validate
Freshly judge exact subject content against bead or caller acceptance, optionally persist verdict.v2 for a declared consumer, and stop. Triggers: "validate", "independently validate", "vibe".
Install
npx skills add https://github.com/boshu2/agentops/tree/main/packs/agentops-executor/agents/validator/skills/validate
claude plugin marketplace add https://llmmart.ai/marketplace.json && claude plugin install boshu2-agentops@llmmart
git clone https://github.com/boshu2/agentops.git
The skills CLI installs just this skill, for any of its supported agents. Claude Code installs the whole boshu2/agentops collection as a plugin from our marketplace. Git is the plain clone.
Skill manifest
Validate
Independently judge one exact subject against the acceptance in its existing
bead or caller source, return one semantic result, and stop. Validate is the
sole verdict.v2 writer when persistence is requested. It never asks the model
to reconstruct Plan or Candidate packets.
Preconditions
- The subject is a nonempty implementation candidate: the manifest lists at
least one entry, and
store-verdictrefuses an empty one. Plans, audits, reviews, and other control artifacts are not completion subjects unless the caller explicitly requested document review. - The intent source is available as a caller-owned artifact or runtime-owned content-addressed snapshot; its acceptance digest is derived automatically.
- The subject manifest still matches the subject.
- Author and validator context IDs are explicit.
- Freshness is explicitly attested with
source: runtime | callerand an attester identity.
Missing, colliding, or unattested identities produce NOT_PROVEN. This is a
declared trust fact, not cryptographic proof that contexts were isolated.
Cross-model fresh validator (caller-elected)
A caller may request that the fresh validator run on a different model than
the author. Dispatch via the controller-session recipe in
the validate model-dispatch recipe (codex-exec and/or NTM panes,
probed at runtime). Record author and validator model_identity in evidence
refs and freshness attestation notes — do not change verdict.v2 schema. If
the requested validator model has no live adapter, disclose the unsatisfied
diversity request and proceed same-model. Single fresh validator remains the
default shape.
Mutating-check quarantine
Before running any acceptance-listed command, classify it as read-only or
subject-mutating. Regen scripts, sync scripts, formatters, and anything with
--force are subject-mutating until proven otherwise. Never run a
subject-mutating check against an uncommitted subject: on 2026-07-15,
scripts/test-ci-deterministic-gates.sh regenerated skills-codex/ from HEAD
mid-validation and destroyed the uncommitted subject, forcing NOT_PROVEN
(verdict b6e759dd...cb6a); only restoring the subject and revalidating in a
fresh context produced the PASS (e9b6cdb8...37b9). If a mutating check is
genuinely required by acceptance, run it against a disposable copy or a
committed subject, never the judged working tree.
Workflow
- Recompute and compare
subject-manifest.v1usingpython3 skills/validate/scripts/validate.py manifest. The helper uses only filesystem content; Git commit/tree IDs are optional metadata. Derive the manifest at the start of validation and re-derive it at the end; any mismatch between the two is subject mutation and returnsNOT_PROVEN. - Confirm the intent-source digest has not changed since implementation. If
the subject changed or complete changed-path coverage cannot be derived,
return
NOT_PROVEN. - Adjudicate the actual diff, not a declared path list: compare
runtime-derived actual changed paths against the intent's scope classes. A
proven out-of-scope path returns
FAIL; incomplete scope evidence returnsNOT_PROVEN. - Inspect the exact subject and factual evidence. Reported exit codes are
claims, not evidence: re-execute the claimed proofs that bear on acceptance
(see the freshness rules below for when a digest-bound receipt suffices).
