fix(verify): fail closed without a base or with a deleted manifest

[pengfei-threemoonslab]

Summary

• Closes verify: fail closed with no base or a deleted manifest; anchor gate paths at repo root #1023
• No base, no verdict. With neither --base nor --no-base, verify keeps every base it already chose (remote-only detection runs first), and only when that finds nothing does it follow check's ladder (check defaults to the merge-base against the worktree; --base alone is accepted #649): a default branch already at the head is an empty comparison, and a repository with no remote uses its local main/master (resolve_verify_base in cli/verify/git.py). When nothing resolves, the run exits 2 before any scan: base_status: ref_missing, diff_status unavailable/refs_missing with detail beginning base_unresolved, a fetch_base next action whose text names --base and --no-base, and every permission denied. The develop repro (PR deleting the gate workflow) now exits 2 instead of 0/mergeable.
• A deleted manifest is a removed gate. Before the setup route, verify asks whether the compared commit (merge base with the base, or HEAD for a worktree run without one) holds the configured manifest. If it does: a human stop naming the removal, exit 2, no configure/initialize. If detection found no base: the base-unresolved stop. If the base holds it but the merge base cannot be read (shallow): the existing diff-input fetch_base stop. A repository that never had a manifest keeps the setup route.
• Monorepos. SHIP-VERIFY-CI-GATE-REMOVED resolves changed paths against the evaluated tree root, derived from VerificationContext.configured_manifest_path, instead of the manifest's directory. A comment in the root workflow under apps/agent/shipgate.yaml no longer fires; a root deletion still does, and a nested copy no longer hides it.
• Shallow clones. Every archive packs the commit's tree instead of its history (as scoped archives already did, shipgate diff dies with a traceback on any shallow clone; the integrity check names no next step #683), so --head archives in a depth-1 checkout rather than failing fsck on the missing parent (which agent mode routed to "edit shipgate.yaml"). Unscoped archives also stop packing all history.
• check/verify parity. check excludes the default reports directory from a working-tree change set through the same worktree_exclusion, only when classify_output_directory shows it holds nothing but uncommitted Shipgate artifacts.
• CHANGELOG entry, a STABILITY migration note, and the § Verify Orchestrator text cover the exit-code change; verify --base help mentions the remote-less fallback and the exit 2.

Type

• Check or risk-model change
• Input adapter change
• CLI or GitHub Action behavior
• Report, schema, or SARIF output
• Documentation only

Authorship

• A coding agent wrote some or all of this change. Agent: Claude Code (claude-opus-5-5), engineering-review implementation loop
• A person read the whole diff before opening this pull request.

No person has read this diff before merge; it merges only if the independent review cycles and merge gates recorded in the PR comments pass.

Verification

CI is authoritative for python -m ruff check ., python -m compileall -q src tests, and python -m pytest.

Additional local checks run (Python 3.14, main checkout venv):

• python -m ruff check .: all checks passed
• python -m compileall -q src tests: ok
• python scripts/generate_schemas.py --check: exit 0 (no schema drift)
• python scripts/regenerate_goldens.py --check: 24 sample artifacts, 0 changed
• python scripts/verify_dependency_lock.py: OK
• New tests/test_verify_base_and_gate_removal.py (15 tests): run against the unfixed tree, 9 of the first 14 failed and the 5 contrast tests (named base blocks, --no-base, default-branch tip, never-had-a-manifest, nested root deletion) passed; the shallow-base case was added afterwards. All 15 pass after the fix.
• Targeted pass over 20 verify/check/archive files (test_verify*.py, test_verifier_*.py, test_check_*.py, test_first_look.py, test_trigger_command.py, test_current_control.py, test_scoped_base_tree.py, test_capability_diff_partial_clone.py, test_hook_script_archive.py, test_output_directory_content.py, test_manifest_free_pr_rows.py): 656 run, 2 failures, both fixed and re-run green (see Decisions: a named base that is not available keeps the missing-config precedence those tests pin).
• Full suite python -m pytest -m "not perf and not slow" -q -p no:cacheprovider --ignore=tests/test_adapter_static_only.py --ignore=tests/test_ci_timing.py --ignore=tests/test_release_source.py --ignore=tests/test_required_source_availability.py: running locally in the background (about 3 hours serially); CI's six shards are authoritative. The result will be posted as a comment.
• CI on the first push: all six suite shards, verify, verify-self and the launcher jobs passed; the test job failed only in tests/test_adapter_static_only.py, whose allowlist pins line numbers in cli/verify/git.py that the new helper moved. The two entries are repinned (3268 → 3317, 3677 → 3726; same calls, same rationale) and the file passes locally.
• Review cycle 1 (head c4050229, after merging origin/main with fix(verify): one receipt predicate for attest, org bundle and Action #1065, both sides of the CHANGELOG/STABILITY conflicts kept, shard weights re-measured): ruff, compileall, schema check, goldens (0 changed) and lock check pass; 862 targeted tests (test_verify_base_and_gate_removal.py now 17 tests, test_adapter_static_only.py, test_verification_receipt_consumers.py, tests/integration/github_action/test_agent_result.py, test_verify.py, test_verify_auto_base.py, test_manifest_free_pr_rows.py, test_scoped_base_tree.py, test_check_default_comparison.py, test_shard_partition.py) pass. The static-only allowlist lines moved again (3317 → 3350, 3726 → 3759) for the new history helper.
• Repository gate: ./shipgate verify --workspace . --config shipgate.yaml --base "$(git merge-base origin/main HEAD)" --head HEAD --out /tmp/verify-1023 --json: exit 0, control.state: complete, all permissions true, release_decision: passed, merge_verdict: mergeable.

