jev-branch-check

Jev branch check

Run this example from the Cheshi repository root. It uses the exported branchCheck(ctx, question) function in workflow.mts. SKILL.md tells the agent what to run; its code blocks are not automatically executed. This example is repository-local and is not automatically installed in the skill catalog.

The workflow calls ctx.jev({ state, condition }). The adapter asks Jev to choose yes or no with a Choice question and follows the returned choice directly:

Jev choice	await ctx.jev(...)	Markdown first line
yes	true	success
no	false	fail
unavailable key, request failure, or malformed response	throws SkillFlowDecisionError	fail

Use the boolean directly in if (judgment). The runner keeps decision metadata for reports; workflows return only { outcome: 'success' | 'fail' }. Request errors throw before either branch executes. The runner records them as judge_error, with a distinct error reason and null value. The adapter does not read, compare, or expose response probabilities or confidence. No numeric decision policy is applied.

Run the deterministic local cases:

bun run scripts/skill-flow-demo.mts

This uses mock responses for the research cases, HTTP failure, and malformed response cases. It never contacts TypeSafe. The summary's success means all expected branches were observed, including cases whose own output is fail.

To test Jev itself using the key already saved in Cheshi Settings, run:

bun run scripts/skill-flow-demo.mts --live

The CLI starts this checkout's Electron runtime without opening a window. It reads the saved encrypted key using safeStorage, as the previous Jev benchmarks did. The key stays inside that process; it is never printed or copied to a file. The runtime uses a temporary data directory so the app's existing data is unchanged. An explicit TYPESAFE_API_KEY or TYPE_SAFE_AI environment value takes precedence.

Live mode sends three variants of the official-source notes in workflow.mts: all three sources, only the Noul source, and no sources. These notes were gathered from TypeSafe's Noul/HTTP API documentation and LangChain's Interpreter Skills article on 2026-09-21. The requested presentation covers Noul semantics, the HTTP contract, and executable-skill roles. Expected results are yes/no/no, fixed before calling the model. This small check does not establish general research quality.

The runner evaluates those notes; it does not browse or refresh sources itself. Each run saves the exact notes and question in inputs.md beside the summary. It uses jev-latest, a 20-second request timeout, and no retries. The returned model version and available token counts appear in the case reports. Missing usage stays unknown. Do not put keys in commands, reports, or code.

Both modes print the summary path under a new out/skill-flow/<mode>-<id>/ directory. Use --output-root DIRECTORY to change its parent. Existing reports are never overwritten. A failed expectation sets exit code 1 after saving the report. Report-write failures also exit with an error, rather than claiming a saved result. No research/PPT tools or desktop chat behavior are changed.

Validation:

bun test desktop/test/skill-flow-judge.test.ts desktop/test/skill-flow-runtime.test.ts
bun run desktop:typecheck
bun run codegraph:server:typecheck
node --input-type=module -e "await import('./desktop/lib/skill-flow-runtime.mts'); await import('./scripts/skill-flow-demo.mts')"

API references: HTTP API, Choice selection.