d
Jev request router: validates the requested outcome, then dispatches to the matched agent, skill, and pipeline.
Install
npx skills add https://github.com/notque/vexjoy-agent/tree/main/skills/meta/d
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install notque-vexjoy-agent@llmmart
git clone https://github.com/notque/vexjoy-agent.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole notque/vexjoy-agent collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
/d — Jev Router
Routes a request in four parts. A deterministic pre-route.py guard keeps
git/PR and security work on their quality gates. You, the session model, pick
the agent, skill, and pipeline from the full routing manifest with /do's rule
text. Jev answers the attachment questions that add extra skills, and validates
that the restated outcome survives the route. A script validates your pick and
computes the attachments.
Design rationale: ${CLAUDE_SKILL_DIR}/references/jev-classifier-design.md.
Phase Banners
Every phase: /d > Phase N: PHASE_NAME — description...
After intent alignment resolves: === routing banner. Both required.
Phase 1: CLASSIFY
Resolve $SDIR: ${HOME}/.claude/scripts, falling back through
.gemini/.hermes/.factory/.codex/.reasonix, or the repo's scripts/
directory.
Step 1 — JEV_RESULT. When JEV_RESULT is already in context (the
jev-route-injector hook ran the script before your first token), use it and
skip the command below.
REQUEST_FILE=$(mktemp); printf '%s' "{user_request}" > "$REQUEST_FILE"
python3 "$SDIR/jev-route.py" --request-file "$REQUEST_FILE" --json-compact
Fields: available, jev_called, source, forced, stack, project,
signals, signal_scores, domain_scores, error, latency_ms, usage.
forced is the safety force route (pr-workflow, pr-pipeline, security)
or null. project holds the request repository's languages, frameworks, and
datastores (names only) or null. error is set when Jev was unavailable or
its answers were unusable. Routing continues, and attach then carries only
the pre-route stack and the agent domain floor.
Step 2 — Pick. Set the complexity from /do Phase 1 (Trivial, Simple,
Medium, Complex). Trivial: go to Phase 1T.
For Simple and above, load /do's routing rules and the manifest. Read the rule
text in full and apply every rule:
python3 "$SDIR/print-route-rules.py"
bash "$SDIR/get-routing-manifest.sh" --request-file "$REQUEST_FILE"
The manifest is large. If Bash truncates it or saves it to a file, read the
whole file before you pick. Use JEV_RESULT.project when the request names no
language or framework.
Return the decision as JSON. Do not print it:
{"agent": "<AGENTS: name or null>", "agents": ["<extra fan-out agents>"],
"skill": "<SKILLS: name or null>", "pipeline": "<PIPELINES: name or null>",
"complexity": "simple|medium|complex", "confidence": "high|medium|low",
"reasoning": "<one sentence>",
"fallback_reason": "general-purpose: <why no listed agent covers this>"}
fallback_reason is required when agent is general-purpose, and only then.
When JEV_RESULT.forced is set, its skill and pipeline are final: step 3
overrides your values for them. Pick the agent and agents that fit, as
/do Phase 2 Step 1(a) does.
Step 3 — Finalize. Write JEV_RESULT and the decision to temporary files
and run:
python3 "$SDIR/jev-route.py" --finalize --result-file "$RESULT_FILE" --decision-file "$DECISION_FILE" --json-compact
rm -f "$REQUEST_FILE" "$RESULT_FILE" "$DECISION_FILE"
Hold the output as ROUTE: agent, agent_source, skill, skill_source,
pipeline, agents, attach, stack, complexity, confidence,
reasoning, fallback_reason, source, signals, domain_scores, and
problems. The script checks every name against the live manifest, applies the
safety force route, fills an empty skill with workflow, and gives a null or
general-purpose agent the owning agent of the skill. attach is the list of
extra skills that ride with skill (for example programming for Go work,
testing when tests are part of the work). It is the whole Phase 4 stack.
Gate: problems empty → Phase 2. problems set → fix the decision, run
step 3 once more, and carry any problem that remains into task_spec.gaps. The
output has an error key and no route fields → use your decision as ROUTE
with attach set to [].