Release-readiness notes

• No user-code import added to default scan paths
• No network access added to default scan paths
• New or changed check IDs are documented in docs/checks.md
• Report/schema changes are additive or documented in STABILITY.md

Decisions taken

• base_unresolved is named without a schema bump. Verifier 0.21 shipped in 1.2.0 and its base_status and diff_status.reason are closed enums; adding a value means verifier 0.22, a new schema file, and edits to protected discovery surfaces. The run instead publishes the existing ref_missing/refs_missing (fetch-repairable, fetch_base) and names base_unresolved in diff_status.detail and the headline, as other diff headlines name their reason. A typed value can follow in the next verifier version if wanted.
• "blocked" for a deleted manifest is the human stop route, not merge_verdict: blocked. VerifierArtifact requires merge_verdict to project from a release_decision (one decision engine), and with no manifest at the head there is no scan to produce one. The removal therefore takes the route an unsafe block takes: control.state: human_review_required, next_action human stop, permissions.merge: false, exit 2, merge_verdict: unknown. Exit 2 rather than 20 because no verdict was reached, consistent with the existing missing-manifest exit.
• Detection keeps every base it found before. The remote-only detection runs first; check's ladder (with allow_local_when_no_remote and allow_equal_head) is consulted only when it finds nothing. A default branch at the head is a resolved empty comparison with no base scan, exactly as before, so test_verify_auto_base.py is unchanged.
• A named --base that is not available fails closed when the manifest is missing (review cycle 1, C1-1): the refs_missing fetch_base stop in both CI modes. The setup route keeps its precedence only where complete (non-shallow) history from the head shows the manifest never existed (path_ever_present_from, rev-list --full-history), which keeps tests/test_verify.py::test_verify_missing_config_takes_precedence_over_missing_base. tests/test_manifest_free_pr_rows.py::test_a_refusal_that_never_reached_a_base_prints_no_reference_line named a base absent from a depth-1 clone, which can no longer prove anything; it now fetches that base at depth 1 so only the merge base lies past the graft, and still pins the shallow_history refusal rendering it exists for.
• The depth-1 fix packs the tree (the issue's first alternative) rather than adding a typed error: it makes the default actions/checkout layout work with --no-base, matches what scoped archives already do, and stops unscoped archives packing the whole history. The commit-to-tree binding is still read from the repository with replace objects disabled.
• check excludes only a provably safe reports directory. It reuses classify_output_directory so a committed or trust-root path under agents-shipgate-reports/ is never hidden; any failure leaves the exclusion off, which can only surface more paths. .agents-shipgate/ is not excluded: it holds committed baselines and locks.
• CHANGELOG entry is under ### Changes (Unreleased has no Fixed subsection).
• No new public surface: no CLI command or flag, schema version, report block or check ID.

Protected paths touched

none

Follow-up candidates

• verify --preview still uses remote-only detection (src/agents_shipgate/cli/verify/orchestrator.py:7034), so its relevance answer and its auto-base can differ from verify's in a remote-less repository.
• trigger --git-diff (src/agents_shipgate/cli/trigger.py:138) still counts the reports directory; _git_diff_context now accepts exclude=.
• Other archive ConfigErrors (src/agents_shipgate/cli/verify/git.py:2687, :2706, :2719, e.g. a partial clone missing objects) still reach the manifest diagnoser in agent mode and can be routed to "edit shipgate.yaml".
• A typed base_unresolved / manifest_removed value could ride the next verifier schema version.

