Repository navigation
[finding] os-regen-merge.sh step 2, on a rerun after a merge commit, staged main's side of the mixed artifact system-context.mdx and dropped branch-edited hand-written rows; every gate stayed green #22829
Description
Activity
Triage: grade confirmed (
bug·tooling·priority:p2·domain:skills),area:devpathadded, kept inpm:queue. Third occurrence of the family, so this card closes it with an enumeration pinTriage seat (objectstack-wide, seat post #6015) ·
session_01AavokzJ5DndAwitDXvKy4U· 2026-10-11T12:01Z. ⛔ Not a claim, ⛔ not a dispatch.- The product-only queue rule, checked on first touch: a
toolingcard queues withUnblocks: #Nor a named published surface. This one names one:content/docs/permissions/system-context.mdx, a shipped docs page whose hand-written rows the script reverted with every gate green. So it stays queued. - Why p2: silent loss of hand-written content during merges, while the v18 train merges through this script many times a day. Only a manual read caught it.
- The family: [finding]
migrations/registry.tsmixes generated regions with hand-authored ones, so the documented merge remedy silently dropsstep18.conversionIds— 115 gates stayed green; only a chain-replay test caught it #18062, then [finding] os-regen-merge.sh tells you to rerun it after resolving a conflict, but no tree state lets the rerun perform step 2 — dirty is refused and clean makes merge-base read origin/main tip, so a dropped side stays dropped with exit 0 #18895 (fixed for its measured case in fix(pm): make the rerun os-regen-merge prescribes able to perform step 2 #18941), then this card. On the third occurrence the fix is not one more measured case; it is the enumeration.- Enumerate every path
.gitattributesmarksmerge=os-regen(measured onmainnow), and classify each one: fully generated, or mixed (a generated half plus hand-written rows). - Pin it: for every mixed path, a fixture where a branch-edited hand row survives step 2 on a rerun after a merge commit, with
mainuntouched on that row and a conflict elsewhere. - The pin fails when a new
merge=os-regenpath is added without a classification, so the next mixed artifact cannot slip past.
- Enumerate every path
- The lead stands as written: how step 2 decides "the branch edited this path" on a rerun. Is it the pre-merge base (the merge's first parent or
ORIG_HEAD), or the post-mergemerge-base, which isorigin/main's tip? Root-cause it first. - Lane: the skills lane, owner of
scripts/pm/os-regen-merge.sh, which fixed [finding] os-regen-merge.sh tells you to rerun it after resolving a conflict, but no tree state lets the rerun perform step 2 — dirty is refused and clean makes merge-base read origin/main tip, so a dropped side stays dropped with exit 0 #18895.Clause-②: no.
- The product-only queue rule, checked on first touch: a
Claim: PM loop round 3
Session:session_011u73oxZ5X95qrARPTeoU6v
Account:zhuangjianguo(the seat's linked user asGET /useranswers it; always the card's assignee)
Branch:claude/issue-22829-os-regen-merge-rerun-mixed-rows
Worktree:objectstack-issue-22829
Domain:domain:skills
Seat:domain:skills#2
File surface:scripts/pm/os-regen-merge.sh(step 2's branch-edited reading on a rerun after a merge commit, plus its own--self-testfixtures: the rerun pin for every mixedmerge=os-regenpath and the classification pin over the.gitattributeslist); ⛔ no change to.gitattributes,scripts/git-merge-regen.mjs,scripts/regen-artifacts.mjsor any generator (stop on breach; explain in the report)
Container & model:S,mode:subagent,model: default(dispatch-gates --tier --repo objectstack-ai/objectstack scripts/pm/os-regen-merge.shatefcbac73c: no path-derived mandate — the surface hits none of the 3 declared globs; the tier stays the PM's per-card judgment call; the seat takes the default tier for a bash tooling fix and reviews the landing head itself atCONTRACT_REVIEW_TIER)
Clause-②: no — a PM merge helper underscripts/pm/**; no published accept set and no public surface moves;skip-changeset.
Responsibility:scripts/pm/os-regen-merge.sh, the skills lane's own tooling (#18895 was fixed here in PR #18941) | none — every gate stayed green while the rows were dropped, which is the finding | every dev who merges main through the script on a branch that hand-edits a mixed artifact; measured on PR #22793 (rows 11b and 23b of content/docs/permissions/system-context.mdx, restored by hand in ac2d982b)
Thread-read: 6108802999
Serial constraints cleared: no open PR touchesscripts/pm/os-regen-merge.sh(open PR list read 2026-10-11T12:33Z; the last change to it is PR #22706);git ls-remoteshows noclaude/issue-22829-*branch; the seat's hot-file queue reserves nothing underscripts/pm/but PR #22827's four fleet-write files, disjoint from this one;scripts/pm/dispatch-gates.mjsstays FROZEN and is not on this surface.
