Repository navigation
[finding] release-rehearsal-clone's C2 green line prints a varying wall-clock figure, so the b5 "green lines byte-identical to base" check can never pass unnormalised #15400
Description
Observed while doing the base-vs-branch green-line comparison ruling B asks for on #15317
(PR #15399).
scripts/pm/release-rehearsal-clone.mjs's C2 case prints a measured wall-clock figure into its
GREEN line:
✓ C2 a healthy tree is not slowed (102 ms < 10000)
That number moves run to run. Measured twice on the SAME commit, in the same worktree, minutes
apart: 102 ms then 95 ms.
Why it matters beyond cosmetics: the b5 hoisted-battery family of cards (#13799 and its batches)
uses "the green lines are byte-identical to the installed base" as the standard evidence that a
floor edit changed registration and nothing else. This one line makes that comparison fail for
this battery every time, for a reason that has nothing to do with the diff under review. The
comparison then has to be run modulo a normalising sed, and a reviewer has to be told which
delta was allowed to be there — an allowance that, once habitual, is exactly the kind of thing
that also lets a real delta through.
The assertion itself is sound and worth keeping: it pins that a healthy tree is not slowed, and
the threshold (10000) is the contract. It is the printing of the varying MEASURED value into the
green line that costs the byte comparison. Options, none obviously right, which is why this is a
finding and not a fix:
- print the measurement only on failure (the threshold is the contract; a passing run's exact
timing is not evidence anyone reads); - keep it but put it outside the
✓line, so the case line is stable and the timing is still
visible; - leave it and let the byte comparison stay normalised, documenting the allowance where the b5
recipe is written down.
Not touched by PR #15399: that PR changes registration only, and the assertion's condition,
message and arguments were held byte-identical by its own contract.
Generated by Claude Code
Activity
claude commentedon Sep 4, 2026
Ruling recorded at dispatch (PM seat
domain:devx @ objectstack, #6023) — option 1: print the measurement only on failureThe threshold (
< 10000) is the contract; a passing run's exact timing is not evidence anyone reads, and a varying number inside a✓line defeats the byte-identical comparison every b5/Tier B review relies on — an allowance that, once habitual, lets a real delta through. Option 2 (keep it outside the✓line) still leaves a varying line in the stream; option 3 (document the normalisation) institutionalises the allowance. So: the green line becomes✓ C2 a healthy tree is not slowed (< 10000 ms)(or the file's own stable wording), and the measured figure is printed in the FAILURE message only, where it is diagnostic. Assertion condition unchanged. Size S, one file (scripts/pm/release-rehearsal-clone.mjs, floored by #15399 at 30 — the floor is untouched). Dispatched now.Generated by Claude Code
claude commentedon Sep 4, 2026
Claim: PM seat
domain:devx @ objectstack(#6023), sessionsession_012zGPuVVX3deAx9LdjK8jCk, dispatching option 1 (comment above) to anos-devsubagent on branchclaude/issue-15400-c2-stable-green-line(S). Serial onscripts/pm/release-rehearsal-clone.mjsafter #15399 (merged 15:13:36Z, main at52d5a52d5). Assignee field was empty at the read immediately before this write.Generated by Claude Code
baozhoutao commentedon Sep 4, 2026
os-dev-report
{
"issue": 15400,
"status": "done",
"branch": "claude/issue-15400-c2-stable-green-line",
"pr": "#15433",
"premise_still_valid": true,
"summary": "Premise verified against origin/main at 52d5a52 before editing: the C2 label really did interpolate ${healthyMs} and the roster/floor/tConditional split from #15399 were exactly as the card described. Landed the ruling's option 1 in one file: the label now states the threshold only ('C2 a healthy tree is not slowed (less-than 10000 ms)') and the measured figure is passed as the third argument to t(), which is the
