check-durability-degradation-log-level: loggerLevel cannot see the (logger.error ?? logger.warn)(…) fallback, so a loud catch reads as silent-swallow — and the spelling it CAN see prints nothing

Sub-issue of #8897, which records the receiver-name half of this same function's narrowness. This is the call-shape half. Filed unassigned. Not a claim.

Found while implementing #9609, which adds runWideningAlters to DURABILITY_CRITICAL_CALLEES. Unlike #8897's finding, this one is not latent — it produced a false red on real, loud code, and it is measured rather than reasoned.

Note for triage: #8897's own restart-when trigger reads "any PR touches scripts/check-durability-degradation-log-level.mjs" — the #9609 PR does.

The blind spot

loggerLevel() requires the call expression's callee to be a property access:

const expr = node.expression;
if (!ts.isPropertyAccessExpression(expr) || !ts.isIdentifier(expr.name)) return undefined;

(this.logger.error ?? this.logger.warn)(msg, meta) is a call on a parenthesized expression, so the guard returns undefined and the catch collects no levels at all. The gate then reports it as catch swallows the failure with no log at all — the harshest verdict in the file, on a site that is loud at runtime.

Measured, with the reproduction

On the #9609 branch, with runWideningAlters in the vocabulary and both widening catches calling the fallback inline:

✗ 2 durability-critical catch(es) degrade quietly
  packages/drivers/driver-sql/src/sql-driver.ts:8216
    found   : catch swallows the failure with no log at all

Routing the same two calls through a named same-file helper whose body is if (this.logger.error) this.logger.error(...) else this.logger.warn(...) turns them green and correctly classified:

✓ durability-degradation log levels: 29 durability-critical catch seam(s), all loud …
  packages/drivers/driver-sql/src/sql-driver.ts:8255 → recovers on one branch, loud (error@4021 via logDurabilityFailure())

Why this direction of error is the dangerous one

⛔ The spelling the matcher does accept is this.logger.error?.(…) — an optional call, whose callee is a plain property access. It is by far the cheapest way to turn this red green, and against a sink that has no error it prints nothing at all. SqlDriver.logger declares error optional by design, and hosts do inject { warn } sinks — the #9609 test fixture was one until that PR changed it.

So the gate's cheapest satisfaction here converts a loud degradation into a genuinely silent one. That is the same "a gate whose cheapest satisfaction is harmful has the wrong shape" argument the file's own FAILURE_PROPAGATION_CALLEES header makes about bolting on a logger.error. Measured: with the helper body replaced by this.logger.error?.(msg, meta), the gate stays green while the no-error-sink test goes red.

The population

7 call sites use the fallback idiom today, none currently reachable by the vocabulary, so nothing is red on main:

packages/drivers/driver-sql/src/sql-driver.ts — 4 sites (the ADR-0120 D4 NULL-safe unique refusal and its restore, plus two more)
packages/drivers/driver-turso/src/turso-driver.ts:396 — (this.logger.error ?? this.logger.warn).call(this.logger, message), a third shape again
2 sites on the driver-sql: the boot widening's swallowed failure logs warn, but AGENTS.md's degradation rule names DDL-that-did-not-run as error #9609 branch, now routed through the helper

The idiom exists because logger.error is optional on these driver classes. It is the correct way to write the call; the matcher is what cannot read it.

Options, none free

Leave it, note it in the header next to the file's other honest limitations, and let the same-file-helper indirection be the documented answer — which is what driver-sql: the boot widening's swallowed failure logs warn, but AGENTS.md's degradation rule names DDL-that-did-not-run as error #9609 did, and which check-durability-degradation-log-level: collectLoggedLevels only recognises a logger named logger/log/console, so a catch that reports through an injected logger reads as silent to BOTH rules #8897's author gave a weak preference to for the sibling case.
Teach loggerLevel the fallback chain: for a call whose callee is a parenthesized ??/||, collect the level of every branch. (error ?? warn) then contributes both, and the existing levels.filter(LOUD) decides — no new policy, since the file already passes a catch containing both a warn and an error. Does not cover turso-driver.ts's .call(…) spelling.
Declare the fallback as a propagation-style vocabulary entry, which is a third declared list to keep from going stale.

