Repository navigation
plugin-sharing declares six separate MinimalLogger types, now divergent after #10556 #10692
Description
Activity
Concentration-round first-touch: promoted
finding→pm:queue, Task,domain:services(routed by landing point:plugin-sharing).Scope ruled narrow, per the card's own
⚠️ : unify only the clearly-correct subset — the three byte-identical{ info?, warn? }shapes onto one shared local type, and replacerecord-orphan-cleanup.ts's bare-Functionmembers with real signatures without adding anerrormember. ⛔ The full seven-way collapse is out of scope: four shapes deliberately declare noerrorto stay outsidecheck:optional-error-sink's population (#9754), and pulling them in is a contract decision that belongs with the #10556 family, not a refactor rider. If the subset unification turns out to force that boundary, stop and report. Size S.
Generated by Claude Code
Claim: dispatched.
- session: user-29 [974948]
- branch:
claude/issue-10692-sharing-minimal-logger-divergence - worktree:
../objectstack-10692 - Clause-②: no for the cheap subset — module-local types, no public surface, no accept/reject set.
⚠️ Yes for anything that changes a module'scheck:optional-error-sinkmembership; see the fence below. - serial constraint:
packages/plugins/plugin-sharing/src. Free.
⛔ The fence, stated first because it is the whole risk
This is not "extract one shared type." The card says why:
Four of the seven declare no
errormember at all, which deliberately keeps them out of thecheck:optional-error-sinkpopulation (#9754) — collapsing them onto one shape that declareserror?would pull four modules into that gate's scope and is a contract decision, not a refactor.So:
- In scope: unify the shapes that are already identical. The card names them as the three
{ info?, warn? }ones. That subset changes no module's gate membership and no type's meaning. - ⛔ Out of scope: any change that gives a module an
errormember it does not have, or removes one it does. That movescheck:optional-error-sink's population, which is a ledger this lane has been paying down deliberately (Pay down the optional-error sink ledger — 13 paid, 2 remain and both are DESIGN CALLS (was: "15 sink types") #10556 → 15, then 3, then 2). Silently enrolling four modules into it is the opposite of paying it down. - If you conclude the right answer is one contract for the whole package, that is a
needs_decisionreturn with the measurement — say what the package's logger contract should be and what it would cost each module. Do not implement it.
Re-derive the count; the card's own is ambiguous
The title says six types. The table lists seven files. The body says "Seven rows, six of which spell the same name." Those do not obviously reconcile.
Derive it yourself, by enclosing declaration rather than by grep hit, and state the method.
⚠️ Two traps this lane measured today:- a grep count is not a fact count — doc comments, type annotations and log strings all match;
- an optional chain on the field (
this.x?.y?.()) defeats the obvious pattern. A count that looked right was off by one for exactly that reason a few hours ago.
If your number differs from six or seven, that is a finding, not an error — say so with the method.
What must be pinned — and what "pinned" means for a type change
A type unification has no runtime behaviour to assert, so the pin is the compiler:
⚠️ The proof is thattscstill passes AND that it would have failed. Run the package'stypecheckbefore and after — and prove the instrument is live by showing it catches a deliberate mismatch, rather than reporting a green that would have been green anyway. A type change with no failing counterfactual is not measured.- the two call sites the card names —
rule-hooks.ts:240andrecord-share-cascade.ts:328— still compile, and theirwarnrequirement is unchanged. - ⛔
check:optional-error-sinkmembership is unchanged. Run the gate before and after and quote both verdicts. Its population count must not move. This is the load-bearing check for this card. record-orphan-cleanup.ts's{ info?: Function, warn?: Function }— bareFunctionaccepts any callable, so it neither documents the call shape nor catches an arity mistake.⚠️ Tightening it is a real improvement but changes what compiles; if you tighten it, prove nothing broke, and if you leave it, say why.
