Diagnose

guild:diagnose

Diagnoses Guild's own runtime failures and creates a gated self-fix path. This skill is allowed to inspect repository files and .guild/ state, but it must not edit anything until the user approves a concrete fix plan.

Inputs

Accept a raw argument string from /guild:fix:

• A run id, e.g. run-2026-05-01-a62fc657.
• Free-text operator context, e.g. "PostToolUse audit log is empty".
• --review=cross to request Codex adversarial review of the diagnosis/fix plan.

If no run id is supplied, inspect the most recent .guild/runs/* directories. If no context is supplied, infer symptoms from telemetry gaps and error rows.

Read Scope

Read only the smallest useful set first:

1. .guild/runs/current-run-id, if present.
2. The requested run directory, or the latest 3 .guild/runs/* directories by mtime.
3. For each inspected run, read if present:
   • metadata.json
   • events.ndjson
   • logs/v1.4-events.jsonl
   • logs/summary.md
   • review.md, verify.md, assumptions.md
   • handoffs/*.md
   • codex-review/*.md
   • captured subprocess logs under benchmark artifacts when the run is a benchmark import (_subprocess.stdout.log, _subprocess.stderr.log)
4. Relevant plugin surfaces only after evidence points there:
   • hooks/hooks.json
   • hook handlers under hooks/
   • command files under commands/
   • meta-skills under skills/meta/
   • scripts under scripts/

Do not bulk-read unrelated project source. Do not inspect credentials or files outside the repo and .guild/ state.

Diagnosis Workflow

1. Normalize arguments. Separate run id candidates, free-text context, and flags. A run id matches run-* or an existing .guild/runs/<id> directory.
2. Collect evidence. Summarize event counts, last timestamps, hook names, tool-call events, loop events, error rows, missing expected files, and recent failed verification/review findings.
3. Classify failure. Use the first matching category:
   • telemetry-scope — events from multiple invocations share one run id, or current-run-id is missing/stale.
   • hook-registration — bootstrap appears but expected hook handlers do not write events.ndjson or logs/v1.4-events.jsonl.
   • audit-log-gap — v1.3 events.ndjson exists but v1.4 JSONL is absent, or Pre/Post tool pairing sidecar is orphaned.
   • loop-telemetry-gap — loops ran but loop_round_start / loop_round_end rows are missing or unpaired.
   • agent-team-stall — task-created/task-completed/teammate-idle evidence shows blocked ownership or missing handoff receipt.
   • codex-review-gap — --review=cross / codex_review requested but trail is missing, invalid, or explicitly skipped.
   • verification-failure — verify.md failed for unmet success criteria.
   • unknown — evidence is insufficient; list the missing evidence.
4. Root cause. State the most likely cause, confidence (high / medium / low), and exact evidence paths.
5. Fix plan. Propose the smallest fix. Each fix item must include:
   • files to edit
   • intended behavior change
   • risk/blast radius
   • validation command
   • rollback note

Adversarial review — G-diagnose (broker)

G-diagnose is a skill-internal gate (docs/v2/adversarial-review.html §Gate ownership): it fires here inside the fix/diagnose flow at the diagnosis→approval boundary. Wire the review broker at this boundary, not guild:codex-review directly: the broker is the host-agnostic front door, and guild:codex-review survives only as the internal Codex adapter the broker dispatches to (docs/v2/adversarial-review.html §The review broker).

If --review=cross is present, or project config resolves codex_review: true, review the diagnosis and fix plan before asking the user to approve edits.

Config resolution:

npx tsx ${GUILD_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-$HOME/.local/share/guild/dist/claude-code}}/scripts/read-guild-config.ts [--cwd <repo-root>] [raw /guild:fix flags]

When codex_review resolves true the G-diagnose review is required, so the broker obeys the SK-9 gate-pass rule (skills/meta/review-broker/SKILL.md §"Gate-pass rule"): a required review never clean-skippeds.