Judge every acceptance criterion and record criterion-level results,
findings, evidence references,
checked, andnot_checked. - Choose exactly one semantic result:
PASS,FAIL, orNOT_PROVEN. Return it with criterion results, findings, evidence references,checked,not_checked, the acceptance and subject identities, distinct author and validator context IDs, and the freshness attestation. PASS requires distinct identities, explicit freshness, nonempty checked scope, top-level evidence, and evidence for every criterion. - Only when the caller requests machine-readable evidence or a declared
downstream consumer requires it, persist canonical
verdict.v2withstore-verdict --draft <draft.json> --intent-source <resolved-intent> --subject-manifest <manifest.json> --author-context-id <id> --validator-context-id <id> --freshness-source <runtime|caller> --freshness-attester-id <id> --scope-result <PASS|FAIL|NOT_PROVEN>. The helper snapshots the exact resolved intent under<workspace>/.agents/ao/intents/sha256/<digest>.intent, then computes and injects intent and subject digests plus author, validator, and freshness facts. Identity and changed-path facts come from runtime-derived inputs and receipts, not model transcription. Storage defaults to<workspace>/.agents/ao/verdicts/sha256/<digest>.json; callers may provideverdict_dir. - Return the semantic result and, when persisted, the artifact path and digest. Stop.
The digest is SHA-256 over canonical JSON with artifact_digest omitted. Writes
use a same-directory temporary file, flush, fsync, and atomic rename. Identical
existing content is idempotent success; conflicting content is an integrity
failure represented by NOT_PROVEN.
Freshness without duplication
Fresh validation means independent judgment over the exact subject. It does not require mechanically replaying every author command. Verify intent identity, scope, evidence digests, and every acceptance criterion; independently rerun the risk-critical, uncertain, or insufficiently evidenced checks. A digest-bound deterministic receipt may prove routine facts. Replay an expensive full suite only when acceptance requires that result or the supplied receipt cannot establish it.
Boundary
Validate emits no WARN, confidence, disposition, briefing learning, owner, next action, repair, retry, replan, helper, escalation, tracker, Git, release, closure, or delivery state. Generic provenance may record a verdict later, but ledger availability cannot change its validity.
Files (agentops)
-
references
-
validate.feature 1.9 KB · in bundle
-
-
scripts
-
check_contract_corpus.py 5.4 KB
Loading… -
test_validate.py 13.1 KB
Loading… -
validate.py 26 KB
Loading… -
validate.sh 1 KB
Loading…
-
-
SKILL.md 6.8 KB
--- name: validate description: 'Freshly judge exact subject content against bead or caller acceptance, optionally persist verdict.v2 for a declared consumer, and stop. Triggers: "validate", "independently validate", "vibe".' practices: - design-by-contract - llm-eval-harness - content-addressed-storage hexagonal_role: driving-adapter consumes: - subject-manifest.v1 produces: - subject-manifest.v1 - validation-result - verdict.v2 context_rel: - kind: customer-of with: plan - kind: customer-of with: implement skill_api_version: 1 user-invocable: true metadata: graph_root: true tier: judgment dependencies: [] capabilities: [compute_subject_identity, judge_acceptance, return_validation_result, persist_verdict] effects: [write_verdict_artifact] canonical_status: canonical disposition: keep output_contract: 'PASS | FAIL | NOT_PROVEN with criteria, evidence, checked/not_checked, identity, and freshness; optional schemas/verdict.v2.schema.json persistence' --- # Validate Independently judge one exact subject against the acceptance in its existing bead or caller source, return one semantic result, and stop. Validate is the sole `verdict.v2` writer when persistence is requested. It never asks the model to reconstruct Plan or Candidate packets. ## Preconditions - The subject is a nonempty implementation candidate: the manifest lists at least one entry, and `store-verdict` refuses an empty one. Plans, audits, reviews, and other control artifacts are not completion subjects unless the caller explicitly requested document review. - The intent source is available as a caller-owned artifact or runtime-owned content-addressed snapshot; its acceptance digest is derived automatically. - The subject manifest still matches the subject. - Author and validator context IDs are explicit. - Freshness is explicitly attested with `source: runtime | caller` and an attester identity. Missing, colliding, or unattested identities produce `NOT_PROVEN`. This is a declared trust fact, not cryptographic proof that contexts were isolated. ## Cross-model fresh validator (caller-elected) A caller may request that the fresh