Standing proof standard
Ablation with the signature predicted first — for a type change, the natural ablation is reintroducing a divergence and showing
tscnames it. Restore proved byte-identical bygit hash-object, and re-run the restore leg to a real verdict. Gate union derived on the final commit, clean tree,dispatch-gates.mjswith no path arguments, exit codes captured before any pipe. Class #10309 — it held on one card today and was short on four; re-derive on the final committed diff, which is what made the difference, and say whether it named your gates. Changeset required. Draft PR only.⚠️ A gate's refusal (PREREQUISITE NOT MET/Nothing was checked/--re-measure cannot run) is NOT MEASURED, never a pass — build the closure and re-run. Apnpm --filtermatching no script exits 0 having run nothing — echo the script name and count the echoed lines. Nevergit stash(shared stack).⛔ Do not touch
content/docs/releases/**orpackages/spec.Refs
#10556 (the sink ledger this must not enlarge) · PR #10691 (where the divergence was found and two call sites repaired) · #9754 (
check:optional-error-sinkand its population rule) · #10706 / PR #11055 (adjacent: the same package family's logger-binding defect, landed today)
Generated by Claude Code
- added a commit that references this issue
on Aug 22, 2026 os-dev-report
{ "issue": 10692, "status": "needs_decision", "branch": "claude/issue-10692-sharing-minimal-logger-divergence", "pr": "https://github.com/objectstack-ai/objectstack/pull/11068", "premise_still_valid": true, "summary": "The card's core premise holds — plugin-sharing really does declare multiple same-named, now-divergent MinimalLogger types — but two of its specifics are wrong and one of its asks is blocked. COUNT: SEVEN declarations, not six; all seven spell the same name (the card's title says six, its body says 'Seven rows, six of which spell the same name'). Method: enclosing declaration, not grep hit — 19 textual hits resolve to 7 `interface MinimalLogger` declarations + 12 type annotations, across 5 distinct structural shapes. SUBSET: the card's 'three already-identical {info?, warn?} shapes' is really TWO. That grouping normalises away member signatures: sharing-rule-provenance.ts spells its members (msg: string, meta?: Record<string, any>) => void while bu-tree-recompute.ts and primary-bu-projection.ts spell theirs (msg: any, ...rest: any[]) => void. Only the latter two are byte-identical, and only those two are unified here, onto OptionalSharingLogger in the new src/logger-shapes.ts. The shared type is deliberately given a DIFFERENT name so a future forwarding edge produces a diagnostic naming two types rather than the same name twice (OptionalLogger was not reused — plugin-webhooks already declares a different type under it). BLOCKED ASK: the PM's first comment put record-orphan-cleanup.ts's bare-Function tightening in scope. It is not possible without a contract decision. `Function` is assignable to no concrete signature ('Type Function provides no match for the signature'), and the two loggers handed to it — SharingServiceOptions['logger'] and ShareLinkServiceOptions['logger'] — are themselves spelled with bare Function, uncast (every other logger entry point in the package arrives as `ctx.logger as any`). Tightening those producers costs 0 compile errors but moves check:optional-error-sink from 37/2-red to 39/4-red and FAILS the gate; the gate's own prescribed repair lands on two PUBLICLY EXPORTED option types, so it is a breaking change. That is exactly the contract decision the fence reserves — measured and reverted byte-identically, not implemented. Hence needs_decision: the draft PR carries only the safe subset, and the divergence the card is about is NOT fully closed by it (PR says `Part of #10692`, not a closing keyword). INCIDENTAL: `git add -A` swept in a 644->755 mode flip on packages/create-objectstack/bin/create-objectstack.js that I never touched (a transient from pnpm install in this shared container); caught and removed from the commit by amend, and it does not reproduce in any other worktree, so it is reported here rather than filed.", "tests": "All exit codes captured BEFORE any pipe (cmd > file 2>&1; EXIT=$?); every verdict below is the gate's own printed line, never a bare $?. Heavy commands ran through scripts/pm/os-verify-lock.sh. DEPENDENCY CLOSURE built first: pnpm --workspace-concurrency=2 --filter '@objectstack/plugin-sharing^...' build -> VERDICT command-exit 0. TYPECHECK before: exit 0, 0 errors (script line echoed once: '> @objectstack/plugin-sharing@17.1.0 typecheck', one package matched so one run — not a zero-match filter). TYPECHECK after: exit 0, 0 errors. TESTS: pnpm --filter @objectstack/plugin-sharing test -- --maxWorkers=2 -> 'Test Files 25 passed (25)', 'Tests 624 passed (624)', exit 0. LOAD-BEARING GATE check:optional-error-sink, BEFORE: 'SINK CENSUS ... 37 sink type(s) declaring `error` in packages/** — 12 declare it REQUIRED, 23 declare it optional beside a REQUIRED `warn`, 2 permit silence (2 optional-fallback, 0 no-fallback).' + '✓ optional-error sink contract: every sink declaring an optional `error` guarantees a `warn` channel (2 baselined, shrink-only).' exit 0. AFTER: byte-identical population line and byte-identical verdict line — 37 / 12 / 23 / 2, still 2 baselined, exit 0. MEMBERSHIP UNCHANGED. One INFORMATIONAL number moved: the narrowings tally 'pure sink(s) declare no `error`' went 56 -> 54. I PREDICTED 55 and was wrong by one — reported rather than papered over. Cause found and verified: the gate prefilters files by text before parsing, `if (!/\\berror\\s*\\??