Claude Code + Jev

You (the current Claude Code session) implement, review and test the change with your normal tools. Jev only judges which repository files are relevant. Do not start Codex, Astra or another coding agent for generation. This Skill does not change your model or effort.

Entrypoint and credentials

Harness v0.7.3 pins new Jev requests to jev-1.14.0, including requests made from older plans. It does not follow model aliases or change the active Claude model. The same native runtime serves Desktop, CLI and Claude Code. Keep the executable and checkout at the same release. After an upgrade, use a fresh plan and new output directory; preserve old judgments and failure records. Checking a saved 1.13 result does not turn it into a 1.14 evaluation. Model-specific cache keys separate the requests without deleting old caches. For upgrade/rollback details, read Model updates and saved judgments.

The supported runtime is the prebuilt Go executable in bin/astra-jev. Python and Go are not needed during coding. Build once with scripts/build-native.sh (Go 1.26+ and a C compiler), or use the matching platform release archive. Do not rebuild, benchmark or run model comparisons during normal coding. The historical .py files remain development references.

Run "<base directory of this Skill>/scripts/context.sh" <command> ... using the absolute base directory shown when the Skill loads. Quote it; it may contain spaces. The wrapper resolves its symlink to bin/astra-jev claude-code in the linked Harness checkout. doctor reports the checkout and whether a key is available, without calling a provider or printing the key.

An existing TYPESAFE_API_KEY takes precedence. On macOS the helper can read login Keychain service astra-jev-harness, account TYPESAFE_API_KEY, into its own process only. Never print keys, pass them as arguments, search .env files, or forward them. A missing key blocks new Jev requests, not valid cached judgments or otherwise authorized local work. A Keychain timeout does not prove the key is absent.

Workflow choice

Jev supplies relevance judgments; it is not a prerequisite for every authorized coding action. A user's no-Jev/local choice applies to the current task and its in-scope follow-ups until they change it. Do not ask for the same permission again or run Jev planning or credential checks merely to satisfy this Skill.

When required files exceed planning limits, credentials are unavailable, or a provider fails, stop the affected Jev operation. State the reason and “Jev selection unverified / Jev未検証 — continuing locally” once, then continue already-authorized investigation, implementation and tests using Local continuation. If successful Jev evaluation is an explicit acceptance condition, keep that item incomplete while independent local work continues. Repository safety rules and external-action boundaries still apply.

Context workflow

1. Confirm the task and checkout. Read its CLAUDE.md/AGENTS.md, check branch, HEAD and status, and preserve existing changes.
2. On the Jev path, before reading implementation file bodies, write a focused task file outside the target and run plan --repo <absolute-root> --task-file <file> --out <new-dir-outside-target>. Read PLAN.md only (not plan.json). Untracked files need repeated --include-file; pin known task-critical files with --focus-file. For large repositories, scoping, or planning failures, read references/workflow.md. During local continuation, a sufficient existing plan is optional; bounded native reads may supply the real task scope.
3. On the Jev path, run select --plan <plan-dir> --out <new-selection-dir> --max-calls <n> --mode auto within the existing authorization and request cap. auto keeps all candidates and skips Jev under 12,000 source bytes; it is not an offline guarantee. Use --mode jev when a real Jev evaluation is requested, and --mode local for local continuation with a sufficient plan. Claude's wrapper supports local select/check/read directly. The cap defaults to 4 (maximum 24). An external-call ban rules out new Jev requests. The helper makes no automatic retries, and a failed attempt may still be billed. Preserve failure receipts before continuing locally.
4. When using a selection, read REPORT.md and only the metadata you need from selection.json: decisions, unjudged_paths and metrics. Do not dump context.json. Read bodies with read --selection <dir> --path <relative-path> --start-line <n> --end-line <n> (default 80 lines, maximum 200 lines or 24 KB). Uncertain, unjudged and dependency files are retained on purpose. A relevance score is not a security check.
5. Run check --selection <dir> before editing from a selection, or recheck local contents when using native reads. Then implement and test directly in this session within the user's scope. The plan and selection grant no permission to edit, commit, push or take external actions.
6. On the Jev path, source changes require a fresh plan for more context. During local continuation, refresh local source/freshness checks without forcing a return to Jev. Inspect saved attempts before deciding whether another run is necessary; never blindly repeat a potentially completed operation.

Keep routine replies focused on changes, verification and actionable blockers. Do not append or aggregate Jev request counts, token usage, cache reuse, cost or output-byte statistics for narration. Show accounting only when explicitly requested, using saved receipts without extra calls or benchmarks. Keep attempts, completed calls, reuse and missing usage distinct. Internal receipts remain available for limits and failure recovery. Report implementation and tests independently of selection; the helper cannot measure Claude Code conversation tokens or cost, and a smaller context is not a proven saving.

The helper rejects plans and selections from the Codex Desktop or CLI surfaces. It has no run, verify or apply commands. Selection artifacts contain source text. Keep them outside the target repository and out of Git.

For offline policy comparison (compare), evidence packets, caches and scope limits, read references/workflow.md.