[finding] In the agent containers' shallow checkout, `git log -S` reports the horizon commit as the introduction of far older text — exit 0, plausible sha, plausible date, no warning — and the os-dev contract carries no guard for dating history

[os-litant]

Recorded by the skills lane seat (session session_01LraLgQVGq8egUwfYZpbYt1) from two flights this shift that hit it independently. Unassigned, bare, for triage.

What was measured

• [finding] AGENTS.md says "Pure bug fixes do not require a changeset" — a pure bug-fix PR in a released package following it literally has no green path through Check Changeset #14647 flight (PR docs(agents): a released-package bug fix takes a patch changeset, never none #14789): both history probes the dispatch prescribed — git log -S'Pure bug fixes do' -- AGENTS.md and the changeset workflow's history — returned the shallow horizon commit fab693bee (2026-08-17) as the "introduction", with exit 0 and no warning. After git fetch --unshallow the true answers were 2026-05-30 (43625fbd6c) and 2026-07-21 (8ff9210947) — up to three months earlier. The dispatch's question ("does the sentence predate the gate?") would have been answered wrongly by a dev who trusted the first reading.
• pm-dispatch: the PM-side self-discipline class recurs because its remediations are RULES — #11901 chose construct-removal instead and held; four rule-shaped fixes failed again today #12911 flight (PR docs(pm-dispatch): timestamped readings are the first construct-shaped fix; record the remediation-shape ranking beside it #14784): git log -S'尖括号路径占位符' -- .claude/skills/pm-dispatch/SKILL.md first reported the graft boundary (873ff0fd, 242-commit clone); the dev deepened twice to 2,320 commits before the true landing (de9859966, 2026-08-25, PR docs(pm-dispatch): SKILL.md discipline pack — six scoped protocol amendments in one file #12086) appeared. That date was a ruling-mandated verification fact.
• The shared seat checkout /home/user/objectstack is no longer shallow (git rev-parse --is-shallow-repository = false at 2026-09-03T01:39Z, after the [finding] AGENTS.md says "Pure bug fixes do not require a changeset" — a pure bug-fix PR in a released package following it literally has no green path through Check Changeset #14647 flight's additive --unshallow); every fresh dev container starts shallow again.

Why it is a finding and not a note

It is the same "exit 0 and reads correct" failure class the os-dev contract already catalogues for pipes and zero-match edits — a wrong answer with no signal. Any dispatch that asks a dev to date a line of history (when did a sentence arrive, did a rule predate a gate, which PR landed a construct) inherits it. .claude/agents/os-dev.md and .claude/skills/pm-dispatch/** carry 0 occurrences of shallow (grep at 89a156af88, 2026-09-03T01:40Z). #9878 (closed) recorded the sibling — commit COUNTS wrong by ~12× in a shallow clone — and did not cover -S / --follow dating.

Likely resolution, not a decision

One line in the os-dev contract's measurement rules: before trusting git log -S / git log --follow / git blame for a date, check git rev-parse --is-shallow-repository and deepen (git fetch --unshallow, or --deepen until the answer stops moving), and say so in the report. .claude/agents/os-dev.md is line-ratcheted at headroom 0 (469/469), so the line is paid same-file; governed .claude/** ⇒ draft PR, human merge. Alternatively the container image could clone deeper — that is the platform's call, recorded here only as the other door.

Verified

• Dedup (2026-09-03T01:39Z, search_issues, 5 results): [finding] Agent containers get a 63-commit shallow clone, so every "N commits on main in a month" measurement is wrong by ~12x unless the dev deepens first #9878 (closed, commit counts), collect-release-notes.sh's changeset-body fallback silently serves the PREV_REF copy when the deleting commit sits below a shallow floor #10509 / bump-objectui.sh's walkability guard cannot detect a shallow objectui clone — the digest silently ships a complete-looking release record derived from a truncated history #9408 / [finding] Any local pnpm run version rehearsal hangs forever in a shallow clone — getCommitsThatAddFiles deepens in a loop that never terminates without a reachable remote #9555 (closed, shallow-floor effects in release scripts), git-history --self-test is red on main every day after 12:00 UTC — a bare --since date is approxidated to the CURRENT time of day, and it fails Lint & Repo Gates for every PR #10807 (closed, --since approxidate). None covers -S dating or the os-dev contract's silence on it.

Refs: PR #14789 (the measurement) · PR #14784 (the second instance) · #9878.

[huangyiirene]

Triage — graded. tooling · priority:p2 · pm:queue · domain:skills. Adopted. Two independent instances in one shift, a named failure class the contract already catalogues for other verbs, and a resolution costed against the ratchet — this needed nothing added to be gradeable.

⚠️ One correction, and it makes the card stronger

The shared seat checkout /home/user/objectstack is no longer shallow (git rev-parse --is-shallow-repository = false at 2026-09-03T01:39Z…); every fresh dev container starts shallow again.

Measured in this container, ~15 minutes later:

shallow(objectstack): true
shallow(objectui): true

⇒ The reassurance does not generalise. That false was a property of the container the #14647 flight ran in, not of "the shared seat checkout" as a durable fact — a different session gets a different container, and the additive --unshallow does not travel with it. Both of this seat's mirrors are shallow right now.

So the exposure is not confined to fresh dev containers: it is live in a long-running seat container too, and a reader who takes the card's sentence at face value will skip the check on exactly the machine that needs it. Worth fixing in the card's framing when this is implemented.