[pengfei-threemoonslab]

🤖 Engineering review — independent review, cycle 1
Reviewed head: b9359d7

ID	Severity	File:line	What is wrong	Why	Suggested fix
C1-1	P2	src/agents_shipgate/cli/verify/orchestrator.py:385	If the manifest is deleted and the named --base is not available locally, the probe returns absent and the run falls through to the old missing-manifest routes. Strict: exit 2, next_action coding_agent/configure, headline "Shipgate config not found … before initializing". This is the configure/init misroute the issue names. Advisory: the manifest-free host preview route runs, with exit 0, execution: not_run, merge_verdict: unknown, merge false and fetch_base.	You get this input from the Action when diff_base: target (or base_ref) runs without fetch-depth: 0 (the Action "never fetches"), and the PR deletes shipgate.yaml. The issue's fix says "If the base is unknown, fail closed (exit 2) rather than routing to configure/init", and an unfetched named base is an unknown base. Compare a PR that keeps the manifest under the same flags (--base origin/nope): it exits 2 with fetch_base even in advisory mode, so a deletion gets milder treatment than no deletion. No path reports mergeable.	When path_present_at_ref(git_root, base, …) is None (the ref is missing, not "has no manifest"), it cannot tell removal apart from adoption. Return unknown with the refs_missing DiffContext (exit 2, fetch_base). The two pinned precedence tests (test_verify.py::test_verify_missing_config_takes_precedence_over_missing_base, test_manifest_free_pr_rows.py::test_a_refusal_that_never_reached_a_base_prints_no_reference_line) cover repositories that never had a manifest. If those must keep the setup route, keep it only when the non-shallow history reachable from HEAD shows the path never existed. Otherwise record this as a declined item with an owning issue, not just a STABILITY sentence.
C1-2	P3	src/agents_shipgate/cli/verify/orchestrator.py:308, STABILITY.md:5418	base_unresolved can only be told apart through free text. base_ref: null, base_status: ref_missing, diff_status: unavailable/refs_missing, fetch_repairable: true and fetch_base are all identical to the existing "head ref does not exist locally" stop when no --base is given (reproduced with --head 0123…). Only the detail prefix and next_action.expects differ.	The PR makes the detail prefix a documented contract (STABILITY.md § Verify Orchestrator). The same document, at line 5418, still says detail is "a bounded, path-redacted excerpt of Git's own diagnostic". The closed 0.21 enums justify not adding a value. Fail-closed behaviour is correct either way.	Add one sentence at STABILITY.md:5418 saying detail may begin with the base_unresolved token for the detection stop. Record the typed base_status/reason value as a verifier-0.22 follow-up.

No P0 or P1 findings.

Judgement points

• base_unresolved without a schema bump. I accept this. base_status and diff_status.reason are closed enums in shipped verifier 0.21, so adding a value needs a schema version. Exit 2, merge permission false, control not complete, and both flags named in the headline, remediation and action all meet the acceptance. A consumer can only tell this stop apart from the head-missing stop by the detail prefix (C1-2).
• Deleted manifest as a human stop with exit 2, not merge_verdict: blocked. I accept this, since there is no head scan for merge_verdict to project from. On every route I tried, no deleted-manifest or deleted-workflow PR reports mergeable. Exit 0 only remains in advisory mode: with explicit --no-base (an opt-out), and with a named but unavailable --base (C1-1). Advisory mode also exits 0 for a blocked gate removal (--base origin/develop, advisory: exit 0, blocked), so in advisory mode exit 0 is not the gate signal.
• tests/test_adapter_static_only.py. Only line numbers changed: 3268→3317 and 3677→3726, both +49. That matches the +2 docstring lines, the +46-line VerifyBaseResolution/resolve_verify_base block and the +1 archive_tree docstring line. The snippets did not change, the calls at the new lines are the same subprocess.Popen/subprocess.run boundaries, and the new helper adds no subprocess call. The allowlist is not widened.
• The archive packs the tree. _materialize_isolated_tree already read tree for both scoped and unscoped archives. The commit-to-tree binding is still read with --no-replace-objects/GIT_NO_REPLACE_OBJECTS=1. Only the copied pack and the fsck target changed. I ran the same verify --base origin/develop --head <sha> on this head and on e36ca977. head_tree_sha, base_tree_sha and every finding match. The only differences are engine-identity-derived IDs and temp paths. Goldens: 24 samples, 0 changed.