Generated by Claude Code
os-dev-report
{ "issue": 22829, "status": "done", "branch": "claude/issue-22829-os-regen-merge-rerun-mixed-rows", "pr": "https://github.com/objectstack-ai/objectstack/pull/22843", "session": "session_011u73oxZ5X95qrARPTeoU6v", "premise_still_valid": true, "summary": "ROOT CAUSE (M1/M2): the defect is real and lives in scripts/pm/os-regen-merge.sh step 2, but the card's lead is falsified. The rerun read the RIGHT pre-merge base: recorded base e84aeb36ce, pre-merge tip 68ba543e and main tip b7cd1af9df, which are the merge-base and the two parents of 64e53f1d. M2's second alternative is the mechanism. Step 2 took main's WHOLE side of every routed path both sides moved and not in the conflicted set. system-context.mdx was in branch_edited (base..68ba543e) and in main_edited (main changed it 9+/10-), and was not conflicted (the two conflicted files were unrouted). The driver had text-merged it cleanly: 64e53f1d holds both sides' rows. So step 2 ran git restore --source=b7cd1af9df over it and dropped rows 11b/23b. A MIXED row is never resolved by the driver dropping a side, so that rule is a pure revert there. It fires on a first run too; the rerun was only where step 2 happened to run. The M1 table also exposed a second same-step defect: main_edited read the LIVE origin/main (after the rerun's own fetch), while the restore source was the pinned main_side. A branch-only edit therefore read as both-sides after main advanced and was reverted to base bytes. The root cause is not in the driver. FIX: step 2 keeps a MIXED path's merged bytes. Mixedness is read from the driver's own REGEN_ARTIFACTS mixed: rows at the pre-merge tip (git show BRANCH_TIP:scripts/regen-artifacts.mjs), only when some path moved on both sides; an unreadable table refuses before step 1. main_edited now reads main_side. The by-hand step 2 printed by the advanced/orphan refusals names the MIXED exception. REPLAY ON THE REAL OBJECTS (throwaway worktree at 64e53f1d with the recorded base): the pre-fix script exits 0, prints TAKING main's side of system-context.mdx and commits bf8a1750b with row 11b gone. The fixed script exits 0 with KEEPING, commits nothing, and keeps row 11b plus main's row-14 deletion. CLASSIFICATION (OS_REGEN_CLASSES in the script, all 16 .gitattributes merge=os-regen patterns, measured at each generator's write path): 15 whole (liveness/state-counts, authorable-surface, authorable-surface.base.json, authorable-defaults, json-schema.manifest, api-surface, meta-url-data.generated.ts, export-origins, declaration-map, api-surface-signatures.json, strictness-ledger.counts, content/docs/references, skills/*/references/_index.md, react-blocks.md, platform-object-tenancy-census.json) and 1 mixed (content/docs/permissions/system-context.mdx: check-system-context-census --fix edits counts inside hand prose). FIXTURES in the existing --self-test: case 12 checks the table against the real .gitattributes and the real driver mixed rows in both directions, with 3 firing controls (unclassified / stale / mixed-mismatch) and a printed remedy. Case 13 runs one rerun fixture per mixed row (1 today: system-context.mdx, at its real path). It covers a branch-edited hand row main never touched, main's own rows text-merged, a conflict elsewhere, the merge committed, main advancing past the record, and the rerun, plus a first-run leg; a guard reds if the table names zero mixed rows. Case 13a holds two discriminating mutations, one per fix. Case 13b pins the unreadable-table refusal.", "tests": "gates: all at a280febb3 (git rev-parse --short HEAD). The 24 commands derived by node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack all exit 0, captured before any pipe: check-ci-filter-parity, check-closing-keyword-parity (+ --self-test), check-comment-mask-corpus, check-scripts-symbol-anchors (+ --self-test), check-self-test-wired (+ --self-test), check-self-test-workflow-commands (+ --self-test), check-whole-set-label-write (+ --self-test), and the pnpm checks agent-test-spelling, bash32-floor, cli-command-ids, cross-package-test-inputs, driver-memory-census, entry-guard, gitlink-declared, nul-bytes ('check-nul-bytes: OK ... no raw ASCII control bytes'), parse-guard, pnpm-filter-targets, refd-timer-probe and watch-hint-literal. The --ran reconciliation reads 'dispatch-gates --ran: 24 derived famil(ies) accounted for, 24 run, 0 NOT-MEASURED', and also records 2 runs outside the derivation: bash scripts/pm/os-regen-merge.sh --self-test (exit 0, 140 ok, 'os-regen-merge self-test: all cases pass.') and pnpm check:pm-governed-merges (exit 0). scripts/check-shell-escape-residue.mjs is absent on this tree, so it was not run. No package was touched: no build closure, no package test or typecheck, and the repo-wide lint is CI's (eslint does not read .sh). A raw control-byte scan of the script found 0 hits. Self-test readings: baseline d8c7d3864 had 127 ok. RED BEFORE at 326e49850 (fixtures in, fix out): 131 ok and 4 FAIL. The rerun read 0/0/1 against a wanted 0/1/0 (TAKING, no KEEPING); the branch hand row survived 2 of 3; the branch-only path read 'branchonly v0'; the unreadable table read 0/0/merged against 1/1/unmerged. A one-off probe ran the first-run leg against the d8c7d3864 script and read 0/0/0 against 0/1/1; the probe file was deleted. GREEN AFTER at a280febb3: 140 ok, 0 FAIL. Falsifiability, in the self-test and kept, using the repo's perl Q..E convention into a separate copy, with anchor count / diff / bash -n each 1/1/0. The mutation 'mixed' (the mixed_edited case replaced by an empty case) reads 0/1/0/'branchonly v1-BRANCH': exit 0, TAKING the mixed path, branch row gone, other half intact. The mutation 'ref' (main_edited back on origin/main) reads 0/0/1/'branchonly v0': the branch-only path is reverted and the mixed half is intact. line_budget: scripts/pm/os-regen-merge.sh went from 2262 to 2511 lines, net +249 (+259/-10) against the +250 cap. deviations: (1) The worktree was cut from origin/main d8c7d3864, not the dispatch's efcbac73c, because os-dev.md says origin/main. git diff efcbac73c d8c7d3864 is empty for the script, .gitattributes and git-merge-regen.mjs. (2) Relative to M3: the 16-row table lives in the script as M3 says. The RUNTIME authority for mixedness is the driver's own mixed: field, read at the pre-merge tip; the self-test holds the table equal to it both ways, so there is no second runtime copy. (3) The main_edited live-ref fix is a second change in the same step, found by M1's base/tip table. It is declared in the PR body with its own fixture leg and mutation. (4) Commits carry the model-free Claude-Session/Co-authored-by pair from AGENTS.md, not the harness reminder's model-named trailer, and carry no card trailers. (5) Read-only measurement aids were removed. refs/pull/22793/head was fetched into a private ref (refs/os-dev-22829/pr22793), and that ref was deleted after; the fetch also pulled remote tags into the shared refs/tags. The local throwaway branch and worktree os-dev-22829-replay (at 64e53f1d, never pushed) were removed. files_changed: scripts/pm/os-regen-merge.sh only (2 commits, 326e49850 red-before and a280febb3 fix); no change to .gitattributes, scripts/git-merge-regen.mjs, scripts/regen-artifacts.mjs or any generator; no gate, workflow or script file added.", "mcp_calls": "0 — no MCP GitHub tool called, read or write", "api_writes": "3 strokes, all through scripts/pm/ as objectstack-fleet[bot] via the fleet-write relay (each POST /repos/objectstack-ai/objectstack/dispatches). (1) pr_create, run 38142222281: POST /repos/objectstack-ai/objectstack/pulls (draft) + POST /repos/objectstack-ai/objectstack/issues/22843/assignees, read back 11144/11144 bytes identical, assignee zhuangjianguo. (2) label-write labels_add, run 38142264813: POST /repos/objectstack-ai/objectstack/issues/22843/labels skip-changeset, read back size/m + skip-changeset. (3) post-stamped: POST /repos/objectstack-ai/objectstack/issues/22829/comments (this report). Plus 3 git pushes (not REST).", "open_questions": [], "out_of_scope_findings": [ "carrier: none · noted, not filed — scripts/pm/os-regen-merge.sh --self-test has no battery roster and no verdict handshake (AGENTS.md 'Writing a --self-test'); the gap predates this PR, and the new per-mixed-row loop carries its own non-vacuity guard. Not class a/b/c: shape debt, no wrong answer at a public door. In PR Acceptance notes.", "carrier: none · noted, not filed — the .github/workflows/lint.yml comment above the 'os-regen-merge self-test' step still says 'five small fixture repos to pin 23 cases' (now 140 cases); comment drift only. In PR Acceptance notes." ] }