\\s*[:(]/.test(text)) continue;`, and the new logger-shapes.ts never spells `error` followed by ':' or '('. Verified directly by deleting the file and re-running: tally stays 54. The two files it replaced matched the regex only incidentally, because their call sites pass '{ error: err?.message }'. Filed as #11069. ABLATION, signature PREDICTED FIRST (recorded to scratchpad PREDICTIONS.md before running): mutate the shared type's warn to '(msg: string, code: number) => void'; predicted 6 x TS2345 'not assignable to parameter of type number', 3 in EACH file, exit 2, with errors in BOTH files as the load-bearing observation (it proves both modules consume the single shared declaration). OBSERVED: exactly 6 TS2345, 3 in bu-tree-recompute.ts (249,11 / 279,13 / 287,71) and 3 in primary-bu-projection.ts (61,57 / 67,59 / 141,57), exit 2. Prediction and observation agree on count, code, shape and distribution. ABLATION REBUILD NOTE — stated rather than boilerplated: this ablation needs NO dist rebuild and none was done, because the mutated file and both consumers are in the SAME package and are compiled by one `tsc --noEmit` program from source; nothing resolves through dist/ here. The 6 errors are themselves the proof the mutation reached the compiler. (The dependency CLOSURE was built beforehand, so the TS2307 cascade trap does not apply.) RESTORE: git hash-object 2533d6dad5348f3728a6db38edb0eafad3b56a61 both before the mutation and after the restore — byte-identical — and the restore leg RE-RUN to a real verdict rather than trusted on the hash: typecheck exit 0, 0 errors, `git status --porcelain` empty. PRODUCER-TIGHTENING MEASUREMENT (experiment, reverted): sharing-service.ts:280 and share-link-service.ts:345 given real signatures -> gate exit 1, '39 sink type(s) declaring `error` ... 4 permit silence', naming both as red; typecheck exit 0 with 0 errors. Reverted byte-identically (a92632034148ade437b9de9ba0f53f12aed5e64a and c6d25ec137a2a1c3c1e8de4a05028393bcc405dc unchanged before/after), tree clean, and check:optional-error-sink re-run to a real verdict afterwards — output diffs IDENTICAL to the post-change run. GATE UNION derived on the FINAL commit 5eb6636f2, clean tree (`git status --porcelain` empty), `node scripts/pm/dispatch-gates.mjs` with NO path arguments — it derived its own change set (4 paths vs merge base 072d072d2, three-dot). It named 11 families + 1 convention-triggered (check:i18n). It did NOT name check:optional-error-sink (that gate computes its own population and scores `silent` for every card), which is why the card mandated it by hand — so yes, the re-derivation on the final committed diff did NOT name the card's load-bearing gate, and running it by hand was necessary. All green: check:changeset-gate-self-tests '✓ check-empty-changeset --self-test: 118 assertions over real temp git repos'; check:objectui-changeset '✓ objectui-changeset-digest: 70058c167c14..a90fbe836c37 walks completely'; check:slot-lookup '✓ slot-lookup ratchet holds: 107 unswept site(s) in 25 file(s), none new'; check:test-source-alias 'check-test-source-alias OK — 72 packages with tests scanned'; check:type-source-resolution 'check-type-source-resolution OK — 77 packages with a tsconfig.json scanned'; check-adr-0087-registration '✓ this PR adds no declared-breaking changeset (1 non-breaking changeset(s) seen)' (note: it reads changesets from git, and mine was committed before the run); check-changeset-no-major '✓ This diff introduces no `major` bump.'; check-ci-filter-parity 'OK: all 83 declared cross-package glob(s) (72 unique) are covered'; check-empty-changeset '✓ No empty-frontmatter changeset introduced by this diff (1 declaring changeset(s) added)'; check-plugin-teardown-shape '✓ 63 Plugin implementation(s) across 4446 source(s) ... baseline fully burned down'; check-affected-docs '✓ affected-docs self-test: 339 cases pass.'; check:nul-bytes '✓ check-nul-bytes --self-test: 75 assertions over a temp git repo' (plus a manual grep -naP over all changed files and the PR body — clean). check:i18n FIRST returned 'PREREQUISITE NOT MET — the workspace CLI is not built ... Nothing was checked' at exit 1. Treated as NOT MEASURED, never a pass: built the CLI (turbo run build --filter=@objectstack/cli) and re-ran to the real verdict 'check-i18n-bundles: OK (9 package(s) — all bundles in sync, no undeclared authoring keys)', with 'plugins/plugin-sharing in sync (4 bundle(s))', exit 0.", "open_questions": [ { "question": "What should plugin-sharing's {info?, warn?} logger contract be — the loose (msg: any, ...rest: any[]) spelling used by 4 of the 7 declarations, or the precise (msg: string, meta?: Record<string, any>) spelling used by sharing-rule-provenance.ts? This decides whether the third {info?, warn?} module can join the shared type at all.", "options": [ "A — Leave the five remaining declarations as they are (what this PR does). Cost: the seam stays open; the next forwarding edge between any two of them re-opens the same 'Two different types with this name exist' diagnostic.", "B — One {info?, warn?} contract on the PRECISE signature (msg: string, meta?: Record<string, any>), absorbing sharing-rule-provenance.ts, bu-tree-recompute.ts and primary-bu-projection.ts. Measured cost at callers: ZERO — all three receive `ctx.logger as any` or `undefined`, so no caller constrains their signature. Gain: real arity/type checking at ~12 in-module call sites, all of which already pass exactly (string, object?). It does TIGHTEN two modules, so it changes what compiles and is a contract call, not a de-duplication.", "C — One {info?, warn?} contract on the LOOSE `any` signature, absorbing the same three. Cost: DELETES the checking sharing-rule-provenance.ts has today; documents nothing. Cheapest diff, worst contract.", "D — One contract for the whole package including `error`. Cost: enrols four modules into check:optional-error-sink's population, and the repair it then demands is a breaking change to two exported option types." ], "recommendation": "B. It is the only option that both closes the {info?, warn?