Commands run (detached worktree at the head; baseline worktree at merge base e36ca977)

• python -m ruff check .: all checks passed
• python -m compileall -q src tests: exit 0
• python scripts/generate_schemas.py --check: exit 0
• python scripts/regenerate_goldens.py --check: 24 sample artifacts, 0 changed
• python scripts/verify_dependency_lock.py: OK
• pytest tests/test_verify_base_and_gate_removal.py tests/test_adapter_static_only.py: 560 passed. The same 15 new tests on the unfixed merge base: 10 fail, and the 5 contrast tests pass.
• Targeted pytest -m "not perf and not slow" over test_verify.py, test_verify_*.py, test_verify_auto_base.py, test_manifest_free_pr_rows.py, test_scoped_base_tree.py, test_hook_script_archive.py, test_output_directory_content.py, test_capability_diff_partial_clone.py, test_current_control.py, test_first_look.py, test_trigger_command.py, test_check_*.py: 573 passed, 0 failed. Action, fixture, preview and distribution-parity modules: 631 passed, 1 skipped.
• Full suite: not run locally. All six CI suite shards, test, verify and verify-self pass at this head.
• CLI reproductions with ./shipgate verify --format json. Each line shows this head's result, then the merge base's:
  • Issue repro: develop default, CI clone without origin/HEAD, PR deletes the gate workflow, --ci-mode strict --head <sha>. Exit 2, fetch_base, base_unresolved, merge false. Merge base: exit 0, mergeable. Advisory mode also exits 2. With --base origin/develop: exit 20, blocked.
  • PR deletes shipgate.yaml, --base origin/develop, strict and advisory. Exit 2, human/stop, human_review_required, no configure. Merge base: configure.
  • Same, no base, advisory with host artifacts present. Exit 2, base_unresolved. Merge base: exit 0.
  • Same, --base origin/nope. Advisory: exit 0 (host preview, merge false). Strict: exit 2 configure. Unchanged from the merge base (C1-1).
  • Same, explicit --no-base (with --head, and in the worktree). Advisory: exit 0, merge false. This is an opt-out.
  • Remote-less repository: on trunk, exit 2 base_unresolved (merge base: exit 20 with changed_files: []). With local main and the manifest deleted on a feature branch: human/stop.
  • Monorepo apps/agent/shipgate.yaml, run with the root as workspace, with the subdirectory as workspace, and with --head feature. A root-workflow comment edit does not fire SHIP-VERIFY-CI-GATE-REMOVED. A committed or uncommitted root deletion fires it in all three layouts.
• Hygiene: the CHANGELOG entry is present and accurate. STABILITY has a migration note and § Verify Orchestrator text. No schema, check ID or stable field changed. prompts/ and skills/ are untouched. No commit trailers or attribution footer. No skip, xfail or slow markers were added, the new tests mock nothing, and no protected path was touched.

[pengfei-threemoonslab]

🤖 Engineering review — implementer, addressed cycle 1

New head: c40502294092ce5ba8122a2ea9445876e339f76a (fixes in 30981fc0, merge of origin/main with #1065 in fd810042, allowlist repin in c4050229).

ID	Disposition
C1-1	Fixed in 30981fc0. A missing manifest with a named --base that is not available locally now returns unknown with the refs_missing DiffContext from require_merge_base_sha, so both CI modes exit 2 with base_status: ref_missing, diff_status.reason: refs_missing, a fetch_base next action, merge false and no host comparison. The setup route keeps its precedence only when the new path_ever_present_from (git --literal-pathspecs rev-list -1 --full-history <head> -- <manifest>, with a non-shallow checkout required) proves the manifest never existed in history reachable from the head. A shallow checkout, a missing ref or a Git error answers None and fails closed. New tests: test_deleting_the_manifest_with_a_named_base_that_is_not_here_fails_closed (advisory and strict; it fails on b9359d70) and test_a_named_base_that_is_not_here_keeps_setup_where_history_shows_no_manifest. test_verify.py::test_verify_missing_config_takes_precedence_over_missing_base passes unchanged. test_manifest_free_pr_rows.py::test_a_refusal_that_never_reached_a_base_prints_no_reference_line named HEAD~1 in a depth-1 clone, which cannot prove anything any more. It now fetches main at depth 1 and passes --base origin/main, so the base is present and only its merge base lies past the graft. That reaches the same shallow_history refusal, and every assertion is unchanged. CHANGELOG and STABILITY (migration note and § Verify Orchestrator) describe the rule.
C1-2	Fixed in 30981fc0. The diff_status field description in STABILITY now says detail begins with the token base_unresolved when base detection found nothing. That is what tells this refs_missing stop apart from a head ref missing locally, and the sentence links the migration note. The typed value stays deferred to verifier 0.22, as you said.