Phase 1T: TRIVIAL (complexity == Trivial)
Direct-handling path. It still goes through Phase 2 so the user outcome is
restated and Jev validates it; use {"complexity": "trivial"} as the route
file. After an aligned Phase 2 result, show Classification: Trivial, then
answer or do the one-line action directly. Do not run Phases 3–5 or call
build-dispatch.py. Stop there.
Phase 2: ALIGN INTENT (required for every /d route)
MANDATORY STOP: For every /d invocation, write PROPOSED_INTENT and run
the validator on that exact text before any routing banner, dispatch, answer,
edit, or other action. This requirement has no exception for force routes or
trivial routes.
Before selecting the work method, write PROPOSED_INTENT: a concise one- or
two-sentence restatement of what the user wants accomplished. State the
outcome and each deliverable the request names or directly requires. Copy
every explicit constraint in the user's words: limits ("only", "at most"),
exclusions ("don't touch", "do not deploy"), required methods, and
authorization boundaries. Do not add deliverables the user did not ask for,
such as extra tests, docs, cleanup, refactors, or verification steps. Do not
add notes about missing inputs or preconditions; the validator decides
whether clarification is needed. Do not describe the selected agent, skill,
or implementation mechanics as the outcome. Preserve the user's words where
precision matters.
Put the request, the ROUTE JSON, and the proposed intent in temporary files
rather than shell-splicing user text, then call:
python3 "$SDIR/jev-intent-align.py" \
--request-file "$REQUEST_FILE" \
--route-file "$ROUTE_FILE" \
--proposed-intent-file "$INTENT_FILE" \
--json-compact
The validator sends one bounded state and all independent questions together
through the selected Jev transport. It checks whether the outcome and constraints
are preserved, the route can cover the material scope, the restatement is too
narrow, it introduces unrequested work, and essential clarification is needed.
It returns aligned, clarification_needed, issues, and raw scores.
Show this before the routing banner:
Intent alignment (/d):
-> Restated outcome: [PROPOSED_INTENT]
-> Jev: [aligned|review|unavailable] [issues, if any]
Gate:
clarification_needed == true→ ask one concise question that names the essential ambiguity; do not dispatch until answered.alignment == alignedand complexity is Trivial → direct handling in Phase 1T; otherwise → Phase 3.alignment == reviewbecause scope is lost, work was added, or the route cannot cover the request → correctPROPOSED_INTENTor the route and run this validator once more. Carry unresolved issues intotask_spec.gaps; do not silently proceed as though Jev approved it.alignment == unavailableorerror→ state that validation was unavailable, preserve the verbatim request and proposed intent in the task spec, then continue under theROUTEresult. Gateway outage must not become a false request rejection.
This is an instruction gate enforced by the /d contract, not a hook-enforced
technical boundary. The user remains the final backstop if an agent violates it.
Phase 3: DECIDE (after aligned intent)
Use ROUTE's agent, skill, pipeline, and complexity as-is. The script
validated them against the live manifest. agent_source: skill-default means
the pick named no domain agent and the script used the skill's owning agent. A
general-purpose agent carries ROUTE.fallback_reason. When complexity is
null, use medium, except a single one-line trivial fix → simple.
Routing banner (Phase 2 intent block + routing block, both required, printed together):
===================================================================
ROUTING (/d): [brief summary]
===================================================================
Intent (/d):
-> Restated: [PROPOSED_INTENT]
-> Alignment: [aligned|review|unavailable] [— issues, if any]
Selected:
-> Agent: [ROUTE.agent] - [ROUTE.reasoning]
-> Skill: [ROUTE.skill] - [ROUTE.reasoning]
-> Attached: [ROUTE.attach, comma-separated, or "none"]
-> Pipeline: [ROUTE.pipeline, if set]
-> Source: [ROUTE.source] (confidence: [ROUTE.confidence])
Invoking...
===================================================================
The Intent block must be populated from the Phase 2 validator run. Printing
the banner with a placeholder or omitting the Intent block is a Phase 2 skip
and is not allowed.
Gate: Agent+skill set, banner shown. Phase 4.
