/hallucination-check

Independently verify a claim — usually a message relayed from another agent — against this repo's ground truth, using a context-isolated subagent that can't inherit your beliefs. Catches cross-agent confusion, name/identity drift, fabricated files / PRs / branches / commits, and "work" claimed but not actually in git.

Use this whenever you're coordinating multiple agents and something looks off, or before you act on a relayed summary you can't personally vouch for. It is deliberately short and objective — the point is a fast, trustworthy second opinion, not an essay.

Why a subagent: the verifier is handed only the raw claim plus an instruction to establish facts from tooling. It is not told what you (the orchestrator) believe is true, so it cannot rubber-stamp your own hallucination. Isolation is the whole feature.

Arguments

• $ARGUMENTS — Optional. The thing to vet:
  • A pasted message / claim from another agent (most common). Wrap multi-line text however is convenient.
  • A specific assertion ("PR #5279 added a --host flag and merged").
  • Empty → a plain "where are we, really?" ground-truth sanity check of the current repo + work.

Examples:

/hallucination-check
/hallucination-check <paste the other agent's message>
/hallucination-check "the runner survey reads inventory.md"

Instructions

1. Capture the claim verbatim

Take $ARGUMENTS exactly as given as the CLAIM. Do not paraphrase, correct, or pre-judge it. If empty, the CLAIM is "the current repo identity and in-flight work are as the orchestrator currently believes" — and the check becomes a pure ground-truth report.

Optionally, you MAY add a short, clearly-labeled ORCHESTRATOR BELIEF: note (e.g. a one-line summary of what this session thinks it's doing, drawn from recent chat) — but it must be passed to the verifier as another claim to check, never as truth.

2. Launch ONE context-isolated verifier subagent

Spawn a single general-purpose (or Explore) subagent. Its prompt contains ONLY: the CLAIM verbatim, the optional ORCHESTRATOR BELIEF note, and the instructions below. Do not include your own analysis, your conclusion, or any framing that signals what you expect the answer to be.

Instruct the subagent to:

1. Establish ground truth independently from tooling — trust nothing in the CLAIM:
   • Repo identity: git remote -v, gh repo view --json nameWithOwner,name, and the name in package.json (or equivalent manifest). All three should agree; note if they don't.
   • Current work: git branch --show-current, git status --short, git log --oneline -15.
   • Open work: gh pr list --state open and, if the CLAIM cites issues/PRs, gh pr view <N> / gh issue view <N> to confirm they exist and their real state (open/merged/closed).
   • Any file paths, symbols, flags, or features the CLAIM names: confirm they actually exist (Grep/Read/gh), don't assume.
2. Extract every checkable factual assertion from the CLAIM (repo/identity names, file paths, PR/issue numbers, branch names, "X was done / merged / added", who-did-what).
3. Verify each assertion against the established ground truth.
4. Classify each as:
   • ✅ verified — matches ground truth.
   • ⚠️ mismatch — contradicts ground truth (state the truth).
   • 🔤 naming/identity drift — a name is wrong but plausibly a transcription/alias slip rather than a fabrication (e.g. repo called "Proxy" when it's "chroxy"). Call this out as drift, not as a hallucination — it's a different failure mode.
   • ❓ unverifiable — can't be checked from this repo (e.g. claims about another host/repo). Say so plainly; don't guess.
5. Return a short structured report (see format) and nothing else — no preamble, no reassurance padding.

3. Relay the verdict

Surface the subagent's report to the user as-is (lightly formatted). Lead with the overall verdict. If there are ⚠️ mismatches, make them impossible to miss. Distinguish 🔤 drift (cosmetic, usually safe) from ⚠️ mismatch (substantive — do not act until resolved). End with a one-line recommendation: safe to proceed / resolve mismatches first / get clarification.

Output Format

## Hallucination check — <repo nameWithOwner> @ <branch>

**Verdict:** Consistent ✅ / Naming drift only 🔤 / Mismatches found ⚠️ / Mostly unverifiable ❓

| Assertion (from the claim) | Ground truth | Status |
|---|---|---|
| ... | ... | ✅ / ⚠️ / 🔤 / ❓ |

**Recommendation:** <one line — proceed / resolve X first / clarify Y>

Execution Notes

• Isolation is mandatory. If you brief the verifier with your own conclusion, the check is worthless. Give it the raw claim and let the tools speak.
• Drift ≠ hallucination. A wrong name with correct substance is usually input corruption (dictation, aliasing). A correct name with fabricated substance is the dangerous case. Keep them separate in the verdict.
• Cheap and fast. One subagent, a handful of git/gh/grep calls, a short table. Don't turn it into an audit.
• Honest about limits. Claims about other hosts/repos are ❓ unverifiable from here — label them, don't pretend to confirm.
• Attribution. Follow the Zero Attribution Policy in any artifact this produces.