Issue Assessment

issue-assessment — triage a gflow-cli issue honestly

Read-only conductor. Verify the reporter's claim against the code, tests, docs, KNOWN_ISSUES.md, and auto-memory; classify it; judge whether it can be verified end-to-end in the current environment; and draft a reply. The output is a standard artifact a human or the issue-resolve skill can act on.

Core principle: never assert more than the evidence supports. A claim is CONFIRMED only with line-level code evidence or a reproduction; a fix is "verified" only after running it on the affected surface. Honest "can't verify here" beats a false green check — a bounced fix costs more trust than an accurate "not yet."

When to invoke

• A new or updated GitHub issue needs a verdict before anyone spends effort.
• An autonomous run (hermes-ops) reacts to an issue labelled for triage.
• You're about to "just fix" a reported bug — assess first; the scope decision (reply-only vs hand to issue-resolve) depends on this.

Skip for: issues that are obviously feature requests routed elsewhere, or already-triaged issues entering implementation.

Invocation

/gflow:issue-assessment <issue number or URL>

This repo's skills/*/SKILL.md are plain Markdown — invoke by reading the file (via the .claude/commands/gflow/* wrapper), never Skill(skill=...).

Protocol

1. Ingest

gh issue view <N> --json title,body,comments,labels,author,state. Extract: the claimed symptom, environment (OS, version, install method), exact repro steps, and any logs/error classes the reporter pasted.

2. Verify (read-only)

Dispatch a search/Explore agent (keep your own context clean) to corroborate or refute the claim against the real tree. Always check, in order:

• the source path(s) the symptom implicates — cite file_path:line_number;
• KNOWN_ISSUES.md (is this Open / Mitigated / Resolved already?);
• open issues/PRs (gh pr list, gh issue list) for duplicates or in-flight fixes;
• auto-memory for prior context on the surface. Disprove parts of the reporter's framing where the code says otherwise (e.g. browser_engine: playwright is the engine axis, not the channel) — a precise correction is more useful than agreement.

Name the affected SURFACES, not just the affected code. A reporter hits one surface; the defect usually spans both. gflow ships most capabilities twice — CLI command and MCP tool — so state explicitly whether the issue reproduces on the CLI, on the MCP tool, or on both, and whether a fix in one automatically fixes the other (it does when both route through the same transport; it does not when the MCP path carries its own params through worker/codec.py). Getting this wrong scopes the whole downstream fix wrong: a "CLI bug" that is really a shared-transport bug leaves MCP users broken after the issue is closed.

3. Classify — exactly one verdict

4. e2e-gate — what would verification actually require?

Classify the verification cost before claiming anything is fixed:

• Browser-free (pure-Python logic, Gemini tool-path, unit/lint/type, recording-verif) → verifiable anywhere, including the headless VPS. Run it.
• Headed-Flow-browser required (generation, selector, auth, reCAPTCHA) → not verifiable on a headless or wrong-OS host. Verdict tilts to LIKELY-BUG / NEEDS-E2E; never claim success; the final check is a human on the affected surface. (See memory: done-means-e2e-verified, pr-must-verify-on-affected-surface.)

5. Report (the artifact)

Produce the reply below. Post it only if the autonomy gate allows (autonomous runs may comment; otherwise surface for a human to send).

**Assessment of #<N>: <verdict>** (confidence <N>/10)

Restated claim: <one line>.

Findings:
- <evidence as file:line> …
- <corrections to the reporter's framing, if any> …

Root cause / hypothesis: <what and why, or "unconfirmed because …">.

What we need next:
- <the single discriminating diagnostic — for NEEDS-INFO/NEEDS-E2E>, or
- <draft PR link — if issue-resolve ran>, or
- <why this is a dup/invalid/wontfix>.

6. Hand-off decision (graded, not a hard stop)

• Verdict ∈ {CONFIRMED-BUG, LIKELY-BUG} and scope is single-surface/localized and a fix is verifiable in this environment → chain to issue-resolve (Phase 9) or predict (Phase 2).
• Otherwise → reply only; the next step is a human or more info.