Weak preference for (2): it introduces no new declared names, reuses the file's existing "a loud level anywhere in the catch counts" semantic, and removes the pressure toward the harmful ?. spelling. But it IS a loosening of a gate matcher and should be measured against the whole scan population before landing — the file's own 收窄先行 discipline.

⛔ Deliberately not resolved inside the #9609 PR: choosing among these is a policy call on a gate's classification surface, and #9609's contained answer (a named helper the gate already follows) needed none of it.

check-durability-degradation-log-level: collectLoggedLevels only recognises a logger named logger/log/console, so a catch that reports through an injected logger reads as silent to BOTH rules #8897 — parent; the receiver-name half of loggerLevel's narrowness.
check-durability-degradation-log-level: the read-seam invention rule keys on return, so a catch that degrades by FALLING THROUGH into an empty accumulator is structurally invisible #8845 — the read-seam rule's return-keying blind spot; same meta-shape, third recorded instance there.
check:durability-log-level 结构性看不见「读接缝把故障答成空值」这一类 —— #4825 / #5108 全家都在闸门盲区里 #5186 / finding(devx): #5186 读接缝规则的「编造空答案」词表看不见「原样返回入参」这一形状 —— #6116 是实证样本 #6451 — the read-seam rule and its prior measured extension.

Claiming this card.

session: session_01XqDQYVU5smx29ts9pAErja
branch: claude/issue-9657-logger-level-matcher
worktree: /home/user/objectstack-9657 (from origin/main @ ed4ca5999, which contains the blocker 3b3f67d31 / PR fix(driver-sql): report an un-run MySQL widening ALTER at error, naming the fix #9665)

File surface I expect to touch: scripts/check-durability-degradation-log-level.mjs (the loggerLevel() matcher and its self-test harness). Nothing under packages/drivers/** unless the census argues for it — per the dispatch ruling, the 7 existing call sites are not red today and will not be swept.

Order of work per the dispatch: H1 census of call shapes first (design follows the census, not the other way round), H2 verify the logDurabilityFailure helper indirection really survives the matcher on main, then a chosen option, then the H3 demonstration of what happens to the logger.error?.(…) harmful gradient. Read #9165's PM disposition (Q4, the six-instance family of "recognizer narrower than the repo's shapes, shortfall reported as a verdict rather than as unrecognised") before deciding.