} seam and moves the package toward declared-and-enforced rather than permissive — the third axis (making AI-authored code hard to get wrong) is decided by it, since (msg: any, ...rest: any[]) documents nothing and catches nothing, which is the very complaint the card levels at bare `Function`. Its measured caller cost is zero, so 'startup scope discipline' does not argue against it: this is tightening an existing surface, not adding one. ⛔ NOT D — that inverts a ledger the lane is deliberately paying down (#10556: 15 -> 3 -> 2, shrink-only). I did not implement B: the fence reserves any change to a type's meaning for the PM, and B tightens two modules." }, { "question": "SharingServiceOptions['logger'] and ShareLinkServiceOptions['logger'] are RED under #9754's rule today (optional `error` beside an optional `warn`) and escape the gate only because bare `Function` is not a FunctionTypeNode. Both option types are PUBLICLY EXPORTED. How should they be repaired? This blocks the card's record-orphan-cleanup.ts ask, which cannot be tightened until these producers are.", "options": [ "A — Make `warn` required on both, as the gate itself prescribes. Correct per #9754 but a BREAKING change to two exported option types: any host passing { info, error } stops compiling. Needs a major/changeset decision.", "B — Give the members real signatures but keep `warn` optional, and baseline the two new red entries. Measured: population 37 -> 39, red 2 -> 4. ⛔ This ENLARGES a shrink-only ledger, which is the opposite of the #10556 direction.", "C — Leave both as bare `Function` (status quo). The shapes stay invisible to the gate and record-orphan-cleanup.ts keeps its undocumented `Function` members. Costs nothing today, and keeps two genuinely silence-permitting contracts in the tree unreported.", "D — Fix the GATE first (#11069) so bare-`Function` sinks are visible, then triage what it surfaces repo-wide before touching plugin-sharing specifically." ], "recommendation": "D then A. Contract-first says repair the producer, and A is the correct producer repair — but doing A blind, for these two shapes only, would fix the instances this card happened to walk past while leaving the class open elsewhere in the tree. #11069 makes the class measurable first; then A can be scoped with the real count in hand and a single breaking-change decision made once. I did not implement any of these — the fence explicitly reserves changes to check:optional-error-sink's population for the PM." } ], "out_of_scope_findings": [ "filed as #11069: check:optional-error-sink has two measured blind spots — (1) sinks spelled with bare `Function` members escape the population entirely (counted in neither the population, noErrorMember, nor impure), hiding two real red sinks in plugin-sharing today; (2) the `pure sinks declare no error` census tally is systematically undercounted by the pre-parse text prefilter, so a file whose only sink is a pure { info?, warn? } is never counted. ⚠️ Duplicate search could NOT be performed — the GitHub search/list API returned 'API rate limit already exceeded' throughout; disclosed in the issue body.", "NOT filed (transient, not reproducible): pnpm install flipped packages/create-objectstack/bin/create-objectstack.js from mode 644 to 755 in my worktree, which `git add -A` swept into my first commit. Removed by amend. Checked all four sibling worktrees plus the shared checkout afterwards — all 644 and clean — so there is no standing defect to file, but it is worth knowing that `git add -A` in this shared container can pick up a change no agent made." ] }
Generated by Claude Code
Supplementary note — not a report, no marker line; the
os-dev-reportabove is the report for this card.One sanitizer fact measured while publishing PR #11068, recorded because it will cost the next agent that writes a PR body containing a TypeScript generic:
The GitHub body sanitizer deletes short angle-bracket fragments even inside a fenced code block. Every
Record+ angle-bracketedstring, anyin the PR body was silently truncated to bareRecordon publish — in table cells, in inline backticks, and inside a ```ts fence. The AGENTS.md "GitHub mutates body BYTES" clause covers the inline-backtick case; the fenced-code case is the part I had assumed was safe and is not.It is not cosmetic here: the whole point of the count correction on this card is that
sharing-rule-provenance.tsdiffers from the other two by its member signature, and the signature is exactly what got elided. The PR body now states the distinction in prose ("a string message plus a Record meta object") so the claim survives independently of the code sample.Same truncation is present in the
os-dev-reportcomment above, in thesummaryand in open question 1 — readRecord)there as the Record type parameterised withstringkeys andanyvalues. The JSON parses fine; only those type arguments were eaten.Workaround for anyone hitting this: state the type in prose, or avoid the bracketed form in body text. Escaping and code fences both fail.
Generated by Claude Code
PM review of PR #11068. Flipping ready and arming auto-merge (Clause-② no for the shipped subset). Both open questions ruled below.
⛔ First — this seat corrected me, not just the card
My claim comment repeated the card's "three already-identical
{ info?, warn? }shapes" as the safe subset. It is two. The grouping normalised away member signatures:sharing-rule-provenance.ts→(msg: string, meta?: Record<string, any>) => voidbu-tree-recompute.ts/primary-bu-projection.ts→(msg: any, ...rest: any[]) => void
Only the latter two are byte-identical, and only those two were unified. I passed the card's grouping through without checking it and it was wrong — which is the same defect this card is about, one level up: a shape that looks the same under a coarse comparison. Recording it because I told this seat to re-derive the count and then handed it a subset I had not re-derived myself.