Phase 4: ENHANCE (attach skills)
stack = ROUTE.attach, in order, plus anti-rationalization-core.
Copy the names exactly. Do not add, rename, or drop skills: the script built
attach from these rules, and build-dispatch.py rejects any name absent
from skills/INDEX.json.
| Source | Attaches |
|---|---|
JEV_RESULT.stack (pre-route, e.g. a .go file with PR or security work) |
its entries, first |
| Agent domain floor | programming for Go, Kotlin, PHP, and Swift agents; kubernetes for kubernetes-helm-engineer; frontend for ui-design-engineer |
domain_scores at 0.6 or higher |
programming, frontend, kubernetes, testing, building-with-jev, research |
tests_requested / comprehensive_review / objective_loop_worthy |
testing / review / workflow |
local_only |
local-only shared pattern |
At most three skills are attached beyond skill. Two adjustments stay with
you:
comprehensive_reviewattachedreviewand a real multi-file diff exists:right-size-review.pyoutranks it, so dropreviewfromstack.signals.research_neededis true: addresearch-coordinator-engineerto the fan-out agents.
Fan-out agents: union ROUTE.agents (your MULTI-AGENT RULE picks, each
validated against the manifest) into the research_needed agent list,
deduped. Dispatch fan-out agents as separate parallel Agent tool calls
alongside the primary build-dispatch.py dispatch.
Gate: Stack applied. Phase 5.
Phase 5: EXECUTE
Build the task spec with request_verbatim unchanged and intent exactly
PROPOSED_INTENT; include any unresolved alignment issue in gaps, then invoke
build-dispatch.py:
python3 "$SDIR/build-dispatch.py" --json '{
"agent": "<ROUTE.agent>", "skill": "<ROUTE.skill; omit when agent-only>",
"pipeline": "<ROUTE.pipeline; omit when null>",
"complexity": "<ROUTE.complexity>",
"model": "<haiku|opus|fable from Model choice, or gemini-3.8-flash|gemini-4-argon|gemini-4-argon-high>",
"context_mode": "summary",
"provider": "<anthropic|openai|google|antigravity|other>",
"manual_model_override": false,
"health": "-",
"fallback_reason": "<REQUIRED when agent=general-purpose; omit otherwise>",
"stack": ["<ROUTE.attach, in order>", "anti-rationalization-core"],
"task_spec": {"request_verbatim": "<user message, unchanged>", "intent": "...",
"constraints": "<applicable rules, limits, and authorization>",
"decisions": "...",
"gaps": "...",
"acceptance": "<command> -> <expected>",
"files": "<owned paths; optional line ranges>", "ownership": "<worker scope>",
"operator_context": "..."},
"flags": {"worktree": false, "local_only": false, "thinking_override": null},
"token_remaining": 480000
}'
Model choice
Set model from the task type, not inherit. Pass the same model on every Agent call, including Explore and Plan.
| Task type | Model / effort |
|---|---|
| read-only gathering: locate, lookup, cite lines, excerpts, research fetch, log reading | haiku |
| code writing and edits, debugging, features, reviews that decide, decisions, no row fits | opus / medium, then high |
| high-risk: auth, data, security, concurrency, irreversible | opus / medium, then high |
| major refactor or migration, architecture planning, lead of a many-agent fan-out, deep research synthesis, quantitative analysis, long unattended run, or Opus at high fell short | fable / medium, then high |
Sonnet is no row's default. When the owner names a model for the task ("use fable for everything"), pass that model on every dispatch with manual_model_override=true.
Tiers (model_policy): low-risk = haiku/low (read-only gathering only), standard = opus/medium, high-risk = opus/medium, max-power = fable/high. Owner, 2026-11-09: "Also I prefer opus 5.5 medium over high". Opus starts at medium on every row and tier, high-risk included. opus/high is no tier's default: reach it as an explicit pick (model: "opus", model_effort: "high", no model_policy) or as the step after a failed opus/medium attempt. max-power, fable/max, opus/max, and explicit opus/xhigh need manual_model_override=true; test fable at high before opus/xhigh. Use inherit only when no row or tier fits and the session model is acceptable.