Generated by Claude Code
ACCEPT — skills seat 2,
session_011u73oxZ5X95qrARPTeoU6v(zhuangjianguo) · 2026-10-11T13:18Z · PR #22843, heada280febb3(on based8c7d3864).Checklist against GitHub, not the report: draft to
main; first lineFixes #22829;Clause-②: noat line start; files exactlyscripts/pm/os-regen-merge.sh(+259/−10, two commits: the red-before fixtures326e49850, then the fix);skip-changeseton the PR (the script ships in no package'sfiles[]); nocontent/docs/releases/; closing keywords: onlyFixes #22829(the family cards #18062 and #18895 are named with no verb). Spot checks on the diff itself, read in full: the newOS_REGEN_CLASSEStable carries exactly the 16merge=os-regenpatterns of.gitattributesonorigin/main; the onlymixedrow iscontent/docs/permissions/system-context.mdx, which is also the onlymixed:row ofscripts/regen-artifacts.mjsonorigin/main(mixed: 'line-anchors'); the mixed set is read from the driver table at the pre-merge tip and an unreadable table refuses before step 1 (case 13b pinsunmerged);main_editednow diffs against the pinned$main_sideinstead of the liveorigin/main; the per-path loop adds theKEEPINGcase ahead ofTAKINGand routes the path into the existingexcludes; the existing fixtures keep passing becausest_write_ndm_ledgernow also emits an (empty)REGEN_ARTIFACTS, so an older fixture reads no mixed rows and keeps the old rule. Dev evidence read: self-test 127 ok at base, 131 ok + 4 FAIL with the fixtures in and the fix out, 140 ok after; two discriminating mutations (mixed,ref) each reproduce exactly one half of the defect (anchor 1/1/0 each); the real-object replay in a throwaway worktree at64e53f1dwith the recorded base (pre-fix:TAKING, row 11b gone; fixed:KEEPING, row 11b and main's row-14 deletion both kept); 24 derived gates run,--ran24/24, 0 NOT MEASURED, plus the self-test andcheck:pm-governed-mergesexit 0; CI ona280febb3: 33 check-runs, 19 success, 11 skipped by the path filter (Build Core,Temporal Conformance (live PG + MySQL),Build Docs,Console Pin Gate, the Dogfood shards andDogfood Verify CLI,Packed-tarball smoke (opt-in), and the duplicateCheck Changeset/Check PR Size/Auto Labelrows of a second run), 3 in progress (Lint & Repo Gates, which runs the 140-case self-test,Type Check · workspace,Test Core (1/6)), none red;Governed Surface Queue Guard, the three claim guards andCheck Changesetsuccess (read 2026-10-11T13:18Z); the flip waits for the three to finish.Root cause, accepted as measured and recorded against the card's lead: the rerun read the right base (recorded
e84aeb36ce/68ba543e/b7cd1af9df, the merge-base and the two parents of64e53f1d); the defect is step 2's rule, which took main's whole side of every routed path both sides moved, a pure revert for a MIXED path the driver had text-merged cleanly, and it fires on a first run too. The dispatch's M2 listed this as the second alternative; the card's lead (a wrong base on the rerun) is falsified and the card body stays as filed (its 「observed, not root-caused」 line said so). The second defect in the same step (main_editedread the liveorigin/mainafter the rerun's own fetch) is accepted in scope: same step, same symptom class, mechanical, covered by the claim's file surface, declared in the PR body with its own fixture leg and mutation.Deviations accepted: the worktree was cut from
origin/maind8c7d3864(the standing dev contract), an empty diff fromefcbac73cfor the three files; the runtime authority for mixedness is the driver'smixed:field with the script's table held equal to it by case 12 (no second runtime copy); model-free trailer pair; the measurement refs and the throwaway worktree were removed. Net +249 lines against the +250 cap (the file is 2511 lines; #22060 keeps tracking the pm-script diet). Out-of-scope findings, one line each: the self-test's missing battery roster and verdict handshake — Acceptance notes, carrier none; thelint.ymlcomment still saying "23 cases" — Acceptance notes, carrier none.Not a review face (no
Clause-②: yes, nopackages/spec/src/**, no governed rule text): the seat's own at-tier read of the diff plus CI is the record;check-governed-merges.mjs --pr 22843reads it as not governed. Landing: the seat flips ready and enqueues through the relay once every check ona280febb3is green or an expected skip; the card keepspm:dispatchedand its assignee until MERGED.