validator run on a different model than the author. Dispatch via the controller-session recipe in the `validate` model-dispatch recipe (`codex-exec` and/or NTM panes, probed at runtime). Record author and validator `model_identity` in evidence refs and freshness attestation notes — do not change `verdict.v2` schema. If the requested validator model has no live adapter, disclose the unsatisfied diversity request and proceed same-model. Single fresh validator remains the default shape. ## Mutating-check quarantine Before running any acceptance-listed command, classify it as read-only or subject-mutating. Regen scripts, sync scripts, formatters, and anything with `--force` are subject-mutating until proven otherwise. Never run a subject-mutating check against an uncommitted subject: on 2026-07-15, `scripts/test-ci-deterministic-gates.sh` regenerated `skills-codex/` from HEAD mid-validation and destroyed the uncommitted subject, forcing `NOT_PROVEN` (verdict `b6e759dd...cb6a`); only restoring the subject and revalidating in a fresh context produced the PASS (`e9b6cdb8...37b9`). If a mutating check is genuinely required by acceptance, run it against a disposable copy or a committed subject, never the judged working tree. ## Workflow 1. Recompute and compare `subject-manifest.v1` using `python3 skills/validate/scripts/validate.py manifest`. The helper uses only filesystem content; Git commit/tree IDs are optional metadata. Derive the manifest at the start of validation and re-derive it at the end; any mismatch between the two is subject mutation and returns `NOT_PROVEN`. 2. Confirm the intent-source digest has not changed since implementation. If the subject changed or complete changed-path coverage cannot be derived, return `NOT_PROVEN`. 3. Adjudicate the actual diff, not a declared path list: compare runtime-derived actual changed paths against the intent's scope classes. A proven out-of-scope path returns `FAIL`; incomplete scope evidence returns `NOT_PROVEN`. 4. Inspect the exact subject and factual evidence. Reported exit codes are claims, not evidence: re-execute the claimed proofs that bear on acceptance (see the freshness rules below for when a digest-bound receipt suffices). Judge every acceptance criterion and record criterion-level results, findings, evidence references, `checked`, and `not_checked`. 5. Choose exactly one semantic result: `PASS`, `FAIL`, or `NOT_PROVEN`. Return it with criterion results, findings, evidence references, `checked`, `not_checked`, the acceptance and subject identities, distinct author and validator context IDs, and the freshness attestation. PASS requires distinct identities, explicit freshness, nonempty checked scope, top-level evidence, and evidence for every criterion. 6. Only when the caller requests machine-readable evidence or a declared downstream consumer requires it, persist canonical `verdict.v2` with `store-verdict --draft <draft.json> --intent-source <resolved-intent> --subject-manifest <manifest.json> --author-context-id <id> --validator-context-id <id> --freshness-source <runtime|caller> --freshness-attester-id <id> --scope-result <PASS|FAIL|NOT_PROVEN>`. The helper snapshots the exact resolved intent under `<workspace>/.agents/ao/intents/sha256/<digest>.intent`, then computes and injects intent and subject digests plus author, validator, and freshness facts. Identity and changed-path facts come from runtime-derived inputs and receipts, not model transcription. Storage defaults to `<workspace>/.agents/ao/verdicts/sha256/<digest>.json`; callers may provide `verdict_dir`. 7. Return the semantic result and, when persisted, the artifact path and digest. Stop. The digest is SHA-256 over canonical JSON with `artifact_digest` omitted. Writes use a same-directory temporary file, flush, fsync, and atomic rename. Identical existing content is idempotent success; conflicting content is an integrity failure represented by `NOT_PROVEN`. ## Freshness without duplication Fresh validation means independent judgment over the exact subject. It does not require mechanically replaying every author command. Verify intent identity, scope, evidence digests, and every acceptance criterion; independently rerun the risk-critical, uncertain, or insufficiently evidenced checks. A digest-bound deterministic receipt may prove routine facts. Replay an expensive full suite only when acceptance requires that result or the supplied receipt cannot establish it. ## Boundary Validate emits no WARN, confidence, disposition, briefing learning, owner, next action, repair, retry, replan, helper, escalation, tracker, Git, release, closure, or delivery state. Generic provenance may record a verdict later, but ledger availability cannot change its validity.
Comments (0)
Sign in to join the conversation.
Reviews (0)
No reviews yet.
No comments yet.