1. Invoke guild-review-broker on the diagnosis report path: gate=G-diagnose artifact_path=<report-path> run_id=<diagnose-run-id> author_host=<run author host>.
2. The broker resolves the reviewer host. When a different-family reviewer is reachable, the review is STRONG. When no cross-host reviewer is available (Codex absent / same-family / non-selectable), the broker degrades to the WEAK same-host path — a fresh-context same-host subagent reviews, independence: weak, recorded, and still gated by the 5 conditions. It does not resolve status: "skipped": a required review that found no cross-host reviewer is a recorded weak review, never a clean skip (that would be the gate-bypass killed in SK-9). skipped happens only when review is not required (no --review=cross and config off → review_required == false). author_host is the host that produced the diagnosis (resolved from the run-start preflight snapshot; claude on a Claude-hosted run).
3. If the broker returns status: "rework", revise the fix plan before the user approval gate. On "satisfied" (STRONG or WEAK) or "force_passed", proceed to the approval gate normally; a true "skipped" only occurs when review was not required.

Adversarial review is advisory for diagnosis. The user approval gate remains required whether the broker is satisfied (STRONG or WEAK), force-passed, or — when review was not required — skipped.

Report

Write the diagnosis report before any edits:

.guild/diagnose/<timestamp>-<slug>.md

Use this structure:

# Guild Diagnose — <timestamp>

## Input
- args:
- selected_run_ids:
- operator_context:
- codex_review:

## Evidence
- files_read:
- event_counts:
- missing_expected_artifacts:
- errors:

## Diagnosis
- category:
- confidence:
- root_cause:
- evidence:

## Fix Plan
1. ...

## Approval Gate
status: pending

User Approval Gate

Before any edit, show the report path and the fix plan, then ask:

Apply this Guild self-fix plan? [approve / edit-plan / diagnose-only / abort]

Allowed responses:

• approve — apply only the listed edits.
• edit-plan — revise the plan, update the report, and ask again.
• diagnose-only — stop after writing the report; no edits.
• abort — stop without edits; leave the report as the audit trail.

Do not treat silence, ambiguous replies, or Codex approval as user approval.

Applying Fixes

When approved:

1. Edit only files listed in the approved plan.
2. Preserve unrelated user/team changes in the working tree.
3. Run the validation commands listed in the plan when practical.
4. Update the report's Approval Gate section to approved and append:
   • files changed
   • checks run
   • results
   • residual gaps

If an edit would touch high-trust surfaces (hooks/, commands/, .claude-plugin/plugin.json, mcp-servers/), call that out in the approval prompt.

Upstream escalation — plugin-level diagnoses (consent-gated)

A confirmed diagnosis whose root cause lives in the Guild plugin itself (any category except unknown where the evidence points at plugin code/hooks/commands — not at project .guild/ state or user config) should ALSO be offered upstream, whether or not a local fix was applied — a locally-patched install still leaves every other installation broken. Route it deterministically, never by hand:

1. Write the diagnosis as a one-element RunLearningFinding[] ARRAY in .guild/feedback/<run-id-or-timestamp>/findings.json (the CLI rejects a bare object) (summary = root cause one-liner, details = the report's Diagnosis section, evidence_refs = the plugin paths implicated, proposed_change = the fix plan or applied fix).
2. Run npx tsx ${GUILD_PLUGIN_ROOT:-${CLAUDE_PLUGIN_ROOT:-$HOME/.local/share/guild/dist/claude-code}}/scripts/feedback-triage.ts triage --run-id <id> --findings <that file> — it classifies (code, not judgement) and writes a sanitized issue draft for plugin/mixed findings.
3. Ask the operator whether Guild may file the draft as a GitHub issue (it contains no user-specific information — the redaction stack strips private paths, tokens, emails). Yes → … file --finding <id> --approve "<operator>"; no → --deny. Non-interactive ⇒ leave the draft pending and name it in the summary. The CLI refuses to file without the explicit approval.

Output

Return a concise summary:

/guild:fix — <diagnose-only | fixed | aborted>
Report: .guild/diagnose/<timestamp>-<slug>.md
Diagnosis: <category> (<confidence>)
Root cause: <one sentence>
Changed files: <list or none>
Checks: <list>
Residual gaps: <list or none>
Pending plugin feedback: <.guild/feedback/<id>/<finding>.draft.md … or none>