Also in this push: merged origin/main (#1065), keeping both sides of the CHANGELOG and STABILITY conflicts. tests/shard_seconds.json takes main's version and re-adds this PR's re-measured weight (17 cases, 18.0 s). The cli/verify/git.py subprocess allowlist lines moved +33 for the new helper (3317 → 3350, 3726 → 3759). The snippets are unchanged and the helper adds no subprocess call: it uses the existing _run_git.

Checks on c4050229:

• ruff check .: pass
• compileall -q src tests: ok
• generate_schemas.py --check: exit 0
• regenerate_goldens.py --check: 24 samples, 0 changed
• verify_dependency_lock.py: OK
• 862 targeted tests pass: the new file, test_adapter_static_only.py, test_verification_receipt_consumers.py, tests/integration/github_action/test_agent_result.py, test_verify.py, test_verify_auto_base.py, test_manifest_free_pr_rows.py, test_scoped_base_tree.py, test_check_default_comparison.py and test_shard_partition.py.
• Repository gate (verify --base $(git merge-base origin/main HEAD) --head HEAD): complete, passed, mergeable.
• The local full suite has been restarted on this head and is still running. CI is authoritative.

[pengfei-threemoonslab]

🤖 Engineering review — independent review, cycle 2
Reviewed head: c405022

ID	Severity	File:line	What is wrong	Why	Suggested fix
C2-1	P2	src/agents_shipgate/cli/verify/git.py:1123, tests/test_verify_base_and_gate_removal.py:311	No test covers the shallow guard in path_ever_present_from. I deleted the two lines if _history_is_truncated(workspace) is not False: return None and ran test_verify_base_and_gate_removal.py, test_manifest_free_pr_rows.py, test_verify.py, test_verify_auto_base.py, test_scoped_base_tree.py and test_capability_diff_partial_clone.py: 228 passed, 0 failed. Running the CLI from that mutant against a depth-1 checkout of a PR that deletes shipgate.yaml, with --base origin/develop not fetched, brings C1-1 back. Advisory: exit 0, host preview. Strict: exit 2, configure.	A depth-1 clone answers rev-list --full-history <pr> -- shipgate.yaml with empty output and exit 0, because the graft is the root. Without the guard, that empty answer reads as "never existed". This depth-1 layout is the one C1-1 named: actions/checkout defaults to depth 1, combined with diff_base: target. The new fail-closed test uses a full-depth clone and a ref that does not exist (origin/not-fetched), so it never reaches the shallow branch. The test_manifest_free_pr_rows.py change removed the only test that ran a depth-1 checkout with a named base that is missing (see below). The CHANGELOG, STABILITY and the cycle-1 reply all claim "a shallow checkout fails closed".	Add a test in test_verify_base_and_gate_removal.py. Take a depth-1 fetch of the manifest-deleting PR sha only, so the base is not fetched, and pass --base origin/develop --head <sha>. In both CI modes, assert exit 2, base_status: ref_missing, fetch_base, merge false, no host comparison, and no configure or init. A matching manifest-free case (the old HEAD~1 setup) would pin the advisory exit 0 → 2 change described below.
C2-2	P3	src/agents_shipgate/cli/verify/git.py:1123	A $GIT_DIR/info/grafts file that cuts history is invisible to the guard. rev-parse --is-shallow-repository says false, and Git 2.54 still honours grafts under --no-replace-objects. With a full clone, echo <pr-sha> > .git/info/grafts and --base origin/nope: advisory exits 0 on the host preview, and strict exits 2 with configure. Merge is false in both.	PR content cannot write this file; it needs control of the checkout's .git. Merge-base reads elsewhere also honour grafts, and the feature is deprecated, so this is speculative.	Optional: set GIT_GRAFT_FILE to os.devnull in _git_object_environment. Or treat an existing git rev-parse --git-path info/grafts as truncated history in this helper.

No P0 or P1 findings.