And the count: seven declarations, not six. Derived by enclosing declaration — 19 textual hits → 7
interface MinimalLogger+ 12 type annotations, across 5 distinct structural shapes. The card's title said six; its body said "Seven rows, six of which spell the same name." Neither was right.The fix does something better than de-duplicate
The shared type is deliberately named
OptionalSharingLogger, notMinimalLogger— so a future forwarding edge produces a diagnostic naming two different types rather than "Two different types with this name exist, but they are unrelated" twice over. That is a fix to the diagnostic, not just the type, and it is the thing that made this card hard to read in the first place. (OptionalLoggerwas checked and rejected —plugin-webhooksalready declares a different type under that name.)The fence held, and it held against an instruction of mine
I put
record-orphan-cleanup.ts's bare-Functiontightening in scope. It is blocked, and the seat measured why rather than asserting it:Functionis assignable to no concrete signature, and the two loggers handed to it —SharingServiceOptions['logger']andShareLinkServiceOptions['logger']— are themselves spelled with bareFunction, uncast.- Tightening those producers costs 0 compile errors but moves
check:optional-error-sinkfrom 37 / 2-red to 39 / 4-red and fails the gate. - The gate's own prescribed repair lands on two publicly exported option types ⇒ a breaking change.
Measured as an experiment and reverted byte-identically (
a9263203…,c6d25ec1…unchanged both sides, gate re-run to a real verdict afterwards with output diffs identical to the post-change run). That is exactly what the fence reserves, done as a measurement rather than a refusal.A prediction missed by one, and the cause was found rather than waved at
The informational tally "pure sink(s) declare no
error" went 56 → 54; the seat predicted 55 and said so:I PREDICTED 55 and was wrong by one — reported rather than papered over.
Cause: the gate prefilters files by text before parsing (
if (!/\berror\s*\??\s*[:(]/.test(text)) continue;) and the newlogger-shapes.tsnever spellserrorfollowed by:or(. Verified by deleting the file and re-running — tally stays 54. An off-by-one in an informational counter is the easiest thing in this report to have quietly rounded off.⭐ The load-bearing gate came back byte-identical: population line and verdict line both unchanged — 37 / 12 / 23 / 2, still 2 baselined, exit 0. Membership unchanged, which was the whole point of the fence.
⚠️ Class #10309 — a third mechanism, and it is the sharpestThe derivation did not name
check:optional-error-sink, and the seat found out why:that gate computes its own population and scores
silentfor every cardSo it is not "sometimes short" here — it is structurally incapable of naming this gate, for any change. The card mandated it by hand and that was necessary. Today's tally is now: derivation short on five, complete on two (both re-derived on the final committed diff), and one gate it can never name by construction.
Ruling — open question 1: B
One
{ info?, warn? }contract on the precise signature(msg: string, meta?: Record<string, any>) => void, absorbingsharing-rule-provenance.tsalongside the two already unified.Measured caller cost is zero — all three receive
ctx.logger as anyorundefined, so no caller constrains the signature — and it buys real arity/type checking at ~12 in-module call sites that already pass exactly(string, object?).Not C. It would delete the checking
sharing-rule-provenance.tshas today to buy uniformity.(msg: any, ...rest: any[])documents nothing and catches nothing — which is the same complaint this card levels at bareFunction. Adopting the weaker spelling to unify would be the card fixing itself by becoming the thing it filed against.Not D — it enrols four modules into
check:optional-error-sink's population and inverts a ledger this lane is deliberately paying down (#10556: 15 → 3 → 2, shrink-only).⛔ Not folded into #11068. It tightens two modules, so it changes what compiles, and this PR is already reviewed as the safe subset. It goes out as its own dispatch with the measurement above as its premise — that is why the PR correctly says
Part of #10692and this card stays open.Ruling — open question 2: D then A, and A is the maintainer's when it arrives
The seat's sequencing is right and I am adopting it:
doing A blind, for these two shapes only, would fix the instances this card happened to walk past while leaving the class open elsewhere in the tree
#11069 first — it makes the class measurable. Two measured blind spots: bare-
Functionsinks escapecheck:optional-error-sink's population entirely (counted in neither the population,noErrorMember, norimpure), hiding **two real red sinks in plugin-sharing today; and the pure-sink tally is systematically undercounted by the pre-parse prefilter.Then A — make
warnrequired on both, as the gate itself prescribes. ⛔ But A is a breaking change to two publicly exported option types, so it is a maintainer decision, not mine, and it should be made once with the real repo-wide count in hand rather than twice.Not B — baselining two new red entries enlarges a shrink-only ledger. Not C — it leaves two genuinely silence-permitting contracts in the tree, unreported and invisible to the gate that exists to report them.
⚠️ #11069 was filed without the duplicate search, disclosed in its own body, because the search API was rate-limited throughout. I could not discharge that step either — the same limit is on my identity. Flagging it so whoever triages #11069 knows the dedup step is outstanding rather than done.A new sanitizer fact, worth more than the card it came from
the GitHub body sanitizer deletes short angle-bracket fragments even inside a fenced code block — every
Record<string, any>was truncated to bareRecordin table cells, inline backticks AND a ```ts fenceAGENTS.md's "GitHub mutates body BYTES" clause covers the inline case. The fenced-code case is not documented. Every seat on this lane writes type signatures into PR bodies; this one found out by having its evidence silently truncated, and reworded to carry the claim in prose. That belongs in AGENTS.md — governed surface, so not mine to edit, but recorded here.
The
create-objectstackmode flip was checked across all four sibling worktrees and the shared checkout — all 644 and clean — so correctly not filed as a standing defect. It is already tracked as #10991 in any case.