Escalate on a miss, when an objective check fails (tests, tsc, ruff, acceptance command) or a Jev blocker Noul is at least 0.6. In order: (1) give the agent a way to verify (test, build, script); (2) raise effort one level, up to high; (3) move to the next model up at a lower effort: haiku → opus/medium → opus/high → fable/medium → fable/high. Never raise opus past high before trying fable.
Evidence: Anthropic's 2026-11-09 model guide and the owner's 2026-11-01 and 2026-11-09 orders, which win where they differ. Prices, effort trade-offs, mixing rules, history, and the re-measure method: skills/meta/d/references/model-task-fit.md.
On Antigravity / Google AI (provider="google"), the tiers are low-risk = gemini-3.8-flash, standard = gemini-4-argon, high-risk = gemini-4-argon-high, and max-power = gemini-4-argon-high / max with manual_model_override=true; see skills/meta/do/references/antigravity-execution.md.
The builder validates each name against its index, then emits the dispatch
action. For Complex or creation requests, apply creation detection, plan-file
gating, quality-loop, workflow dispatch, fan-out, and the auto-pipeline fallback.
Gate: Agent invoked, results delivered.
Post-execute: GRILL-JEV (plan/spec/design output)
After any execution that produces a plan, spec, or design artifact, run
grill-jev automatically before declaring the work complete. This applies
whenever the agent's output contains phases, steps, checklists, or a
structured implementation plan.
Detection: the agent wrote task_plan.md, a spec file, a design document,
or the response itself is a structured plan with numbered steps or phases.
# File artifact
python3 "$SDIR/grill-jev.py" --file task_plan.md --mode plan
# Inline plan (write to temp file first, then grill)
python3 "$SDIR/grill-jev.py" --file /tmp/plan_output.md --mode plan
Print the findings report. If high-signal findings exist (exit code 1):
- Show findings to the user
- Ask whether to address findings before proceeding or accept and move on
If no high-signal findings (exit code 0): proceed, note "grill-jev: clean".
Skip grill-jev when:
- The output is code only (no plan structure) — use
--mode codeinstead - The output is a pure research response with no actionable steps
- grill-jev is itself the requested action (avoid recursion)
Error handling
Errors inside jev-route.py resolve to an error key and an empty evidence
set. Routing continues; Phase 2 reports an unavailable validation.
When changing how the router or validator builds or sends Jev requests, apply
skills/shared-patterns/jev-production-lessons.md: stage requests at or under the reliable size, send each
stage's requests together, and retry by status code.
References
${CLAUDE_SKILL_DIR}/references/jev-classifier-design.md— request/response contract, what Jev still decides, design decisions${CLAUDE_SKILL_DIR}/references/model-task-fit.md— model per task type, measured cost and Jev scores, re-measure method${CLAUDE_SKILL_DIR}/SPEC.md,${CLAUDE_SKILL_DIR}/EVAL.md— maintenance contract and regression cases (load only when creating, evaluating, or redesigning this skill)scripts/jev-route.py,scripts/print-route-rules.py,scripts/jev-intent-align.py,scripts/jev_transport.py,scripts/jev_vercel.py,scripts/jev_gateway/jev_vercel_gateway.mjs,scripts/pre-route.py,scripts/routing-manifest.py,scripts/build-dispatch.py- Rule text:
skills/meta/do/SKILL.mdPhase 2 (printed byprint-route-rules.py) - Jev hook:
hooks/jev-route-injector-userprompt.py(UserPromptSubmit) precomputesJEV_RESULT
Files (vexjoy-agent)
-
references
-
jev-classifier-design.md 9 KB
Loading… -
model-task-fit.md 13.6 KB
Loading…
-
-
EVAL.md 6.2 KB
Loading… -
SKILL.md 16.9 KB