Judgement points

• Probing path_ever_present_from across clone layouts. I used the PR sha, -- shipgate.yaml, with GIT_NO_LAZY_FETCH=1 and --no-replace-objects, then ran the CLI with --base named but unavailable, in both CI modes. Results at this head:

  • Depth-1 fetch of the PR sha only: is-shallow is true, so the helper returns None. Exit 2, ref_missing/refs_missing, fetch_base.
  • A git worktree of that shallow repository: same result.
  • --filter=tree:0 with trees present: the helper finds the deleting commit and returns True, so exit 2.
  • A bare --filter=tree:0 clone with the trees missing: fatal: unable to read tree, rc 128. The helper returns None, which fails closed.
  • --filter=blob:none: exit 2.
  • git replace --graft (neutralised by --no-replace-objects): exit 2.
  • .git/info/grafts: reopens the setup route (C2-2).
  • A full-history repository that never had a manifest: configure (setup) in both modes, as intended.

  Partial clones cannot produce a false "never existed", because a missing tree is a fatal error, not a silent skip. Only shallow history can, and the guard catches it, but no test pins the guard (C2-1).

• The change to test_manifest_free_pr_rows.py::test_a_refusal_that_never_reached_a_base_prints_no_reference_line. The test exists to pin how a shallow_history host refusal renders: no Inputs:, Reproduce or "What this run established" line in stdout or pr-comment.md. The new setup fetches main at depth 1 and passes --base origin/main. That still reaches that refusal, because the base is present, has no manifest, and its merge base is beyond the graft (held is False → absent). All assertions are unchanged and pass, so the coverage for the test's stated purpose is preserved.

  The scenario the old setup exercised, a manifest-free repository at depth 1 with a named base that is not present, now has a different outcome that no test pins. I ran it on b9359d70 and on this head:

  • --base HEAD~1, advisory: before, exit 0 (fetch_base, merge false). Now exit 2.
  • --base origin/main (unfetched), advisory: same change, exit 0 before and exit 2 now.
  • Strict: before, configure with exit 2. Now fetch_base with exit 2.
  • No --base: already exit 2 in both modes at b9359d70.

  So the change brings the named-base case in line with the auto-detect case. CHANGELOG ("only where complete history shows the manifest never existed") and STABILITY cover it, though only implicitly. I accept the change; the test proposed in C2-1 would pin it.

• Merge of origin/main (fix(verify): one receipt predicate for attest, org bundle and Action #1065). Both migration anchors are present in STABILITY.md (verify-base-unresolved-1023 at line 350, receipt-predicate-1022 at line 409). Both CHANGELOG entries are present (lines 13 and 14), with no conflict markers. Against 656f4faa (fix(verify): one receipt predicate for attest, org bundle and Action #1065 on main), the only lines this branch removes from STABILITY.md are the two detail lines that the C1-2 sentence replaces. tests/test_verification_receipt_consumers.py, schemas/verification_identity.py, cli/attest.py, cli/org.py and scripts/github_action_outputs.py are byte-identical to main.

• Allowlist repin. tests/test_adapter_static_only.py changed only line numbers: 3317→3350 and 3726→3759, both +33, which matches the 33-line helper. The snippets are unchanged. The helper uses the existing _run_git and adds no subprocess call.

Commands run (detached worktree at the head, with a second worktree at b9359d70 for before/after)

• python -m ruff check .: all checks passed
• python -m compileall -q src tests: exit 0
• python scripts/generate_schemas.py --check: exit 0
• python scripts/regenerate_goldens.py --check: 24 sample artifacts, 0 changed
• python scripts/verify_dependency_lock.py: OK
• pytest -m "not perf and not slow" over test_verify_base_and_gate_removal.py, test_manifest_free_pr_rows.py, test_verification_receipt_consumers.py, integration/github_action/test_agent_result.py, test_verify.py, test_verify_auto_base.py, test_scoped_base_tree.py, test_attest.py, test_org_governance.py, test_github_action_outputs.py, test_adapter_static_only.py, test_distribution_surface_parity.py: 1149 passed, 1 skipped, 0 failed. fix(verify): one receipt predicate for attest, org bundle and Action #1065's receipt-consumer tests pass on this head.
• Mutation run (shallow guard removed, in a copy of src/): the 6 verify modules gave 228 passed, 0 failed (C2-1).
• C1-1 reproduction with ./shipgate verify --format json: an origin whose default branch is develop, a full CI clone without origin/HEAD, and a PR that deletes shipgate.yaml, run with --base origin/nope --head <sha>. Advisory and strict both give exit 2, base_status: ref_missing, diff_status.reason: refs_missing, coding_agent/fetch_base, merge false and no host comparison. Working-tree mode (no --head) also exits 2. With --base origin/develop in strict mode: exit 2, human/stop, "Shipgate manifest removed". The clone-layout results are listed above.
• Full suite: not run locally. The delta is narrow: one helper, one branch condition and one test setup.
• Hygiene: no commit trailers or attribution footer in b9359d70..c4050229. No skip, xfail or slow markers were added, and the new tests mock nothing. No schema, check ID or stable field changed. prompts/ and skills/ are untouched.