Generated by Claude Code

{
  "issue": 9657,
  "status": "done",
  "branch": "claude/issue-9657-logger-level-matcher",
  "pr": "https://github.com/objectstack-ai/objectstack/pull/9750",
  "premise_still_valid": true,
  "summary": "H1 FIRST — the census says SIX shape families, not three, so a contained matcher extension was the wrong shape on its own. Across packages/** (non-test .ts): 3,308 log-emitting calls, 658 inside a catch. Visible to the old matcher: `logger.error(...)` and its `?.` variants (3266 / 645 in-catch). Invisible: `(l.error ?? l.warn)(...)` 6/2; `(l.error ?? l.warn).call(l, ...)` 1/0 (driver-turso); `((c.warn ?? c.error))?.(...)` 1/0; a `.bind()` fallback stored in a const and called through the local 8/6 — a FOURTH family the card did not name, in db-job-adapter and two trigger packages, two of whose catches reason explicitly about the durability class in comments; bare `warn(...)`/`info(...)`/`log(...)` helpers 26/5. OPTION TAKEN: option 2 EXTENDED — resolve the callee structurally (parens, ??/||, ternary, .call/.apply/.bind, non-null, logger['error'], same-file const alias) AND stop accepting the one spelling that prints nothing. Option 1 was falsified by the census (five spellings by authors with no reason to know a gate exists; documenting a blind spot does not stop the sixth); option 3 adds a third staleness-prone vocabulary for nothing option 2 lacks. HARMFUL GRADIENT CLOSED: an optional CALL `?.(` no longer counts toward `loud` — that reads the author's own optionality marker rather than guessing, and the distinction is exact: the fallback of `??` is another log, the fallback of `?.` is silence. Receiver optionality (`logger?.error(...)`) is deliberately NOT judged — no alternative sink exists, and judging it would have flagged 22 more in-catch calls with no remedy. Also, per H1's own instruction at that count, 'I could not recognise this callee' is now the distinct verdict `unreadable-report` instead of being folded into `silent-swallow`; it fails the gate but accuses the right thing and its remedy text rules `?.` OUT rather than inviting it. ⭐ TWO THINGS THE CARD DID NOT ANTICIPATE. (1) The tightening found SIX REAL SEAMS, not collateral: plugin-audit x3, plugin-email, plugin-security x2, every one on a sink whose `error` is DECLARED OPTIONAL (SweepLogger, ProjectionLogger — whose own comment calls `error` the 'durability-degradation channel'), i.e. genuine instances of the defect this card is about. Repaired locally in the same defect class using #9665's landed shape (reach for error, fall back to warn, never to silence). (2) PART OF THIS IS A TYPE-DECLARATION DEFECT, NOT A MATCHER DEFECT — `AuthEventAuditLogger` declared `error?` and `debug?` and NO `warn` at all, so at that site there was no fallback channel to reach for and tsc said so; its sibling `ReadAuditLogger` in the same package has always declared `warn?`. An optional `error` with no declared alternative is a CONTRACT THAT PERMITS SILENCE, and the sink type is a fourth axis for the #9165 family census alongside the call shape. Added `warn?` mirroring the sibling — purely additive. The read-seam rule is passed no `unreadable` sink, so its census and verdicts are byte-identical to main.",
  "tests": "All anchored to the final commit 3a6a4b20a (merged origin/main @ 6cb88d9f2 mid-run after main touched audit-writers.ts; census re-measured on the merged tree and the header table updated to match). GATE: `node scripts/check-durability-degradation-log-level.mjs` => '✓ 29 durability-critical catch seam(s), all loud, rethrowing or propagating (5 propagating, declared)' + '✓ 66 read seam(s), none invents an unreported answer (7 type-discriminated) (1 pass an input through, reported) (1 baselined)' — the read-seam line is byte-identical to main's. SELF-TEST: '✓ 51 case(s) passed' (16 new) + '✓ 35 case(s) passed'. Every new PASSING case pins expectSeams:1 so it cannot pass vacuously; every new FLAGGING case pins expectKinds — a boolean 'did it flag' cannot separate a right verdict from a wrong one, which is this card's whole subject, so the harness gained expectKinds. H3 RUNTIME: against a sink `{warn}` with no error — `l.error?.(msg)` => '(NOTHING EMITTED)'; `(l.error ?? l.warn)(msg)`, `.call(l,msg)`, and `if (l.error) ... else l.warn(...)` => 'warn: durability failure'. H3 GATE, two-direction ablation on the real repo, main's matcher (extracted via `git show origin/main:` and run from scripts/) vs mine: (a) #9665's two sites written INLINE as the fallback => MAIN: '✗ 2 durability-critical catch(es) degrade quietly ... found: catch swallows the failure with no log at all' (the card's false red, reproduced exactly); MINE: green, '--list' shows 'loud (error@8264)'. (b) logDurabilityFailure's body replaced with `this.logger.error?.