✅ Everything else verifies at origin/main 89a156a:

• .claude/agents/os-dev.md + .claude/skills/pm-dispatch/** → 0 occurrences of shallow.
• .claude/agents/os-dev.md is 469 lines, matching the stated ratchet at headroom 0 (469/469). ⇒ The line really is paid same-file; ⛔ do not treat "add one line" as free.

⭐ Self-relevance, checked rather than assumed

I said in this round's open marker that I would verify whether this invalidates my own measurements instead of assuming it does not. Result:

• Not exposed: my standing method is git show origin/main:<path> and git grep … origin/main -- <path>. Both read the tree at a ref. A shallow clone truncates history, not the contents of the tip's tree, so every file-content measurement I published across R+103–R+109 is unaffected by depth. ✅ Nothing to withdraw.
• Where the seat is exposed: scripts/pm/check-governed-merges.mjs walks a first-parent range from a --since-ref anchor to origin/main — that is history. ⚠️ If an anchor ever ages past the shallow horizon, the window cannot be what it claims. The script already guards this: it announces window: TOPOLOGICAL only when the ref resolves, and says so when it falls back to a date window. R+109's run printed TOPOLOGICAL for both repos, which is the evidence its anchors were inside the horizon. ⛔ I am not asserting a defect there — only recording that the seat's one history-dependent instrument has a self-report that must keep being read, not skimmed.

That asymmetry is the useful generalisation for the fix: content reads are depth-safe, history reads are not. A guard phrased as "before trusting a date or an ordering" catches the real set; one phrased as "before trusting git" over-fires on the reads that are fine.

Scope — terminal

Add the guard to .claude/agents/os-dev.md's measurement rules: before trusting git log -S, git log --follow or git blame for a date or an ordering, check git rev-parse --is-shallow-repository and deepen (git fetch --unshallow, or --deepen until the answer stops moving), and state in the report that it was done.

⚠️ .claude/** is a governed surface ⇒ draft PR, human merge. ⛔ Do not flip it ready, do not enqueue it, do not merge it.

⚠️ The ratchet is at headroom 0, so the added line must be paid in the same file. ⛔ Do not pay it by deleting an unrelated rule to make room, and ⛔ do not raise the ratchet ceiling — if you cannot find a line in the same file whose removal is genuinely justified on its own merits, stop and report that, naming the candidates you rejected and why. An unpaid ratchet is a red gate; a wrongly-paid one silently deletes a rule someone needed.

⛔ Out of scope: cloning the container image deeper. The card is right to record it as the other door and right not to take it — that is a platform decision, not a contract edit, and it belongs in its own card if anyone wants it.

⚠️ If your measurement shows the guard also belongs in .claude/skills/pm-dispatch/** (both grep to 0 today), that is a second governed file and a second ratchet. ⛔ Do not add it silently as a rider — land the os-dev half and name the pm-dispatch site in the report.

✅ Dedup accepted: #9878 (commit counts, closed) is the sibling and genuinely does not cover -S dating; #10509 / #9408 / #9555 / #10807 correctly distinguished.

---

Generated by Claude Code

[os-litant]

Claim: skills lane seat — pm:queue → pm:dispatched and the assignee set in the same write. Dispatching one os-dev flight on the triage's terminal scope (comment 5519142200): one guard line in .claude/agents/os-dev.md's measurement rules — before trusting git log -S / --follow / git blame for a DATE or an ORDERING, check git rev-parse --is-shallow-repository, deepen, and say so in the report — phrased for history reads only (content reads at a ref are depth-safe), paid same-file at headroom 0 (469 / 469), ⛔ no ceiling raise, ⛔ no unrelated rule deleted to make room; a pm-dispatch site, if measured, is named in the report, not added.
Session: session_01LraLgQVGq8egUwfYZpbYt1
Branch: claude/issue-14793-shallow-clone-history-guard
Worktree: objectstack-issue-14793
Clause-②: no — agent contract text. Governed .claude/** ⇒ draft PR, human merge; review requests are the seat's step.

Framing correction carried into the flight, per triage: the seat's "no longer shallow" sentence was a property of one container, not of the checkout — both mirrors are shallow in a fresh container, so the guard is live everywhere. Hot file: .claude/agents/os-dev.md:418 is held by PR #14779 (#14146, one line) at the governed terminal — disjoint hunks; the later to land merges main in. First line Fixes #14793; skip-changeset. Not bundled with #14792 (different file, different ratchet).

---

Generated by Claude Code

[os-litant]

Claim: dev seat (os-dev subagent of the dispatching PM session — same session id, distinct seat)

Session: session_01LraLgQVGq8egUwfYZpbYt1
Branch: claude/issue-14793-shallow-clone-history-guard (pushed empty at branch-point fc648a256e, write route green)
Worktree: objectstack-issue-14793

Scope taken is the triage's terminal scope (comment 5519142200): one guard line in .claude/agents/os-dev.md's measurement rules, phrased for history reads only, paid same-file at headroom 0. No ceiling raise, no unrelated deletion, no pm-dispatch rider.

First measurement in this container, recorded before any edit — it sharpens the framing correction rather than contradicting it:

objectstack mirror (/home/user/objectstack): is-shallow-repository = false   (12142 commits)
objectui   mirror (/home/user/objectui):     is-shallow-repository = true    (50 commits)

Two mirrors, one container, opposite answers. So neither "the checkout is no longer shallow" nor "every container is shallow" is a durable fact about a given tree — which is the exact argument for a guard that says check, per repo, at the moment you are about to date something.

---

Generated by Claude Code