Jev Advisor

Use the local Jev Advisor as a private decision helper when a task has a meaningful choice between tools or between continuing and stopping. It is one input to judgment, not a controller.

When To Consult

Consult at material decision points:

• understand: decide whether evidence is missing.
• explore or trace: choose among file, shell, browser, network, Git, or external-service reads.
• implement: compare edit or execution candidates.
• verify: choose tests, builds, checks, or runtime validation.
• recover: reassess after failure, timeout, partial success, or conflicting evidence.
• finish: judge whether observable completion evidence is sufficient.

Skip an obvious continuation with no meaningful alternative. Batch all realistic candidates into one request. Never consult the advisor about invoking itself.

Request Construction

Build one compact UTF-8 JSON request from context visible to Codex. Include:

• goal, observed facts, explicitly marked inferences, and open questions;
• recent tool outcomes;
• workspace and Git fingerprints;
• the dynamic tool registry, annotations, schemas, and realistic candidate actions.

Do not include hidden reasoning or claim access to chain-of-thought. Summarize large output and sensitive values.

Use the installed command when available:

jev-advisor request.json

During repository development, use:

python -m jev_advisor.cli request.json

If a calibrated local endpoint is explicitly configured, pass its matching immutable model and threshold files. Do not download weights, install dependencies, or start a service automatically. If the command is missing, times out, or returns invalid data, continue normally. model_version: heuristic-0.1 means the advisor failed open.

Apply Advice

Use recommended_action, candidate success/latency/information-gain/token-cost scores, named risks, and ask_user as evidence for the next step. Keep the choice consistent with the user's request, current evidence, host permissions, and safety policy.

• A recommendation never expands authorization or overrides host approval.
• The advisor can suggest a tool but cannot call, approve, deny, intercept, or retry it.
• Treat external writes, destructive operations, credentials, data exposure, network access, financial actions, stale evidence, incomplete parameters, and command execution as risk-bearing.
• Reassess after an action changes observable state; otherwise allow the state-hash cache to serve repeated decisions.
• Keep advice private. Do not quote it to the user unless they explicitly ask to inspect advisor behavior.

Treat finish as valid only when completion evidence exists: no unresolved question or failed/partial recent result, plus a passed or completed verification event. Model confidence alone is never completion evidence.

Read references/protocol.md when constructing or debugging the JSON protocol.