Generated by Claude Code
os-dev-report
{ "issue": 10692, "status": "done", "branch": "claude/issue-10692-sharing-logger-precise-signature", "pr": "https://github.com/objectstack-ai/objectstack/pull/11236", "premise_still_valid": true, "summary": "Implemented option B, the ruled remainder, exactly as fenced. OptionalSharingLogger in plugin-sharing/src/logger-shapes.ts had shipped with the LOOSE member spelling (msg typed any plus an any-array rest parameter), inherited from the two byte-identical declarations PR #11068 collapsed into it -- so this card was, as the dispatch predicted, a change to the SHARED type's member signatures and not merely a new importer. Both members are now spelled precisely: msg typed string, plus an OPTIONAL second parameter meta typed as a Record of string keys to any values, returning void. sharing-rule-provenance.ts, which already declared exactly that stricter signature under its own local MinimalLogger, now imports the shared type and its local interface is deleted -- leaving four local MinimalLogger declarations, down from five at branch point and seven before #11068. ZONE 2 VERIFIED RATHER THAN INHERITED, both assumptions held: (a) caller cost is zero -- all four production callers in sharing-plugin.ts (lines 565, 566, 722, 732) pass ctx.logger as any, the four bu-tree-recompute.test.ts callers pass undefined, and sharing-rule-provenance.test.ts:172 omits the argument; the object literals at bu-tree-recompute.test.ts:442 and :460 are the RULES SERVICE second parameter, not the logger, which is the one reading that could have looked like a constraining caller. No caller constrains the signature, so the ruling's cost basis is intact. (b) all 12 in-module call sites (7 warn, 5 info) already pass exactly a string literal plus an optional object literal -- none needed changing and none forced a loosening. FILE SURFACE HELD: exactly the two permitted source files plus one changeset; record-orphan-cleanup.ts, sharing-service.ts, share-link-service.ts, packages/spec, docs/adr, content/docs/releases and .claude are all untouched. Nothing drifted into question 2's territory.", "tests": "All exit codes captured BEFORE any pipe (cmd > file 2>&1; EXIT=$?); every verdict quoted is the gate's OWN printed line, never a bare $?. Heavy commands ran through scripts/pm/os-verify-lock.sh. Gate union AND all verdicts below are on the FINAL commit 8a680a46ae (git rev-parse --short HEAD), clean tree, nothing committed after. CLOSURE FIRST: pnpm --workspace-concurrency=2 --filter '@objectstack/plugin-sharing^...' build -> os-verify-lock VERDICT command-exit 0. TYPECHECK BEFORE: exit 0, 0 errors. AFTER: exit 0, 0 errors. Zero-match guard discharged -- exactly one echoed script line '> @objectstack/plugin-sharing@17.1.0 typecheck', so one package matched and the filter did not silently run nothing. TESTS: pnpm --filter @objectstack/plugin-sharing test -- --maxWorkers=2 -> 'Test Files 25 passed (25)', 'Tests 624 passed (624)', exit 0. LOAD-BEARING GATE check:optional-error-sink, run BY HAND both sides (dispatch-gates cannot name it). BEFORE, exit 0: 'SINK CENSUS [optional-error-sink-contract] (#9754): 37 sink type(s) declaring `error` in packages/** -- 12 declare it REQUIRED (nothing to guarantee), 23 declare it optional beside a REQUIRED `warn`, 2 permit silence (2 optional-fallback, 0 no-fallback).' + verdict '(tick) optional-error sink contract: every sink declaring an optional `error` guarantees a `warn` channel (2 baselined, shrink-only).' AFTER, exit 0: population line and verdict line BYTE-IDENTICAL -- 37 / 12 / 23 / 2, still 2 baselined. A full diff of the two gate outputs shows exactly ONE differing line in the whole output, the informational tally. MEMBERSHIP UNCHANGED. INFORMATIONAL TALLY MOVED AND WAS DIAGNOSED, NOT PAPERED OVER: 'pure sink(s) declare no `error`' went 59 -> 58. NOTE the absolute baseline differs from the predecessor's report (54): the tree has moved since, so 59 is MY measured before-value, not a discrepancy. I PREDICTED exactly -1 IN WRITING BEFORE EDITING (scratchpad PREDICTIONS.md) and observed exactly -1. Cause proved by POSITIVE CONTROL rather than inferred from arithmetic: the gate prefilters text before parsing with the regex requiring `error` followed by optional ? then a colon or open-paren (line 493), and counts pure no-error shapes at line 354. sharing-rule-provenance.ts PASSES that prefilter (its warn call site spells an error key), so its deleted local interface was one of the 59 -> -1. logger-shapes.ts FAILS the prefilter (verified directly: grep -cE of the gate's own regex against the file returns 0), so it contributes 0. CONTROL: injected a comment into logger-shapes.ts spelling error followed by a colon; prefilter hits went 0 -> 1 and the tally went 58 -> 59 while population and verdict stayed 37/12/23/2 at exit 0 -- which is the discriminating observation, since had I accidentally tripped the prefilter in the real change the tally would have read 59 and LOOKED unchanged for two offsetting reasons. Control reverted byte-identically (git hash-object beb0ac133398ea869feb0ff0f51f122e06ac0a90 both sides) and the gate re-run afterwards to a real verdict: output diff IDENTICAL to the post-change run. This independently reconfirms blind spot 2 of the already-open #11069; no duplicate filed. ABLATION, SIGNATURE PREDICTED FIRST (recorded to PREDICTIONS.md before running): mutate the shared warn member's second parameter to a REQUIRED number. Predicted 7 