--- name: d version: "2.0.0" description: "Request router: picks the agent, skill, and pipeline with /do's rules, then uses Jev to validate the requested outcome and attach extra skills." user-invocable: true argument-hint: "[request]" allowed-tools: - Read - Bash - Grep - Glob - Skill - Agent routing: triggers: - "jev router" - "route with jev" - "use the d router" not_for: "General-purpose task execution — /d classifies and dispatches, it does not perform the work itself." category: meta-tooling --- # /d — Jev Router Routes a request in four parts. A deterministic `pre-route.py` guard keeps git/PR and security work on their quality gates. You, the session model, pick the agent, skill, and pipeline from the full routing manifest with `/do`'s rule text. Jev answers the attachment questions that add extra skills, and validates that the restated outcome survives the route. A script validates your pick and computes the attachments. Design rationale: `${CLAUDE_SKILL_DIR}/references/jev-classifier-design.md`. ### Phase Banners Every phase: `/d > Phase N: PHASE_NAME — description...` After intent alignment resolves: `===` routing banner. Both required. --- ### Phase 1: CLASSIFY Resolve `$SDIR`: `${HOME}/.claude/scripts`, falling back through `.gemini`/`.hermes`/`.factory`/`.codex`/`.reasonix`, or the repo's `scripts/` directory. **Step 1 — JEV_RESULT.** When `JEV_RESULT` is already in context (the `jev-route-injector` hook ran the script before your first token), use it and skip the command below. ```bash REQUEST_FILE=$(mktemp); printf '%s' "{user_request}" > "$REQUEST_FILE" python3 "$SDIR/jev-route.py" --request-file "$REQUEST_FILE" --json-compact ``` Fields: `available`, `jev_called`, `source`, `forced`, `stack`, `project`, `signals`, `signal_scores`, `domain_scores`, `error`, `latency_ms`, `usage`. `forced` is the safety force route (`pr-workflow`, `pr-pipeline`, `security`) or `null`. `project` holds the request repository's languages, frameworks, and datastores (names only) or `null`. `error` is set when Jev was unavailable or its answers were unusable. Routing continues, and `attach` then carries only the pre-route stack and the agent domain floor. **Step 2 — Pick.** Set the complexity from `/do` Phase 1 (Trivial, Simple, Medium, Complex). Trivial: go to Phase 1T. For Simple and above, load `/do`'s routing rules and the manifest. Read the rule text in full and apply every rule: ```bash python3 "$SDIR/print-route-rules.py" bash "$SDIR/get-routing-manifest.sh" --request-file "$REQUEST_FILE" ``` The manifest is large. If Bash truncates it or saves it to a file, read the whole file before you pick. Use `JEV_RESULT.project` when the request names no language or framework. Return the decision as JSON. Do not print it: ``` {"agent": "<AGENTS: name or null>", "agents": ["<extra fan-out agents>"], "skill": "<SKILLS: name or null>", "pipeline": "<PIPELINES: name or null>", "complexity": "simple|medium|complex", "confidence": "high|medium|low", "reasoning": "<one sentence>", "fallback_reason": "general-purpose: <why no listed agent covers this>"} ``` `fallback_reason` is required when `agent` is `general-purpose`, and only then. When `JEV_RESULT.forced` is set, its `skill` and `pipeline` are final: step 3 overrides your values for them. Pick the `agent` and `agents` that fit, as `/do` Phase 2 Step 1(a) does. **Step 3 — Finalize.** Write `JEV_RESULT` and the decision to temporary files and run: ```bash python3 "$SDIR/jev-route.py" --finalize --result-file "$RESULT_FILE" --decision-file "$DECISION_FILE" --json-compact rm -f "$REQUEST_FILE" "$RESULT_FILE" "$DECISION_FILE" ``` Hold the output as `ROUTE`: `agent`, `agent_source`, `skill`, `skill_source`, `pipeline`, `agents`, `attach`, `stack`, `complexity`, `confidence`, `reasoning`, `fallback_reason`, `source`, `signals`, `domain_scores`, and `problems`. The script checks every name against the live manifest, applies the safety force route, fills an empty skill with `workflow`, and gives a null or `general-purpose` agent the owning agent of the skill. `attach` is the list of extra skills that ride with `skill` (for example `programming` for Go work, `testing` when tests are part of the work). It is the whole Phase 4 stack. **Gate**: `problems` empty → Phase 2. `problems` set → fix the decision, run step 3 once more, and carry any problem that remains into `task_spec.gaps`. The output has an `error` key and no route fields → use your decision as `ROUTE` with `attach` set to `[]`. --- ### Phase 1T: TRIVIAL (complexity == Trivial) Direct-handling path. It still goes through Phase 2 so the user outcome