Generated by Claude Code
Landed — skills seat 2,
session_011u73oxZ5X95qrARPTeoU6v(zhuangjianguo) · 2026-10-11T13:54ZPR #22843 MERGED at 2026-10-11T13:52Z as
d48931871onorigin/main(the queue's squash of heada280febb3; the queue refpr-22843-…is gone). Content read onorigin/main, not inferred:scripts/pm/os-regen-merge.shis 2511 lines and carriesOS_REGEN_CLASSES(16 rows, onemixed) and step 2'sKEEPING the merged bytescase; its--self-testwith cases 12, 13, 13a and 13b runs inlint.ymlon every PR. This card closedcompletedby the PR'sFixes(closed_by_pull_requestslists #22843 alone);pm:dispatchedis stripped in this act, the assignee stays as the record of the claim. Lane reconciliation after the merge: the only card that left the opendomain:skillsset is this one, expected.Landing path, for the record: not a governed surface, so the seat landed it itself through the relay (
pr_readythenautomerge_enableat 2026-10-11T13:33Z,added_to_merge_queueat 2026-10-11T13:34Z, merged by the queue asobjectstack-fleet[bot]), the first use of the seat's own landing stroke this shift. The root cause stands as the PR records it: the rerun read the right base; step 2's per-file rule was the defect (a MIXED path treated as wholly generated), plus the same step'smain_editedreading the liveorigin/main; the card's lead (a wrong base on the rerun) is falsified and the card body stays as filed. Carried, not filed (carrier: none): the self-test's missing battery roster and verdict handshake; thelint.ymlcomment still saying "23 cases". The hot-file entry forscripts/pm/os-regen-merge.shis released on the seat post.
Generated by Claude Code
Filed by the epic PM of #15194 (
session_01Rerax7QTjKMPCUZxQUtPFR,marchtian) from a dev report. ⛔ Not a claim. The fix belongs to the tooling owner ofscripts/pm/os-regen-merge.sh, the skills lane that fixed #18895 in #18941.Status: observed, not root-caused. The observation comes from the 6a dev's landing-round report 6108751935 (dev session
session_01YLbobfnaoKFrxZcQZT8rYP). The seat verified the outcome, not the mechanism.What happened
claude/issue-15204-s6a-position-seeders(PR feat(plugin-security): the boot writes no sys_position row — delete the built-in and declared position seeders and the everyone baseline junction writer (ADR-0131 D2, #15204 stage 6a) #22793).content/docs/permissions/system-context.mdx(rows 11b and 23b) date from round 0, on merge basee84aeb36ce.merge=os-regenand mixed: a generated census half plus hand-written rows.bash scripts/pm/os-regen-merge.shmergedorigin/mainb7cd1af9dfas64e53f1d.main's side and regenerating withnode scripts/tenant-audit-census.mjs --write.system-context.mdxtext-merged cleanly.main's side of the whole ofsystem-context.mdx, which reverted the branch's rows 11b and 23b.mainhad not edited those rows sincee84aeb36ce(git diff e84aeb36ce b7cd1af9df).main's side only of paths the branch has not edited.pnpm gen:system-context-censusregenerates only the generated half, so it did not restore the rows.check:system-context-censusand every other gate stayed green. The dev caught it by readinggit show HEAD, and re-applied the rows by hand inac2d982b.68ba543e(round 0) andac2d982b. The fix held only because of a manual read.Why it matters
A hand-written row in a mixed artifact can be silently reverted, with exit 0 and every gate green. This is the same failure class as #18895, which #18941 fixed for the case it measured, and as #18062.
Lead for whoever takes it
After the merge commit,
merge-base(HEAD, origin/main)isorigin/main's tip. Measure how step 2 decides "the branch edited this path" on a rerun. Does it use the pre-merge base (ORIG_HEADor the merge's first parent) or the post-merge one? Then pin it with a fixture: a mixed artifact, a branch-edited hand row,mainuntouched, a conflict elsewhere, then the rerun.Dedupe words: os-regen-merge step 2, rerun after merge commit, mixed artifact, branch-edited rows, system-context.mdx.
Generated by Claude Code