier higher. User-facing output (docs, prose, reviews the owner reads, design) leans opus one tier up from the task class; bulk/mechanical/parse-heavy work is where the OpenAI lane's cheaper points earn their keep (under Codex harness or explicit cross-provider call).
Anthropic lane (automatic under Claude Code). Effort is advisory — recorded in marker as model@effort for telemetry; the Agent tool has no per-call effort parameter.
Current default: Opus 5 (opus) at every task class. It is the model this session runs and the owner's directed default, adopted across the lane on 2026-07-24.
| Variant | max | xhigh | high | medium | low |
|---|---|---|---|---|---|
| Opus-5 (current default, unmeasured) | n/a — not yet benchmarked | n/a — not yet benchmarked | n/a — not yet benchmarked | n/a — not yet benchmarked | n/a — not yet benchmarked |
| Opus-4.8 (prior measurement) | 59 / 13.22 / 135k / 120 | 54 / 8.01 / 86k / 95 | 52 / 4.28 / 50k / 73 | 49 / 3.44 / 41k / 66 | 41 / 2.29 / 29k / 54 |
| Sonnet-5 (prior measurement) | 54 / 26.40 / 214k / 268 | 50 / 11.89 / 121k / 186 | 48 / 7.43 / 87k / 147 | 40 / 4.08 / 57k / 108 | 31 / 2.19 / 36k / 77 |
| Task class | Selection | pts/$ | Why |
|---|---|---|---|
| deterministic | no LLM | — | Run the script directly. |
| low-risk | opus / low | n/a | Current session model, owner-directed default; effort floor per start-low. |
| standard | opus / medium | n/a | Current session model, owner-directed default; one tier up for standard work. |
| high-risk | opus / high | n/a | Current session model, owner-directed default; high effort for risk-bearing work. |
| max-power | opus / xhigh | n/a | Current session model, owner-directed default; manual_model_override=true; state justification in task_spec intent. |
Effort selection still follows start low, escalate on miss — the effort column is a ceiling by risk class, and a miss against the acceptance bar is what buys the next tier. Opus 5 at max stays manual-only pending measurement. Sonnet-5 and Opus-4.8 points are the manual-only ones: they need manual_model_override=true plus model_effort, and stay available for cost, latency, context-window, and fan-out breadth constraints the benchmark does not measure. Haiku is retired.
OpenAI lane (automatic under Codex CLI).
| Variant | max | xhigh | high | medium | low |
|---|---|---|---|---|---|
| GPT-5.6 Sol | 73 / 8.39 / 60k / 61 | 71 / 4.70 / 41k / 44 | 69 / 3.47 / 28k / 37 | 61 / 1.86 / 18k / 31 | 45 / 1.07 / 11k / 23 |
| GPT-5.6 Terra | 70 / 4.95 / 72k / 76 | 60 / 2.13 / 40k / 43 | 54 / 1.13 / 22k / 34 | 35 / 0.58 / 12k / 25 | 24 / 0.43 / 8.6k / 21 |
| GPT-5.6 Luna | 67 / 3.03 / 73k / 102 | 57 / 1.54 / 45k / 71 | 44 / 0.78 / 26k / 49 | 11 / 0.22 / 8.2k / 24 | 2 / 0.07 / 3.1k / 12 |
| GPT-5.5 legacy | n/a | 67 / 7.23 / 46k / 82 | 64 / 5.10 / 31k / 62 | 54 / 2.75 / 20k / 46 | 27 / 1.20 / 9.4k / 28 |
| Task class | Selection | pts/$ | Why |
|---|---|---|---|
| deterministic | no LLM | — | Run the script directly. |
| low-risk | gpt-5.6-terra / high | 47.8 | 54 Pass@1 at 1.13, 22k tokens, 34 steps. |
| standard | gpt-5.6-sol / high | 19.9 | 69 Pass@1 at 3.47, 28k tokens, 37 steps. |
| high-risk | gpt-5.6-sol / xhigh | 15.1 | 71 Pass@1 at 4.70, 41k tokens, 44 steps. |
| max-power | gpt-5.6-sol / max | 8.7 | 73 Pass@1 at 8.39, 60k tokens, 61 steps; manual_model_override=true; state justification in task_spec intent. |
All GPT-5.5 choices are manual-only. Off-policy GPT-5.6 points (Sol medium/low, Terra max/xhigh/medium/low, all Luna) are manual-only — some are cost trade-offs, not dominated; use with manual_model_override=true for a stated constraint.