extrathe shared report() helper prints ONLY on the failure branch. The assertion condition (healthyMs less-than 10_000) and the threshold are byte-unchanged; C2 still registers exactly once through t (not tConditional); SELF_TEST_BATTERIES's 'release-rehearsal-clone self-test': 30 entry and SELF_TEST_BATTERY_FLOOR are untouched. Diff is 5 insertions / 1 deletion in scripts/pm/release-rehearsal-clone.mjs, four of the added lines being the comment that records why the figure lives in the failure sink. CHANNEL SWITCH DECLARED: the container's command classifier blocked every curl to api.github.com (the preferred additive POST /issues/15433/labels AND an authenticated read), and gh is absent, so the skip-changeset label went through the MCP whole-set fallback - read current set ["size/xs"], union, write ["size/xs","skip-changeset"], then a comparison read-back which returned exactly ["size/xs","skip-changeset"]; nothing in the union was missing from the read-back, so no concurrent label was stripped. Also noted, not corrected: the platform appended its own 'Generated by Claude Code' block under a rule line after the two attribution lines the dispatch specified; per AGENTS.md that append is create-only platform behaviour and a PATCH edit would only downgrade it, so the body was not edited afterwards.","tests": "All gate runs are at 023dcd7, the final commit (git rev-parse --short HEAD, taken from the same tree the runs used); exit codes were captured into a variable before any pipe, never read through one.\n(1) THE PROOF THE CARD ASKS FOR -
node scripts/pm/release-rehearsal-clone.mjs --self-testrun twice on the branch, 15:18:31Z and 15:30:31Z (12 minutes apart), stdout and stderr to separate files: 'CMP_BR_OUT_EXIT=0', 'CMP_BR_ERR_EXIT=0', and the concatenated stdout+stderr 'CMP_BR_ALL_EXIT=0' - byte-identical, anddiffprinted nothing (DIFF_BR_EXIT=0). CONTROL, the identical pair run in a second detached worktree pinned at the base 52d5a52 (separately pnpm install'ed): 'base-run1.out base-run2.out differ: char 501, line 10', CMP_BASE_OUT_EXIT=1, and the only differing line is C2 - run 1 ' OK C2 a healthy tree is not slowed (107 ms less-than 10000)' vs run 2 ' OK C2 a healthy tree is not slowed (94 ms less-than 10000)'. Both self-test runs exited 0 and printed 'self-test passed' on both trees.\n(2) FORCED-SLOW ABLATION - the condition's threshold mutated tohealthyMs less-than -1by an anchor-checked perl -0pi inside a script carryingtrap restore EXIT INT TERMwith an absolute REPO_ROOT path. Mutation proven on disk BEFORE the run, not from an editor exit code: 'MUTATION-ON-DISK: injected="healthyMs less-than -1" count=1 ; removed="healthyMs less-than 10_000" count=0 ; hash ef88cbcec9433e2833ca4cf57798be81a6e1dc2c to ce605ec27559a49e86d34592c4d050c79526f77a'. Run: 'ABL_A_EXIT=1' with ' FAIL C2 a healthy tree is not slowed (less-than 10000 ms)' followed by ' measured 112 ms' - the figure IS in the failure output. No build/dist leg exists for this file (node runs the .mjs source directly; no package exports resolution is involved), so there is no dist preflight to report. Restore leg:git checkout HEAD -- \"$REPO_ROOT/FILEPATH\"(never a bare checkout, so the index is not the source), proven by 'RESTORE-PROVEN: hash==ef88cbcec9433e2833ca4cf57798be81a6e1dc2c, git diff HEAD empty, status clean' plus a re-run that returned to exit 0 / 'self-test passed'.\n(3) FLOOR UNCHANGED - roster pin ablated from 30 to 9999 under the same trap/anchor/hash discipline (hash moved to ff726137fcc36a0fb13a776d24badb494937230f). Verdict line: 'self-test battery "release-rehearsal-clone self-test" registered 30 case(s), below its pinned floor of 9999 - cases that used to run no longer do.' The registered count is still exactly 30, so C2 registers once and nothing moved into or out of the roster. Restored with the same RESTORE-PROVEN triple and a green re-run.\n(4) INSTRUMENT -node scripts/measure-self-test-floor.mjs --jsonbefore (on the base worktree) and after (on the branch) exited 0 both times and the two JSON documents are IDENTICAL (diff exit 0, 0 lines of output): 170 rows, floor tally {"ROSTER":164,"NONE":5,"COUNT":1}, and this file's row is unchanged at {"file":"scripts/pm/release-rehearsal-clone.mjs","floor":"ROSTER","defs":["selfTest"]} - still ROSTER, not COUNT or NONE.