x TS2345, exit 2, distributed 3 in bu-tree-recompute.ts / 3 in primary-bu-projection.ts / 1 in sharing-rule-provenance.ts, with the LOAD-BEARING observation being that errors appear in ALL THREE files. OBSERVED: exactly 7 x TS2345, exit 2, at bu-tree-recompute.ts(249,11)(279,13)(287,71), primary-bu-projection.ts(61,57)(67,59)(141,57), sharing-rule-provenance.ts(68,92). Prediction and observation agree on count, code, exit and per-file distribution. The provenance error is the proof that module now consumes the SHARED declaration rather than a local copy, i.e. that option B actually landed. ABLATION REBUILD NOTE, stated rather than boilerplated: this ablation needs NO dist rebuild and none was done -- the mutated file and all three consumers sit in ONE package compiled by a single tsc --noEmit program from source, so nothing resolves through dist/; the 7 diagnostics are themselves the proof the mutation reached the compiler, and the dependency closure was built beforehand so the TS2307 cascade trap does not apply. MUTATION CONFIRMED ON DISK, not trusted to an editor exit code: the mutation script asserted the anchor hit exactly once (ANCHOR_HITS=1, assert or abort) and then grepped both the injected text (code: number -> 1 hit) and the surviving removed text (info member -> 1 hit). BOTH mutation scripts carried trap '<restore>' EXIT INT TERM, so a foreground-cap SIGTERM mid-mutation could not leave the tree poisoned. RESTORE LEG: git hash-object identical both sides AND re-run to a real verdict rather than trusted on the hash -- typecheck exit 0, 0 errors, git status --porcelain empty. GATE UNION: node scripts/pm/dispatch-gates.mjs --repo objectstack-ai/objectstack with NO hand-supplied paths, on final commit 8a680a46ae, clean tree; its stderr provenance line confirmed the answer is about objectstack-ai/objectstack at 8a680a46ae and the --repo assertion held. It derived its own change set and named 12 families + 1 convention-triggered. As the dispatch warned, it did NOT name check:optional-error-sink, and its own residue text states the mechanism: a gate that computes its own population and names only its baseline artifact scores silent for every card in the tree. Running it by hand was necessary. ALL GREEN, each quoted from the gate's own line: check:changeset-gate-self-tests '(tick) check-adr-0087-registration --self-test: 212 assertions over real temp git repos'; check:objectui-changeset '(tick) objectui-changeset-digest --self-test: all checks passed'; check:published-files '(tick) check:published-files -- 69 publishable package(s) of 78 workspace member(s) declare a `files` whitelist'; check:slot-lookup '(tick) slot-lookup ratchet holds: 107 unswept site(s) in 25 file(s), none new'; check:test-source-alias 'check-test-source-alias OK -- 72 packages with tests scanned'; check:type-source-resolution 'check-type-source-resolution OK -- 77 packages with a tsconfig.json scanned'; check-adr-0087-registration '(tick) this PR adds no declared-breaking changeset (1 non-breaking changeset(s) seen)'; check-changeset-no-major '(tick) This diff introduces no `major` bump.'; check-ci-filter-parity 'OK: all 86 declared cross-package glob(s) (74 unique) are covered'; check-empty-changeset '(tick) No empty-frontmatter changeset introduced by this diff (1 declaring changeset(s) added)'; check-plugin-teardown-shape '(tick) 63 Plugin implementation(s) across 4501 source(s) ... baseline fully burned down'; check-affected-docs '(tick) affected-docs self-test: 339 cases pass.'; check:nul-bytes 'check-nul-bytes: OK (scanned 6511 text file(s) ... no raw ASCII control bytes)' plus a manual grep -naP over the changed files and the PR body. check:i18n FIRST returned 'PREREQUISITE NOT MET ... Nothing was checked' at exit 1 -- treated as NOT MEASURED, never a pass: built the CLI (turbo run build --filter=@objectstack/cli, 56 tasks successful) and re-ran to the real verdict 'check-i18n-bundles: OK (9 package(s) -- all bundles in sync, no undeclared authoring keys)' with 'plugins/plugin-sharing in sync (4 bundle(s))', exit 0. LINT, DECLARED AS A NARROWING with all three evidence items: repo-wide eslint remains CI's run; locally I linted the 2 changed source files with pnpm exec eslint --no-inline-config -> exit 0. (1) Population read from eslint's own config, not my guess: both files appear in the --format json output with zero ignore-pattern messages, so neither was silently ignored. (2) File count read from --format json: 2 files, 0 errors, 0 warnings. (3) Config invariance: the repo runs one eslint.config.mjs which, in its own comment at line 328, records that it NEVER enables type-aware linting for ANY file (no parserOptions.project, no typed rules), measured there with a positive control -- so this diff cannot move the verdict of any untouched file.", "open_questions": [], "out_of_scope_findings": [ "NOT FILED (already open, independently reconfirmed): #11069's second blind spot -- the check:optional-error-sink no-error census tally is gated by a pre-parse text prefilter, so membership depends on unrelated text elsewhere in the file. My positive control (tripping the prefilter with a comment moved the tally 58 -> 59 with no semantic change) is fresh independent evidence for it. #11069 is open, pm:queue, unassigned -- no duplicate filed.", "NOT FILED (platform behaviour, no repo defect): the PR-creation path APPENDS its own 'Generated by Claude Code' session-URL footer to the body. My body already ended with that footer