is restated and Jev validates it; use `{"complexity": "trivial"}` as the route file. After an aligned Phase 2 result, show `Classification: Trivial`, then answer or do the one-line action directly. Do not run Phases 3–5 or call `build-dispatch.py`. Stop there. --- ### Phase 2: ALIGN INTENT (required for every /d route) **MANDATORY STOP:** For every `/d` invocation, write `PROPOSED_INTENT` and run the validator on that exact text before any routing banner, dispatch, answer, edit, or other action. This requirement has no exception for force routes or trivial routes. Before selecting the work method, write `PROPOSED_INTENT`: a concise one- or two-sentence restatement of what the user wants accomplished. State the outcome and each deliverable the request names or directly requires. Copy every explicit constraint in the user's words: limits ("only", "at most"), exclusions ("don't touch", "do not deploy"), required methods, and authorization boundaries. Do not add deliverables the user did not ask for, such as extra tests, docs, cleanup, refactors, or verification steps. Do not add notes about missing inputs or preconditions; the validator decides whether clarification is needed. Do not describe the selected agent, skill, or implementation mechanics as the outcome. Preserve the user's words where precision matters. Put the request, the `ROUTE` JSON, and the proposed intent in temporary files rather than shell-splicing user text, then call: ```bash python3 "$SDIR/jev-intent-align.py" \ --request-file "$REQUEST_FILE" \ --route-file "$ROUTE_FILE" \ --proposed-intent-file "$INTENT_FILE" \ --json-compact ``` The validator sends one bounded state and all independent questions together through the selected Jev transport. It checks whether the outcome and constraints are preserved, the route can cover the material scope, the restatement is too narrow, it introduces unrequested work, and essential clarification is needed. It returns `aligned`, `clarification_needed`, `issues`, and raw `scores`. Show this before the routing banner: ``` Intent alignment (/d): -> Restated outcome: [PROPOSED_INTENT] -> Jev: [aligned|review|unavailable] [issues, if any] ``` **Gate:** - `clarification_needed == true` → ask one concise question that names the essential ambiguity; do not dispatch until answered. - `alignment == aligned` and complexity is Trivial → direct handling in Phase 1T; otherwise → Phase 3. - `alignment == review` because scope is lost, work was added, or the route cannot cover the request → correct `PROPOSED_INTENT` or the route and run this validator once more. Carry unresolved issues into `task_spec.gaps`; do not silently proceed as though Jev approved it. - `alignment == unavailable` or `error` → state that validation was unavailable, preserve the verbatim request and proposed intent in the task spec, then continue under the `ROUTE` result. Gateway outage must not become a false request rejection. This is an instruction gate enforced by the `/d` contract, not a hook-enforced technical boundary. The user remains the final backstop if an agent violates it. --- ### Phase 3: DECIDE (after aligned intent) Use `ROUTE`'s `agent`, `skill`, `pipeline`, and `complexity` as-is. The script validated them against the live manifest. `agent_source: skill-default` means the pick named no domain agent and the script used the skill's owning agent. A `general-purpose` agent carries `ROUTE.fallback_reason`. When `complexity` is `null`, use `medium`, except a single one-line trivial fix → `simple`. **Routing banner** (Phase 2 intent block + routing block, both required, printed together): ``` =================================================================== ROUTING (/d): [brief summary] =================================================================== Intent (/d): -> Restated: [PROPOSED_INTENT] -> Alignment: [aligned|review|unavailable] [— issues, if any] Selected: -> Agent: [ROUTE.agent] - [ROUTE.reasoning] -> Skill: [ROUTE.skill] - [ROUTE.reasoning] -> Attached: [ROUTE.attach, comma-separated, or "none"] -> Pipeline: [ROUTE.pipeline, if set] -> Source: [ROUTE.source] (confidence: [ROUTE.confidence]) Invoking... =================================================================== ``` The `Intent` block must be populated from the Phase 2 validator run. Printing the banner with a placeholder or omitting the `Intent` block is a Phase 2 skip and is not allowed. **Gate**: Agent+skill set, banner shown. Phase 4. --- ### Phase 4: ENHANCE (attach skills) `stack` = `ROUTE.attach`, in order, plus `anti-rationalization-core`. Copy the names exactly. Do not add, rename, or drop skills: the script built `attach` from these rules, and `build-dispatch.py` rejects any name absent from `skills/INDEX.json`. | Source | Attaches | |---|---| | `JEV_RESULT.stack` (pre-route, e.g. a `.go` file with PR or security work) | its entries, first | | Agent domain floor | `programming` for Go, Kotlin, PHP, and Swift agents; `kubernetes` for `kubernetes-helm-engineer`; `frontend` for `ui-design-engineer` | | `domain_scores` at 0.6 or higher | `programming`, `frontend`, `kubernetes`, `testing`, `building-with-jev`, `research` | | `tests_requested` / `comprehensive_review` / `objective_loop_worthy` | `testing` / `review` / `workflow` | | `local_only` | `local-only` shared pattern | At most three skills are attached beyond `skill`. Two adjustments stay with you: - `comprehensive_review` attached `review` and a real multi-file diff exists: `right-size-review.py` outranks it, so drop `review` from `stack`. - `signals.research_needed` is true: add `research-coordinator-engineer` to the fan-out agents. **Fan-out agents**: union `ROUTE.agents` (your MULTI-AGENT RULE picks, each validated against the manifest) into the `research_needed` agent list, deduped. Dispatch fan-out agents as separate parallel `Agent` tool calls alongside the primary `build-dispatch.py` dispatch. **Gate**: Stack applied. Phase 5. --- ### Phase 5: EXECUTE Build the task spec with `request_verbatim` unchanged and `intent` exactly `PROPOSED_INTENT`; include any unresolved alignment issue in `gaps`, then invoke `build-dispatch.py`: ```bash python3 "$SDIR/build-dispatch.py" --json '{ "agent": "<ROUTE.agent>", "skill": "<ROUTE.skill; omit when agent-only>", "pipeline": "<ROUTE.pipeline; omit when null>", "complexity": "<ROUTE.complexity>", "model": "<haiku|opus|fable from Model choice, or gemini-3.8-flash|gemini-4-argon|gemini-4-argon-high>", "context_mode": "summary", "provider": "<anthropic|openai|google|antigravity|other>", "manual_model_override": false, "health": "-", "fallback_reason": "<REQUIRED when agent=general-purpose; omit otherwise>", "stack": ["<ROUTE.attach, in order>", "anti-rationalization-core"], "task_spec": {"request_verbatim": "<user message, unchanged>", "intent": "...", "constraints": "<applicable rules, limits, and authorization>", "decisions": "...", "gaps": "...", "acceptance": "<command> -> <expected>", "files": "<owned paths; optional line ranges>", "ownership": "<worker scope>", "operator_context": "..."}, "flags": {"worktree": false, "local_only": false, "thinking_override": null}, "token_remaining": 480000 }' ``` #### Model choice Set `model` from the task type, not `inherit`. Pass the same `model` on every Agent call, including Explore and Plan. | Task type | Model / effort | |---|---| | read-only gathering: locate, lookup, cite lines, excerpts, research fetch, log reading | haiku | | code writing and edits, debugging, features, reviews that decide, decisions, no row fits | opus / medium, then high | | high-risk: auth, data, security, concurrency, irreversible | opus / medium, then high | | major refactor or migration, architecture planning, lead of a many-agent fan-out, deep research synthesis, quantitative analysis, long unattended run, or Opus at high fell short | fable / medium, then high | Sonnet is no row's default. When the owner names a model for the task ("use fable for everything"), pass that model on every dispatch with `manual_model_override=true`. Tiers (`model_policy`): low-risk = haiku/low (read-only gathering only), standard = opus/medium, high-risk = opus/medium, max-power = fable/high. Owner, 2026-11-09: "Also I prefer opus 5.5 medium over high". Opus starts at medium on every row and