feat(audit): record explicit operator acceptance

contract_tests/test_claiming.py

@@ -73,6 +73,7 @@
     "github.pull-request.comments.read",
     "github.pull-request.reviews.read",
     "github.pull-request.checks.read",
+    "github.repository.required-checks.read",
     "github.pull-request.threads.read",
 )
 

contract_tests/test_host_boundary.py

@@ -22,6 +22,7 @@
     "github.pull-request.comments.read",
     "github.pull-request.reviews.read",
     "github.pull-request.checks.read",
+    "github.repository.required-checks.read",
     "github.pull-request.threads.read",
 ]
 HEAD_SHA = "a" * 40
@@ -149,12 +150,17 @@ def complete_observation() -> dict[str, object]:
                 "pull_request_number": 785,
                 "kind": "CHECK_RUN",
                 "name": "test",
+                "integration_id": 42,
                 "status": "COMPLETED",
                 "conclusion": "SUCCESS",
                 "candidate_sha": HEAD_SHA,
                 "details_url": "https://github.com/shaug/atelier/actions/runs/1",
             }
         ],
+        "required_checks": {
+            "configuration_read": True,
+            "contexts": [{"name": "test", "integration_id": 42}],
+        },
         "threads": [
             {
                 "id": "thread-1",
@@ -175,6 +181,7 @@ def complete_observation() -> dict[str, object]:
             "pull_request_comments": True,
             "reviews": True,
             "checks": True,
+            "required_checks": True,
             "threads": True,
         },
     }

contract_tests/test_mailbox.py

@@ -413,6 +413,18 @@ def add_work(
                 "policy_commit": SHA_A,
                 "candidate_revision": SHA_B,
                 "evidence": {name: "satisfied" for name in EVIDENCE},
+                "audit_evidence": {
+                    "schema": "atelier.audit-evidence/v1",
+                    "review": {
+                        "mechanism": "review-code-change",
+                        "verdict": "clean",
+                        "candidate_revision": SHA_B,
+                        "comparison_base_revision": SHA_A,
+                        "observed_at": TIMESTAMP,
+                        "findings": [],
+                    },
+                    "feedback_dispositions": [],
+                },
             }
         self.works[work_id] = work
         self.write_work(work_id)

contract_tests/test_planning.py

@@ -126,6 +126,7 @@ def observation(*, blocked: bool = False, with_pull_request: bool = False) -> di
         "pull_request_comments": [],
         "reviews": [],
         "checks": [],
+        "required_checks": {"configuration_read": True, "contexts": []},
         "threads": [],
         "completeness": {
             "issue": True,
@@ -135,6 +136,7 @@ def observation(*, blocked: bool = False, with_pull_request: bool = False) -> di
             "pull_request_comments": True,
             "reviews": True,
             "checks": True,
+            "required_checks": True,
             "threads": True,
         },
     }

