Verify Before Claim

Verify Before Claim

Choose the least expensive safe path for the claim width. Never broaden the claim beyond fresh evidence.

The fact and scoped-task routes are self-contained. Do not reread this SKILL.md, load the local reference, inspect runtime source, or call runtime --help unless execution fails or the user requests a full schema.

Route by claim width

fact

A narrow observable statement such as a command result, file existence, config readback, or one named check. Run the smallest observation. Produce fresh Claim Evidence and a Claim Record without a boundary or review. Return claim wording, claim_kind=fact, check/result, scope limit, and stale condition.

scoped-task

Completion of a bounded small code task. Resolve ` from this installed Skill; in this source repository use runtime/handrails_state.py`. Inventory all current candidates:

python /scripts/handrails_state.py inventory \
  --compact --repository . --base  [...] \
  > 

Keep the machine inventory out of the response and tool transcript. Query its digest, counts, paths, and one target set at a time. For classification_target=hunks, classify every runtime hunk_id and never the parent Candidate ID; for classification_target=candidate, classify its candidate_id once. The private inventory path must be outside participating worktrees or ignored and undeclared. Pass exactly this shape through stdin:

{
  "task_id": "stable task ID",
  "outcome_revision": "r1",
  "inventory_digest": "runtime inventory digest",
  "logical_units": [
    {
      "target_id": "runtime candidate or hunk ID",
      "classification": "include-claim | exclude-proven-unrelated",
      "rationale": "task-specific reason",
      "authorization": "authorized | not-applicable"
    }
  ]
}

Use each runtime target ID exactly once. include-claim requires authorized; exclude-proven-unrelated requires not-applicable. Never invent an ID or digest. Rationale is digest-bound semantic content. Choose a private boundary path outside participating worktrees, or an ignored Handrails evidence path that is not passed through --declare-ignored. Never write machine evidence inside a declared ignored candidate; that would immediately stale the state. Confirm once with the same capture arguments:

python /scripts/handrails_state.py confirm \
  --compact --kind claim --classifications - \
  --boundary-out  \
  --repository . --base  [...]

Treat the persisted boundary as opaque. Never open, parse, query, or print that artifact; pass it unchanged only to capture. Read scalar digests from the compact confirm result and later compact capture result, not from the full boundary file.

Run narrow checks. Then validate the exact confirmed artifact:

python /scripts/handrails_state.py capture --compact \
  --boundary 

Require its Boundary, Inventory, and Code State Token digests to match the confirmation. Drift blocks the claim. Do not reconfirm merely to reconstruct a discarded payload. The Claim Record carries claim_boundary_digest, not Review ID.

high-assurance

Merge-ready, integrated, production/deployment safe, no-regressions, or similar. Read [the local v2 contract](references/usable-loop-v2.md). Require a formal Review Record, validate its boundary, and recapture current state:

python /scripts/handrails_state.py capture \
  --compact --boundary 

Outcome Revision, Boundary Digest, and Code State Token must match exactly. The Claim Record carries review_id, not Claim Boundary Digest. If wording is ambiguous, widen the route or explicitly narrow the claim. Never hide a review transition inside a micro fact.

Produce claim evidence

Only this skill produces Claim Evidence with fresh | stale | blocked. Development evidence is orientation only. Record Evidence ID, task/outcome, exact state and environment, boundary/review link, check/result, narrow scope, capture time, and stale conditions. Outcome, boundary, relevant code state, or environment changes stale it; rerun the narrow check.

The Claim Record states exact wording, claim kind, Evidence IDs, supported | limited | unsupported | blocked | stale, scope limits, and invalidation. Never say done, safe, integrated, or regression-free beyond the evidence.

Direct result

For a fact, use at most 12 lines. For a scoped task, use at most 30 lines. Return conclusion, evidence, scope limit, and necessary next step. Expand only on explicit request or a Checkpoint trigger. No mandatory Feedback Loop backflow.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: Aquish-Lee
• Source: Aquish-Lee/agent-handrails
• License: MIT

Install and usage instructions live in the source repository linked above.