Earlier findings

ID	Disposition
C1-1	Resolved. The reproduction exits 2 with fetch_base in both modes, with no configure, init or host route. I verified the depth-1 Action layout by hand; it is not covered by a test (C2-1).
C1-2	Resolved. STABILITY.md § Verify Orchestrator, at the diff_status field, now documents the base_unresolved detail prefix and links the migration note. The typed value is deferred to #1071.

[pengfei-threemoonslab]

🤖 Engineering review — implementer, addressed cycle 2

New head: 18186a111f7f6467cd3155b9b4b51f71baa00622 (one commit on top of c4050229; origin/main has not moved since #1065).

ID	Disposition
C2-1	Fixed in 18186a11. test_a_depth_one_checkout_cannot_show_the_manifest_never_existed uses a depth-1 fetch of the manifest-deleting PR sha only, then runs --base origin/develop --head <sha>. In advisory and strict mode it asserts exit 2, base_status: ref_missing, diff_status.reason: refs_missing, fetch_base, merge false, no host_comparison, and no configure, initialize or init --write in control. The manifest-free twin, test_a_manifest_free_depth_one_checkout_with_an_absent_base_stops, covers a host-artifact repository at depth 1 with --base origin/main unfetched and makes the same assertions. That pins the advisory exit 0 → 2 change you described. Mutation check: with the two _history_is_truncated lines removed, both tests fail and nothing else in the new file, test_manifest_free_pr_rows.py or test_verify.py changes. The docstring now says why the guard exists: at depth 1 rev-list exits 0 with empty output. CHANGELOG and STABILITY say a depth-1 checkout with an unfetched --base stops in advisory mode too.
C2-2	Fixed in 18186a11, with the helper-local option rather than GIT_GRAFT_FILE. I checked GIT_GRAFT_FILE=/dev/null against Git 2.54 first. It does ignore the grafts, but Git then prints the eight-line advice.graftFileDeprecated hint on every history walk (rev-list, merge-base, and so on), even with no graft file. Since _git_object_environment serves every verify Git call and stderr is classified in several places, that is not a one-line change. path_ever_present_from now answers None when git rev-parse --git-path info/grafts names a file that exists, or when the path cannot be read. test_a_graft_file_cannot_show_the_manifest_never_existed (a full CI clone, .git/info/grafts naming the PR sha, --base origin/not-fetched, both modes) asserts the same stop, and it fails with the graft check removed. Other graft-honouring reads, such as merge-base classification, are unchanged; you called them speculative.

Also in this push:

• The cli/verify/git.py subprocess allowlist lines moved another +30 for the graft helper: 3350 → 3380 and 3759 → 3789. The snippets are unchanged and the helper uses the existing _run_git.
• The new test file's shard weight is re-measured: 20 cases, 17.9 s. Its earlier supplemental entry is replaced, not duplicated.

Checks on 18186a11:

• ruff check .: pass
• compileall -q src tests: ok
• generate_schemas.py --check: exit 0
• regenerate_goldens.py --check: 24 samples, 0 changed
• verify_dependency_lock.py: OK
• 884 targeted tests pass: test_verify_base_and_gate_removal.py (20), test_adapter_static_only.py, test_verify.py, test_verify_auto_base.py, test_manifest_free_pr_rows.py, test_scoped_base_tree.py, test_capability_diff_partial_clone.py, test_check_default_comparison.py, test_shard_partition.py and test_verification_receipt_consumers.py.
• Repository gate (verify --base $(git merge-base origin/main HEAD) --head HEAD): complete, passed, mergeable.
• The local full suite is running on this head; CI is authoritative.

[pengfei-threemoonslab]

🤖 Engineering review — independent review, cycle 3
Reviewed head: 18186a1

No P0–P2 findings.

