Claim Check

Claim Check

Investigate a premise (a ticket, a hunch, a question) deeply against the current state of the repo, and report whether it still holds and whether it can be acted on. This skill runs the investigation, including any repro needed to prove or break a claim, and stops at a verdict; it does not implement the fix during the investigation. Preserve the verdict and its evidence, then continue any supported repair the larger task already authorizes.

When to use

Someone hands you something to act on and you should not take it at face value first: a tracker ticket from an earlier audit, a suspicion you are carrying ("I think our cache double-fetches"), or an open question about the code. The code may have moved, a merged change may have addressed it in full or in part, or the framing may be stale, and equally, it may be exactly right. You want to know before anyone builds on it.

The one rule that makes this work

Every claim is a hypothesis to be checked against current repo reality, neither assumed true, nor assumed false. You are not trying to prove the premise wrong, and you are not trying to rubber-stamp it: "the premise still holds" is a first-class outcome, as good as "already handled." Conclusions come only from evidence you went and found. Anything you cannot show is unknown, and unknown means seek reachable discriminating evidence, never "probably." When reasonable paths are exhausted or access is concretely blocked, report the unresolved claim and the limit. Do not promote an unverified claim to a finding.

Access precondition: STOP if you can't reach the source

A claim-check is only as real as the evidence you can examine firsthand, and that starts with the premise's own source. Before investigating, confirm you can actually reach what the premise rests on: the ticket or PR that states it, and the repo, file, doc, or reference it is about. If you cannot: no tracker integration and the URL won't load, the PR or repo isn't accessible to this session, the reference is missing or paywalled, and the operator hasn't pasted the substance: then you do not have that premise to check. Pause that claim and say so. Continue independent claims whose substance and relevant evidence are available.

Report the gap plainly. Name the resource you could not access, what you tried, and the one thing that would unblock you (paste the ticket body, grant repo access, share the doc). Do not reconstruct the premise from the link's slug, the ticket ID, your own memory of it, or inference, and do not run an investigation against a guessed version of the claim: a confident report built on a resource you never saw is exactly the failure this skill exists to prevent.

This STOP is distinct from inconclusive. inconclusive is earned after a genuine investigation hits a wall on a load-bearing claim; the access STOP fires before you start, because the premise's substance never arrived; it is a precondition failure, not one of the verdict buckets. The partial case still proceeds when only the prior-work backlog is unreachable (the tracker won't take a query) you investigate anyway and record that the backlog wasn't swept.

Grounding a verdict

This is deliberately a deep investigation. The failure mode to beat is satisficing: taking the first plausible evidence and emitting a confident verdict. Two guards against it.

The evidence ladder. Rank what a claim rests on, strongest first:

1. a repro that ran, or the exact source lines you read yourself
2. the generating artifact the code conforms against (spec, config, codegen input)
3. a subagent's quoted snippet you can see and check
4. a subagent's summary, or a doc that merely looks consistent
5. inference / "it would make sense if"

A claim earns confirmed or refuted only from rungs 1–2 (a quoted snippet on rung 3 can support it when you can see and trust the quote). Anything resting on rungs 4–5 is unverified: keep digging; do not promote it to a verdict. A headline verdict is only as strong as its weakest load-bearing claim. If one load-bearing claim is still unverified, the verdict cannot be confirmed or refuted.

The contest test. Before you emit any verdict, for each load-bearing claim ask: what exact artifact did I examine that settles this, and would it survive the operator pushing back once? If the honest answer is "I inferred it" or "a subagent said so," you are not done: go read the source. "I found something plausible" is never a stopping condition; the stopping condition is "every load-bearing claim sits at the top of the ladder, or I have hit a real wall."

inconclusive is earned, not an escape. When a genuine deep search hits a real wall: an artifact you cannot access, an ambiguity the repo does not resolve, or context only the operator holds, the honest verdict is inconclusive. Name the specific wall and the one input that would breach it. This is legitimate only after rungs 1–2 are genuinely exhausted; reaching for it to avoid digging is the same premature-conclusion failure in a different coat.

Resolving the premise

Where the claims come from depends on the input:

• A ticket or doc (link or pasted). Get its substance, not just its link: fetch the body and acceptance criteria via a tracker integration or the URL if reachable; otherwise ask the operator to paste them. If neither lands, pause that claim (see Access precondition) rather than checking a premise you never read. Its claims arrive pre-articulated; read them as written rather than recalling them.
• A hunch you are carrying, or a bare question. No claims are stated yet. Articulate atomic, checkable claims. Proceed when the request and scope are clear; ask only when competing interpretations materially change the investigation. This step is what keeps a fuzzy input from producing a fuzzy investigation: everything downstream operates on concrete hypotheses.
• Nothing clear to check. Ask. Do not invent a claim to investigate.

Steps

Right-size the depth to the claim's blast radius before you start: a load-bearing or architectural claim earns a deep sweep; a trivial one does not. Over-investigating a typo and under-investigating a foundation are the same mistake. Then:

1. Resolve the premise (above) and state the atomic claims you are about to check. Resolve materially different interpretations with the operator; do not request confirmation of a clear claim list merely to begin.
2. Check each claim against the current repo. Fan-out is the recommended tool (especially for code and doc scanning) but you orchestrate it and may search directly when that is tighter. A subagent brief must be neutral and specific ("what does the handler at X currently do?", not "confirm X is missing") and must return evidence, not a judgment (file:line, a snippet, a commit). Keep each brief scoped to one claim or one scan. When subagents disagree, do not average them: read the disputed lines yourself and settle it on the evidence. The disagreement is usually the signal.
3. Check the provenance of the premise's evidence, not just the claim. Ask where the premise's own evidence came from (a doc, a spec, a table, a repro) and whether that is the same artifact the repo actually conforms against. A claim can be internally tidy yet rest on the wrong source of truth (a vendor doc instead of the cached spec the code is generated from). The real finding often lives in the basis, not the assertion.
4. For a falsifiable code claim, build the repro. Writing and running a throwaway repro or falsification test is the search: "unknown means go search," and a failing-then-passing probe is the strongest evidence there is. During investigation the fix stays separate from the harness that proves or breaks the claim. Preserve the verdict before any subsequent authorized repair.
5. Scan for prior or parallel work: the case the premise itself cannot see. Always search the repo's git history for commits and merged PRs that already address it in full or in part. If the tracker is queryable (the same integration or web access used to fetch the premise), also search it for sibling or duplicate tickets; if it is not reachable, say the backlog was not swept rather than implying it was. Record what you searched, so "none found" means something.
6. Adversarially re-check your conclusions. For anything that came back confirmed and anything that came back obsolete, take a second pass that tries to falsify it. This guards equally against a wrong "it's already handled" and a lazy rubber-stamp.
7. Ground the verdict, then synthesize. Run the contest test (see Grounding a verdict): confirm every load-bearing claim sits at the top of the evidence ladder. If one does not, keep digging; if a real wall blocks it, the verdict is inconclusive. Only then write the output below.

Output

Keep the report concise and verdict-first, written as plain structured text: never wrapped in a > blockquote (that renders as an annotation block that buries the verdict). It has three parts and nothing else:

1. Verdict: <bucket>: the single most important sentence, then a tight rationale in a sentence or two: why this verdict and how it was verified (the decisive method and its rung on the ladder). The evidence itself follows as short labeled bullets: one per repro case with its observed-vs-expected result, one for the root-cause chain of file:line hops, one per caveat or limit on the evidence, so parallel cases and caveats sit side by side and can be compared. For mis-scoped, add a one-line corrected framing; for inconclusive, name the wall and the one input that would breach it.
2. Prior / parallel work: <status>: the status is one word, so the landscape reads at a glance: clean (the search found nothing that addresses or constrains it), in-flight (parallel work on the same ground is underway: coordinate first), related (nothing touches it directly, but merged or adjacent work bears on the fix), or blocked (other work must land first); when several apply, pick the one the operator must act on first. Then, as prose, only what bears on the verdict: the commits or PRs that already closed part of it, and the sibling tickets that need coordination or that a fix here could regress: plus one line on what was searched (so "none found" means something). Not a catalogue of every related ticket and branch.
3. Readiness: the section an implementer scans back to, so it is shaped for scanning. It opens with a one-line call: actionable, blocked (and on what), or not actionable (and exactly what is missing or what decision unblocks it): plus where to start when there is one. The substance follows as short labeled bullets: one per candidate direction with its one-line trade-off (mark the recommendation when you have one), one per gotcha, dependency, or open unknown, each anchored to the relevant code or docs.

A confirmed premise can still be confirmed-but-blocked on readiness: surface that when it is true.

For mixed independent claims, a compact per-claim table may sit inside Verdict to preserve partial conclusions. Keep the overall verdict, prior/parallel search, and readiness dossier. Do not echo the supplied source, pad the report with unrelated work, or wrap it in a blockquote.

Persist the report only when durability or a handoff is wanted (a repo docs home, or the work scope's folder: .workbench/<work_scope>/<slug>-claim-check.md); otherwise in-chat. Any repro artifacts worth keeping land in that same scope folder, never a per-run temp directory.

Rules

• A claim-check requires firsthand access to the premise's source and the artifact it concerns. If either is unreachable and the operator can't supply it, pause only that claim and report the access gap, never substitute the link slug, the ticket ID, memory, or inference for the resource you could not open, and never investigate a guessed premise. This precondition STOP is not inconclusive (which is earned only after a real investigation hits a wall).
• Never conclude from assumption. A claim earns confirmed/refuted only from direct evidence you examined yourself (the top of the evidence ladder) never a subagent's summary or an inference; unknown resolves to search, and for a falsifiable code claim the search includes building the repro.
• Do not satisfice. The verdict is only as strong as its weakest load-bearing claim; run the contest test before emitting one, and treat "I found something plausible" as a cue to dig, not to stop.
• inconclusive (needs more information) is a first-class, honest outcome, earned only after a genuine deep search hits a real wall and never used to dodge digging. Name the wall and the one input that would breach it.
• Stay unbiased. Confirming the premise and refuting it are equally good outcomes, decided by the evidence, including the provenance of that evidence, not just the claim itself.
• Articulate atomic claims before investigating. Ask only when competing interpretations materially change the investigation; clear claims proceed.
• Subagent briefs are neutral, specific, and return evidence, never a leading question. When subagents disagree, settle it by reading the disputed lines yourself; never average conflicting reports.
• Re-check both "confirmed" and "obsolete" conclusions adversarially before reporting them.
• Right-size depth to the claim's blast radius: neither over-investigate trivia nor under-investigate a load-bearing claim.
• Lead with the verdict; the report is verdict + prior/parallel work + readiness and no unrelated material: a per-claim table may clarify mixed outcomes; no echoed source or blockquote wrapper. Bullet the verdict's evidence; open prior/parallel work with its one-word status and keep it to what bears on the verdict; shape readiness as a one-line call followed by labeled bullets.
• Finish the evidence-grounded investigation before repairing. Build the harness that proves or breaks the claim, then continue any implementation already authorized by the larger task. A standalone investigation ends at the report.