Cheshi: Skill for Claude Code

Jev research report

Run this skill from the Cheshi repository root. The registered workflow executes real web research and Markdown writing; it is not the mock branch demo.

1. Derive a topic and 1–8 concrete requirements from the user's request. Use the requested language. Reuse available source notes when relevant; do not invent evidence to make the initial judgment pass.

2. Write a request JSON file under a task-specific directory in out/skill-flow/:

   {
     "topic": "JevのChoiceとNoulの違い",
     "requirements": ["それぞれの返り値", "コードでの分岐方法"],
     "sources": []
   }
   

   Optional sources are objects with title, public HTTP(S) url, and notes. At most six sources are accepted. Keep notes within 1,600 characters each and the complete evidence within 23,000 UTF-8 bytes. These limits bound request size; they do not decide whether the evidence is sufficient.

3. Invoke the registered skill:

   bun run scripts/skill-flow-run.mts --skill jev-research-report --input out/skill-flow/REQUEST.json
   

   Substitute the actual request path. Preserve the invoking account's CODEX_HOME; do not switch accounts or copy credentials. This requires the checkout's Electron runtime, a Codex subscription with Luna low, and network access. The existing saved Jev key is read inside a headless Electron process.

4. Open the printed result.md and, on success, report.md. Present the report link and any material limitations. fail means the workflow did not produce a completed report; do not substitute a made-up successful result.

Jev selects yes or no; ctx.jev() returns a boolean. Insufficient evidence triggers additional research. A rejected draft triggers a new writing attempt, whose new contents are checked again. Normal no never triggers Luna reassessment of the same judgment. Jev service errors or unavailable credentials trigger one Luna low/default decision using the same state and condition. Cancellation and invalid inputs do not trigger fallback. If both providers fail, the run records an error and stops.

Source URLs are fetched independently and their readable text is stored in evidence files. Jev checks whether each source supports its notes. Unreadable or rejected sources are excluded and recorded in sources-*.json. Previously verified sources are retained while they fit; at the capacity limit, Jev compares old and new sources for usefulness to the requirements. selection-*.json records comparisons and excluded sources. Insufficient evidence triggers the remaining research rounds. The first sufficiency yes proceeds to document writing. Each finished requirement and section is then checked with Jev yes/no. Markdown parsing permits literal code, arrays and placeholders while rejecting active prose links, images and HTML. Final rendered citations are checked against the verified sources. Writing has at most three attempts. Validation failures provide error codes and field paths to the next writing attempt; writing-*.json retains the response, status and issues. Provider errors, cancellation and execution limits do not trigger a rewrite. Raw draft capture is capped at 64,000 UTF-8 bytes and marked when truncated; only untruncated drafts within 18,000 bytes are sent as feedback. Large citation checks split the prose at sentence boundaries while retaining all cited notes. A single sentence that cannot fit with its evidence fails validation without truncation. Only a validated document is saved as report.md. After two additional research rounds, another no stops with fail. Research may use web search; judgment and writing cannot use tools. The runner records input, researched evidence, decisions, provider attempts and Codex usage separately under a unique directory. It preserves prior output files and never installs an autopilot loop.

SKILL.md is discovered through the standard repository skill catalog. The agent invokes the command; workflow.mts exports the executable skill definition. The generic registry loads that module and runs its input, execution and validation functions. These modules are trusted repository code and run with the command’s permissions. The runtime does not equate a Jev no with a failed task. Selecting a skill alone does not silently execute code. The skill currently requires this checkout and is not bundled into released Cheshi applications.