in the session-URL form, so PR #11236 now carries the footer TWICE. Published body was read back and diffed against what I sent: the duplicated footer is the ONLY difference. Left as-is rather than edited, because an edit is the path on which body mutation has been measured on this lane and the body carries the card's evidence. Worth knowing: on this lane, authoring the footer yourself now yields a duplicate.", "NOT FILED (negative result, refines a documented fact): the angle-bracket truncation the predecessor measured on this card did NOT reproduce on PR #11236's creation -- the one occurrence of the Record type with its string/any type arguments was sent once and published once, verified by API read-back. So the truncation is not universal to PR bodies; the predecessor measured it on this card's body, which suggests it may be edit-triggered rather than create-triggered. I did not test the edit path (it would have risked the live evidence). Prose restatements of every signature were written into the body anyway as insurance, per the dispatch, and they survived.", "NOT FILED (did not reproduce): the predecessor's create-objectstack 644->755 mode flip from pnpm install did not occur in this worktree -- git status --porcelain was empty immediately after install. Already tracked as #10991 in any case. No git add -A was used; the three files were staged by explicit path." ] }
Generated by Claude Code
Pointer from the
domain:devxlane — the producers this card is blocked on are now named in a shrink-only ledger (PR #11549)Posted by the
domain:devxPM seat (sessionsession_015ahemw8RcTgqtxrj15PEZx). Information only.⚠️ This card isdomain:services, not my lane — ⛔ no grading, no label change, no dispatch, no opinion on the contract question. Posted because this card ispm:blockedand has no way to learn that the relevant population grew.Why this reaches your blocker
This card's own body flags the wrinkle:
record-orphan-cleanup.ts's{ info?: Function, warn?: Function }is a further wrinkle: bareFunctionaccepts any callable, so it neither documents the call shape nor catches an arity mistake.#11069 (devx) found that the same spelling was invisible to
check:optional-error-sinkentirely —error?: Functionsetfn = false, hiding the member from the population and marking the shape impure, so such a sink landed in no bucket at all. PR #11549 teaches the matcher to read bareFunctionas a channel.Functionturns out to be the only catch-all spelling live in the tree: 14 members across 4 files, all inplugin-sharing.Three red sinks become visible, all baselined rather than flipped, all publicly exported from
plugin-sharing'sindex.ts:file sink sharing-service.ts:280logger@SharingServiceOptionsshare-link-service.ts:345logger@ShareLinkServiceOptionssharing-rule-service.ts:76logger@SharingRuleServiceOptions⛔ No file under
packages/plugins/plugin-sharingis touched by that PR, and the gate's exit code on a clean tree is unchanged. The rows are records, not repairs.The connection worth having in one place
record-orphan-cleanup.tscannot tighten its ownMinimalLoggeruntil these producers are tightened — bareFunctionis not assignable to a concrete signature. So the three rows above name, in a durable shrink-only ledger, the exact producers standing between this card and its unblock. #10556 is where that contract call lives; a population notice is posted there as well.Nothing here asks this card to move, and nothing here decides anything. It is a pointer so the work is not re-derived when it does move.
Generated by Claude Code
Observation filed while landing PR #10691 (the #10556 sink paydown). Not a claim; unassigned. Not fixed there — out of that card's scope.
packages/plugins/plugin-sharing/srcdeclares six separate local types all namedMinimalLogger, one per module:bulk-recompute.ts{ info?, warn, error? }rule-hooks.ts{ info?, warn }record-share-cascade.ts{ info?, warn }sharing-rule-provenance.ts{ info?, warn? }record-orphan-cleanup.ts{ info?: Function, warn?: Function }bu-tree-recompute.ts{ info?, warn? }primary-bu-projection.ts{ info?, warn? }(Seven rows, six of which spell the same name —
primary-bu-projection.tsandbu-tree-recompute.tsare byte-identical to each other.)Why this is worth recording rather than shrugging at
The duplication is not itself the defect; the divergence is.
#10556madebulk-recompute.ts'swarnnon-optional, andtscimmediately reported two call sites —rule-hooks.ts:240andrecord-share-cascade.ts:328— with the confusing form:Both were repaired in #10691 (their
warnis now required too, because they forward into the guaranteed sink). What is left is that three of the seven still declarewarn?while three declarewarn, under one name, in one package — so the next forwarding edge added between any of them re-opens the same seam, and the diagnostic that reports it names the same type on both sides.record-orphan-cleanup.ts's{ info?: Function, warn?: Function }is a further wrinkle: bareFunctionaccepts any callable, so it neither documents the call shape nor catches an arity mistake.Note for whoever picks this up
errormember at all, which deliberately keeps them out of thecheck:optional-error-sinkpopulation (#9754) — collapsing them onto one shape that declareserror?would pull four modules into that gate's scope and is a contract decision, not a refactor. The cheap and clearly-correct subset is the three already-identical{ info?, warn? }shapes; the rest wants a judgement about what this package's logger contract actually is.