docs/git-mailbox-contract.md

@@ -696,12 +696,23 @@ acceptance:
     required-validation-reported: satisfied
     independent-review-current: satisfied
     unresolved-feedback-zero: satisfied
+  audit_evidence:
+    schema: atelier.audit-evidence/v1
+    review:
+      mechanism: review-code-change
+      verdict: clean
+      candidate_revision: 2222222222222222222222222222222222222222
+      comparison_base_revision: 1111111111111111111111111111111111111111
+      observed_at: 2026-07-25T13:14:30Z
+      findings: []
+    feedback_dispositions: []
 ```
 
-Acceptance records what was verified at that commit. If evidence later becomes
-unavailable, stale, or contradictory, history is not rewritten. Audit reports
-the current promise as `unknown`, `stale`, or `violated` and cites the earlier
-acceptance.
+Acceptance records what was verified at that commit, including the exact structured
+review findings and explicit dispositions for top-level review and comment bodies.
+If evidence later becomes unavailable, stale, or contradictory, history is not
+rewritten. Audit reports the current promise as `unknown`, `stale`, or `violated`
+and cites the earlier acceptance.
 
 ## Minimal planner-worker interactions
 

docs/project-policy-contract.md

@@ -244,26 +244,36 @@ Each predicate has one authoritative source and exact evaluation.
 ### `pull-request-open`
 
 - **Source:** live GitHub pull-request state.
-- **Satisfied:** the PR is open.
-- **Violated:** the PR is closed or merged.
+- **Satisfied:** the PR is open and not a draft.
+- **Violated:** the PR is closed, merged, or still a draft.
 - **Unknown:** GitHub cannot be read.
 - **Stale:** not applicable.
 
 ### `pull-request-mergeable`
 
-- **Source:** live GitHub mergeability.
-- **Satisfied:** GitHub reports the PR mergeable.
-- **Violated:** GitHub reports a conflict.
-- **Unknown:** mergeability is unavailable or still being calculated.
-- **Stale:** the PR head changed after the observation.
+- **Source:** live GitHub conflict and merge-state status.
+- **Satisfied:** GitHub reports the exact head mergeable and its current merge state
+  does not block readiness. `UNSTABLE` does not fail this predicate by itself;
+  configured required checks are evaluated separately.
+- **Violated:** GitHub reports a conflict or a policy-blocked merge state.
+- **Unknown:** conflict or merge-state status is unavailable or still being
+  calculated.
+- **Stale:** the PR head or exact comparison base changed after the observation.
 
 ### `required-checks-pass`
 
-- **Source:** live required GitHub checks for the exact PR head.
-- **Satisfied:** every required check completed successfully.
-- **Violated:** a required check completed unsuccessfully.
-- **Unknown:** required-check configuration or a required result cannot be read.
-- **Stale:** the PR head changed after the check result.
+- **Source:** live required-check configuration plus check results for the exact PR
+  head. Required identities are the exact context name plus optional GitHub App or
+  ruleset integration identity obtained from effective branch protection and
+  repository rulesets; observed results retain their check/status kind and
+  integration identity.
+- **Satisfied:** the required configuration was read and every named required
+  check from its configured provider completed successfully; an empty configured
+  set is satisfied.
+- **Violated:** a named required check completed unsuccessfully.
+- **Unknown:** required-check configuration or a named required result from its
+  configured provider cannot be read, is missing, or is ambiguous.
+- **Stale:** a named required result belongs to another PR head.
 
 ### `required-validation-reported`
 
@@ -276,20 +286,26 @@ Each predicate has one authoritative source and exact evaluation.
 
 ### `independent-review-current`
 
-- **Source:** a structured `review-code-change` result recorded in the receipt.
-- **Satisfied:** the result is clean and bound to the delivered SHA and
-  comparison base.
-- **Violated:** the result requires changes.
-- **Unknown:** the result is missing, malformed, or blocked.
-- **Stale:** the delivered head or effective comparison base changed.
+- **Source:** the delivered receipt plus normalized aggregate
+  `review-code-change` audit evidence with structured finding dispositions.
+- **Satisfied:** both records are clean, every finding has a non-unresolved
+  disposition, and both are bound to the delivered SHA and comparison base.
+- **Violated:** the result requires changes or retains an unresolved finding.
+- **Unknown:** either record is missing, malformed, blocked, or contradictory.
+- **Stale:** the delivered head, live PR base, or effective comparison base changed.
 
 ### `unresolved-feedback-zero`
 
-- **Source:** live GitHub review, comment, and thread state.
-- **Satisfied:** no unresolved material review, comment, or thread remains.
-- **Violated:** an unresolved material item remains.
-- **Unknown:** thread-aware feedback cannot be read.
-- **Stale:** the PR head changed after a previously clean observation.
+- **Source:** live GitHub review, comment, and thread state plus the normalized
+  audit-evidence dispositions bound to provider IDs and exact body digests.
+- **Satisfied:** every nonempty top-level review or comment body has an explicit
+  non-unresolved disposition and no unresolved material thread or receipt
+  obligation remains. This predicate is an unconditional v0 acceptance guard,
+  even when an older approval omitted it from `required_evidence`.
+- **Violated:** an item is explicitly unresolved, the live review decision requires
+  changes, or a material thread or receipt obligation remains unresolved.
+- **Unknown:** thread-aware feedback cannot be read or a live body lacks a disposition.
+- **Stale:** a disposition identifies missing feedback or an earlier body revision.
 
 ## Policy drift
 

scripts/validate_repository.py

@@ -82,11 +82,17 @@ def validate_skill(errors: list[str]) -> None:
         errors.append("Atelier skill frontmatter must include a description")
     if "[TODO:" in text:
         errors.append("Atelier skill contains a TODO placeholder")
-    if "Production `audit` behavior is not implemented yet" not in text:
-        errors.append("skill must state that audit is unavailable")
-    for relative in ("references/delegation.md", "scripts/delegation.py"):
+    if "Production `audit` behavior is implemented" not in text:
+        errors.append("skill must publish the audit boundary")
+    required = (
+        ("delegated work", "references/delegation.md"),