tier, high-risk included. opus/high is no tier's default: reach it as an explicit pick (`model: "opus"`, `model_effort: "high"`, no `model_policy`) or as the step after a failed opus/medium attempt. `max-power`, fable/max, opus/max, and explicit opus/xhigh need `manual_model_override=true`; test fable at high before opus/xhigh. Use `inherit` only when no row or tier fits and the session model is acceptable. Escalate on a miss, when an objective check fails (tests, `tsc`, `ruff`, acceptance command) or a Jev blocker Noul is at least 0.6. In order: (1) give the agent a way to verify (test, build, script); (2) raise effort one level, up to high; (3) move to the next model up at a lower effort: haiku → opus/medium → opus/high → fable/medium → fable/high. Never raise opus past high before trying fable. Evidence: Anthropic's 2026-11-09 model guide and the owner's 2026-11-01 and 2026-11-09 orders, which win where they differ. Prices, effort trade-offs, mixing rules, history, and the re-measure method: `skills/meta/d/references/model-task-fit.md`. On Antigravity / Google AI (`provider="google"`), the tiers are `low-risk` = `gemini-3.8-flash`, `standard` = `gemini-4-argon`, `high-risk` = `gemini-4-argon-high`, and `max-power` = `gemini-4-argon-high` / `max` with `manual_model_override=true`; see `skills/meta/do/references/antigravity-execution.md`. The builder validates each name against its index, then emits the dispatch action. For Complex or creation requests, apply creation detection, plan-file gating, quality-loop, workflow dispatch, fan-out, and the `auto-pipeline` fallback. **Gate**: Agent invoked, results delivered. --- ### Post-execute: GRILL-JEV (plan/spec/design output) After any execution that produces a plan, spec, or design artifact, run `grill-jev` automatically before declaring the work complete. This applies whenever the agent's output contains phases, steps, checklists, or a structured implementation plan. Detection: the agent wrote `task_plan.md`, a spec file, a design document, or the response itself is a structured plan with numbered steps or phases. ```bash # File artifact python3 "$SDIR/grill-jev.py" --file task_plan.md --mode plan # Inline plan (write to temp file first, then grill) python3 "$SDIR/grill-jev.py" --file /tmp/plan_output.md --mode plan ``` Print the findings report. If high-signal findings exist (exit code 1): - Show findings to the user - Ask whether to address findings before proceeding or accept and move on If no high-signal findings (exit code 0): proceed, note "grill-jev: clean". Skip grill-jev when: - The output is code only (no plan structure) — use `--mode code` instead - The output is a pure research response with no actionable steps - grill-jev is itself the requested action (avoid recursion) --- ## Error handling Errors inside `jev-route.py` resolve to an `error` key and an empty evidence set. Routing continues; Phase 2 reports an unavailable validation. When changing how the router or validator builds or sends Jev requests, apply `skills/shared-patterns/jev-production-lessons.md`: stage requests at or under the reliable size, send each stage's requests together, and retry by status code. ## References - `${CLAUDE_SKILL_DIR}/references/jev-classifier-design.md` — request/response contract, what Jev still decides, design decisions - `${CLAUDE_SKILL_DIR}/references/model-task-fit.md` — model per task type, measured cost and Jev scores, re-measure method - `${CLAUDE_SKILL_DIR}/SPEC.md`, `${CLAUDE_SKILL_DIR}/EVAL.md` — maintenance contract and regression cases (load only when creating, evaluating, or redesigning this skill) - `scripts/jev-route.py`, `scripts/print-route-rules.py`, `scripts/jev-intent-align.py`, `scripts/jev_transport.py`, `scripts/jev_vercel.py`, `scripts/jev_gateway/jev_vercel_gateway.mjs`, `scripts/pre-route.py`, `scripts/routing-manifest.py`, `scripts/build-dispatch.py` - Rule text: `skills/meta/do/SKILL.md` Phase 2 (printed by `print-route-rules.py`) - Jev hook: `hooks/jev-route-injector-userprompt.py` (UserPromptSubmit) precomputes `JEV_RESULT` -
SPEC.md 6.8 KB
Loading…
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.