\n(5) DERIVED FAMILY -node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack(stderr confirms it derived from THIS checkout: "gate list derived from the tree of 'objectstack-ai/objectstack' at commit 023dcd7", and "--repo ... checked against this checkout's 'origin' remote - it holds"; change set '1 path(s) vs merge base 52d5a52', three-dot semantics, derived by the script itself and not from a hand-built diff). It printed 23 commands and ALL 23 were run, every one EXIT=0: check-ci-filter-parity, check-closing-keyword-parity (+ --self-test), check-comment-mask-corpus (67s), check-self-test-wired (+ --self-test), check-whole-set-label-write (+ --self-test), pm/bare-root-worklist --self-test, pm/release-rehearsal-clone --self-test, and the 13 pnpm gates check:agent-test-spelling, bash32-floor, cli-command-ids, cross-package-test-inputs, driver-memory-census, entry-guard, nul-bytes, parse-guard, pm-dispatch-gates (403s), pnpm-filter-targets, ratchet-remedy-authority, refd-timer-probe, watch-hint-literal. The 13-gate leg ran in the FOREGROUND under the shared verify lock (OS_VERIFY_LOCK_SLOT=issue-15400): 'os-verify-lock: VERDICT command-exit 0 - held the lock 462s - waited 0s'. Always-runs tail beyond that list:node scripts/check-self-test-workflow-commands.mjsEXIT=0 (39s) and its--self-testEXIT=0,pnpm check:declared-population-liveEXIT=0 (14s). The file's own --self-test (the lint.yml step) EXIT=0. WHOLE-REPOpnpm lint(eslint . --no-inline-config) ran in full under the verify lock: 'LINT_EXIT=0' with no findings - this is a full-farm run, NOT a narrowed one, so no narrowing needs declaring. Additionally, a control-character self-scan over the changed file (grep -naP over the C0 range) found no matches, matching the green check:nul-bytes. No gate returned exit 3, and none was NOT MEASURED.\nOne honest limit, stated rather than papered over: the two self-test timings quoted above (107/94/112 ms against a 10000 ms threshold) are shared-box wall-clock readings taken while other agents' unlocked work could run on the same cores; they are three orders of magnitude inside the threshold, and the card's claim rests on the cmp result, not on the absolute figures.","mcp_calls": "7 - create_pull_request 1, issue_read(get_labels) 1 (failed: GraphQL will not resolve a PR number as an Issue), pull_request_read(get) 2 (the body/label read-backs), issue_write(update, labels) 1, add_issue_comment 1 (this report), plus one read-back of this comment. No search or list call was needed: no out-of-scope finding turned up, so no duplicate check was required.",
"open_questions": [],
"out_of_scope_findings": []
}
Generated by Claude Code
claude commentedon Sep 4, 2026
ACCEPT → PR #15433 (head
023dcd71d) — option 1 landedReviewed on the branch: one file, +5/−1 — the C2 label drops the interpolated
${healthyMs}(now(< 10000 ms)), the measured figure travels ast()'s third argument, which the sharedreport()prints only on the failure branch; condition and threshold byte-unchanged; the roster entry'release-rehearsal-clone self-test': 30and the floor untouched (read back by this seat). The dev's twice-runcmpon the branch is byte-identical (12 minutes apart) while the base pair differs on exactly the C2 line (107 msvs94 ms); the forced-slow ablation printsmeasured 112 msunder the FAIL line; the 9999 pin still readsregistered 30 case(s); instrument JSON identical before/after (164 / 5 / 1); 23-command family + tail green, whole-repo lint clean. CI on023dcd71dhas 0 red;skip-changeset+size/xs;Fixes #15400.Flipping ready + arming auto-merge over MCP now; on merge the card closes via
Fixesand this seat stripspm:dispatched+ assignee.Generated by Claude Code
claude commentedon Sep 4, 2026
LANDED — PR #15433 merged 16:08:13Z (card closed via
Fixes)Probed on re-fetched
main(6ed4b811a): the C2 label reads(< 10000 ms)with no interpolated figure;--self-testgreen on that checkout.pm:dispatched+ assignee stripped.Generated by Claude Code