+        ("delegated work", "scripts/delegation.py"),
+        ("audit", "references/audit.md"),
+        ("audit", "scripts/audit.py"),
+    )
+    for boundary, relative in required:
         if not (path.parent / relative).is_file():
-            errors.append(f"delegated work boundary missing: {relative}")
+            errors.append(f"{boundary} boundary missing: {relative}")
 
 
 def validate_reset_boundary(errors: list[str]) -> None:

skills/atelier/SKILL.md

@@ -29,7 +29,7 @@ The preflight must prove:
   the host capability descriptor;
 - the installed and authorized `github@openai-curated` connector;
 - every required read-only issue, relationship, pull-request, comment, review,
-  check, and thread operation; and
+  check, effective required-check configuration, and thread operation; and
 - one complete observation conforming to
   `references/github-observation.schema.json` when live state is requested.
 
@@ -61,15 +61,19 @@ result. Launch one fresh worker through the host with the installed
 substitute CLI process, cache provider observations, or widen Atelier's v0
 authority ceiling.
 
-Production `audit` behavior is not implemented yet. When it is requested:
-
-1. Report that audit is unavailable in the reset scaffold.
-1. Link the owning issue: shaug/atelier#780.
-1. Make no mailbox, repository, ticket, pull-request, or acceptance mutation.
-
-An explicit host-readiness request may complete the read-only host preflight and
-return its exact compatibility or failure result. That does not make audit
-available.
+Production `audit` behavior is implemented for one delivered or accepted
+assignment. Before audit, read `references/audit.md`, repeat the fail-closed host
+preflight, obtain one fresh complete GitHub observation, normalize the exact
+aggregate review and live feedback dispositions as `atelier.audit-evidence/v1`,
+and use `scripts/audit.py audit`. Preserve every predicate verdict, structured
+finding, and visible feedback disposition in the report; do not infer success
+from unavailable, undispositioned, or stale state.
+
+Audit itself is read-only. Only after the operator explicitly confirms the
+report's exact acceptance fence may a strictly newer complete provider snapshot
+be supplied to `scripts/audit.py accept` for the one acceptance commit. Reusing
+report evidence is rejected. Acceptance never implies merge, deployment,
+native-ticket mutation, or issue closure.
 
 ## Mailbox boundary
 

skills/atelier/references/audit.md

@@ -0,0 +1,70 @@
+# Audit and acceptance
+
+Use this boundary only for one validated mailbox assignment whose current state
+is `delivered` or `accepted`.
+
+## Live audit
+
+1. Complete the host preflight in `host-boundary.md`.
+2. Read the canonical mailbox through `GitMailboxWriter.observe`; do not reuse a
+   prior clone or write an audit projection.
+3. Generate one complete `atelier.github-observation/v1` snapshot at the live
+   read boundary. It must include the exact issue relationships, pull request,
+   comments, reviews, checks, effective required-check configuration, and review
+   threads. A failed ruleset or branch-protection read is represented by
+   `required_checks.configuration_read: false`; it is never inferred from the
+   visible check list.
+4. Normalize the exact aggregate `review-code-change` result and every deliberate
+   disposition for a nonempty top-level review or pull-request comment into an
+   `atelier.audit-evidence/v1` JSON object. Each review finding records an ID,
+   summary, disposition, rationale, and optional follow-up. Each live review or
+   comment disposition records its kind, provider ID, exact `sha256:` body digest,
+   disposition, rationale, and optional follow-up. Omitted, duplicate, foreign,
+   or body-stale identities fail closed.
+5. Invoke `scripts/audit.py audit` with the mailbox remote and branch, work ID,
+   policy target, host target, observation path and lower bound, and
+   `--audit-evidence` path.
+6. Present the complete report. Every registry predicate is classified as
+   `satisfied`, `violated`, `unknown`, or `stale`; the report also preserves the
+   structured review findings, receipt obligations, and explicit disposition and
+   body of every review, pull-request comment, and thread.
+
+The report reconstructs the exact delivered receipt, remote candidate, pull
+request, head and comparison base, approved and current policy, native ticket
+observation, required validation, independent `review-code-change` result, and
+current feedback. An accepted assignment additionally derives the Git commit
+that introduced acceptance and the preceding delivery observation. Later drift
+changes the current audit verdict; it never rewrites historical acceptance.
+
+Audit is read-only. Missing host capability or project-policy identity reports
+`authority-unreconstructable`. Missing live evidence reports `unknown`.
+Contradictory state reports `violated`; changed candidate-bound state reports
+`stale`.
+
+## Explicit acceptance
+
+Acceptance is a second operation. Do not infer it from a clean audit.
+
+1. Show the operator the report and its `acceptance_fence`.
+2. Require explicit confirmation of that exact fence.
+3. Start a new provider read only after the confirmed report's `observed_at`.
+   Produce a second complete observation whose live-read lower bound and
+   `observed_at` are both strictly later. Reusing the report snapshot is rejected.
+4. Invoke `scripts/audit.py accept` with that second observation and lower bound,
+   the same normalized audit-evidence path, the complete fence, `--accepted-at`,
+   and `--confirm`.
+5. The transition rereads the canonical mailbox, host boundary, policy, ticket,
+   candidate, pull request base and head, required-check configuration and
+   results, reviews, comments, and threads. It requires the fence's mailbox,
+   receipt, candidate, and semantic evidence digest to remain exact while the
+   observation timestamp advances. Changed facts are rejected as stale or
+   violated; only the observation timestamp may differ.
+6. Only when every currently required predicate and the unconditional
+   no-unresolved-feedback guard are `satisfied` does one verified mailbox commit
+   move `delivered` to `accepted`, clear the claim, and bind the operator record
+   to the exact receipt, candidate, audit evidence, and satisfied verdicts. The
+   accepted timestamp is always checked against the current UTC clock, even when
+   no test clock is supplied.
+
+Acceptance does not merge a pull request, deploy, mutate or close a native
+ticket, delete a branch, or accept a parent initiative.