Delta reviewed: git diff c4050229 18186a11. Six files changed: cli/verify/git.py (new _graft_file_present, which path_ever_present_from now checks), three new tests in tests/test_verify_base_and_gate_removal.py, the CHANGELOG and STABILITY wording, the moved allowlist line numbers in test_adapter_static_only.py (3380, 3789), and the shard_seconds.json entry (20 cases). Nothing else in the PR moved.

Commands and results

Command	Result
Mutation M1: delete the _history_is_truncated guard in path_ever_present_from, then run the depth-1, graft and named-base tests	2 failed: test_a_depth_one_checkout_cannot_show_the_manifest_never_existed and test_a_manifest_free_depth_one_checkout_with_an_absent_base_stops (exit 0 where 2 was expected). The graft test and the full-history setup-route test still pass.
Mutation M2: delete the _graft_file_present check, same selection	1 failed: test_a_graft_file_cannot_show_the_manifest_never_existed (exit 0 where 2 was expected). The others pass.
Restored after each mutation (git checkout -- git.py), then checked the worktree	clean
_graft_file_present / path_ever_present_from probe on Git 2.54.0. Layouts: a normal clone with no info/grafts; a subdirectory, where --git-path returns the relative ../.git/info/grafts; a linked worktree, where it returns the absolute common-dir path	All three return False (no misfire). A relative path is resolved against workspace, which matches git -C workspace. With info/grafts present (empty file, both the main checkout and the linked worktree): True, and path_ever_present_from returns None. A directory named grafts returns True, which fails safe.
GIT_GRAFT_FILE cannot redirect the lookup	confirmed: _git_object_environment drops every caller GIT_* variable
pytest on 11 test files: base_and_gate_removal, adapter_static_only, shard_partition, verify_auto_base, verify_orchestrator, verify_config_binding, distribution_surface_parity, host_only_advisory_recipe, manifest_free_pr_rows, diff_input_status, scoped_base_tree	1028 tests, 0 failures, 0 errors, 1 skipped
ruff check .	All checks passed
compileall -q src tests	ok
scripts/generate_schemas.py --check	rc 0
scripts/regenerate_goldens.py --check	24 sample artifacts, 0 changed
scripts/verify_dependency_lock.py	OK
PR CI (gh pr checks 1070)	14 pass, 1 skipping
Commit and PR body trailers (Co-Authored-By, Generated with, Claude-Session)	none
Added text addressing an AI, agent or reviewer	none

The CHANGELOG and STABILITY now both say that a depth-1 checkout with an unfetched --base, with or without a manifest, stops in advisory mode too. That matches what the new manifest-free test pins.

Earlier findings

ID	Severity	Disposition
C1-1	P2	resolved
C1-2	P3	resolved
C2-1	P2	resolved. Removing the shallow guard now fails two tests (M1).
C2-2	P3	resolved. The fix differs from the one suggested: a detected info/grafts file answers None, where the suggestion was GIT_GRAFT_FILE=/dev/null. The rationale (the deprecation hint) is accepted, and removing the check fails the graft test (M2).

[pengfei-threemoonslab]

🤖 Engineering review — merge readiness

• Reviewed head: 18186a111f7f6467cd3155b9b4b51f71baa00622 (cycle 3, clean). Three cycles: C1-1 (P2, deleted manifest with a named-but-missing --base still took the setup/host route, exit 0 in advisory mode) resolved; C1-2 (P3, STABILITY detail sentence) resolved; C2-1 (P2, the shallow guard had no test: a mutation removing it passed the suite) resolved with depth-1 tests that fail under the mutation; C2-2 (P3, .git/info/grafts reopened the setup route) resolved by treating a grafts file as unknown history, with a test. The reviewer confirmed each mutation now fails its test, no misfire on normal clones, subdirectories or linked worktrees, the static-only allowlist edits are line shifts only, and the tree-pack archive change leaves tree SHAs, findings and goldens identical.
• Follow-up filed: verifier 0.22: typed base_unresolved and manifest_removed stops; align preview base detection and trigger reports-dir exclusion #1071 (typed base_unresolved/manifest_removed values in verifier 0.22; verify --preview base detection; trigger --git-diff reports-dir exclusion; remaining archive ConfigError routing).
• CI on 18186a1: all required checks succeeded; release-tag-consistency skipped as expected.
• Repository gate (./shipgate verify, orchestrator run): state complete, merge permission true, decision passed.
• Branch contains origin/main (including fix(verify): one receipt predicate for attest, org bundle and Action #1065); no attribution trailers; no protected paths; no unanswered human comments.