Other harnesses (provider=other): model_policy is unavailable — choose the highest non-dominated Pass@1 point among models the harness exposes, applying the same start-low-escalate-on-miss discipline. Set model explicitly.
Cross-provider escalation — manual only, never automatic. Escalating anthropic → sol is a cost/limits lever or independent-second-opinion lever, not a quality upgrade. Under Claude Code, codex-wrapper dispatches (codex skill, pr-workflow codex second-opinion review) remain valid as EXPLICIT tools — deliberate cross-provider calls, not defaults. Escalation targets: anthropic max-power miss → sol/xhigh or sol/max (second opinion, cheaper per point); openai max-power miss → opus/xhigh (the Anthropic-lane default). Manual-pick ordering among legacy/manual points: opus-4.8 above gpt-5.5 where they otherwise tie.
Coordinator model. The main-thread coordinator routes and evaluates but never executes; its cost is input-dominated (largest context, short outputs), and DeepSWE Pass@1 measures execution it never does. Picks: anthropic harness → opus (Opus 5, the session model — it replaces the prior sonnet pick); openai harness → gpt-5.6-terra/high. Safe because deterministic scripts (pre-route, manifest, build-dispatch, health weights) absorb routing complexity and the learning loop bounds misroute cost. Downgrade the anthropic coordinator to sonnet only as a deliberate plan-limit measure. Session model is set via harness config (/model), not per-turn.
Medium+ MUST set a model or policy. Codex prompts stay read-only and public unless a task requires otherwise.
Complex (3+ sources):
| Verbs | Mode |
|---|---|
| list/count/extract/inventory/search/check/find/grep | Scripts when deterministic; otherwise harness-native low-risk readers → harness-native high-risk synth |
| review/audit/assess/analyze/debug/investigate/evaluate | Single harness-native high-risk agent |
Simple/Medium: direct. Feature-branch; mods commit. isolation:"worktree"→flags.worktree. Non-org: 3 reviews→fix→PR. Org: confirm git.
Step 3: Multi-part / fan-out — deps sequential; independent parallel (max 10). Phase 2 agents → ONE build-dispatch.py call and ONE Agent dispatch per agent: N agents = N calls = N markers, one marker each. Emit the parallel Agent calls in a single message. Packing several markers into one Bash/Workflow script keeps them recorded but forfeits route-fit scoring, which reads a lone marker per event.
Step 4: Auto-Pipeline Fallback (no match, Simple+) — auto-pipeline. None → closest+objective-loop. Never empty skill.
Lazy-completion check. "Done" on enumerable → compare scope; short → reject, re-dispatch (references/lazy-completion-detector.md). Re-dispatch → route failure.
Gate: Agent invoked, results delivered.
Hooks record routing. On an observed route failure, or to re-derive a quoted telemetry figure → load ${CLAUDE_SKILL_DIR}/references/routing-telemetry.md (hooks table, outcome fidelity, route-failure protocol, figure queries).
On any routing error → load ${CLAUDE_SKILL_DIR}/references/error-handling.md.
${CLAUDE_SKILL_DIR}/references/progressive-depth.mdagents/INDEX.json, skills/INDEX.jsonskills/workflow/SKILL.md, skills/workflow/references/pipeline-index.jsonscripts/routing-manifest.py