(msg, meta)` => MAIN: '✓ 29 ... all loud' (THE TRAP, green over runtime-silent code); MINE: '✗ ... found: catch reaches `error` only CONDITIONALLY (error?.@4022 via logDurabilityFailure())'. Both mutations reverted with `git checkout --` from the commit, `git status` empty after. H2 CONFIRMED on main: '--list' prints 'sql-driver.ts:8258 guards runWideningAlters()@8253 → recovers on one branch, loud (error@4022 via logDurabilityFailure())' — transitive helper-following works, and #9665's sites are green for the right reason; but only because the helper body spells an UNCONDITIONAL `this.logger.error(...)`, so the helper is a sanctioned shape, not a sufficient answer. PACKAGE SUITES (flock-serialized, dependency closure built first with `pnpm --filter '<pkg>^...' build`): plugin-audit 'Tests 270 passed (270)', plugin-email '418 passed (418)', plugin-security '1293 passed (1293)'; `typecheck` EXIT=0 on all three (it first caught the missing `warn` on AuthEventAuditLogger — a real find, not noise). REVERSE-VERIFICATION of the new tests, predicted direction RED and observed RED: reverting the three source repairs with `git checkout origin/main -- <3 files>` (tests untouched) => 'AssertionError: expected [] to have a length of 1 but got +0' (plugin-audit), 'expected undefined to be defined' (plugin-security), plugin-email x — real exit 1 read without a pipe; restored with `git checkout HEAD --`, byte-exact, all three green again. GATE UNION on the final commit, derived with `node scripts/pm/dispatch-gates.mjs <changed paths>` rather than recalled: check:durability-log-level, check:nul-bytes (plus a manual control-byte grep over my diff, clean), check:cross-package-test-inputs, check:test-source-alias, check:type-source-resolution, check:engine-double-contract, check:where-matcher, check:query-options-erasure, check:type-check-coverage, check:empty-changeset ('1 declaring changeset(s) added') — all exit 0. check:i18n first printed 'PREREQUISITE NOT MET — the workspace CLI is not built ... Nothing was checked', which is a refusal and not a pass, so I built @objectstack/cli and re-ran: 'OK (9 package(s) — all bundles in sync)', EXIT=0 read with echo, not through a pipe. NOT run locally: check:type-check-debt --re-measure (needs the whole workspace closure built; left to CI).",
  "open_questions": [
    {
      "question": "Should a sink that declares the durability channel be allowed to declare `error` OPTIONAL with no required alternative? Every seam this card's tightening found was on such a type, and `AuthEventAuditLogger` declared no `warn` at all — the call site could not have been written correctly against the contract it was given. This is the sink-TYPE axis of the same defect, and it is not settled by this PR.",
      "options": [
        "A — leave the sink types as they are (optional `error`, now with a `warn` beside it everywhere) and let the gate keep judging the call sites. Cheapest; the contract still permits a silence the gate then has to catch case by case.",
        "B — require `warn` on any sink type that declares an optional `error`, enforced by a small gate over the logger-ish interfaces. Declared = enforced: it becomes impossible to hand code a channel with no fallback. Costs a new gate and a scan population that has to be drawn.",
        "C — make `error` REQUIRED on these sink interfaces (the kernel `Logger` already requires it). Strongest, and it deletes the whole fallback idiom — but it is a breaking change to three exported plugin types, and it forecloses the legitimate `{ warn }`-only host the drivers were explicitly written for."
      ],
      "recommendation": "B, but not now and not by me. On real business need: hosts DO inject reduced sinks (SqlDriver's `error?` exists for exactly that), so C's premise is false today and A leaves a live foot-gun. On long-term soundness: B is the contract-first version — the defect is in the producer (the type), which is where AGENTS.md says to fix it, rather than in per-site vigilance. On making AI-written code hard to get wrong: an AI writing a new durability report reads the interface, sees `error?` and nothing else, and writes `error?.(…)` — B is what makes that impossible at the point of authoring rather than catching it one gate-run later. Not in this PR because drawing the population of 'logger-ish interfaces' is precisely the fuzzy-scope problem this file's own header warns produces a gate people disable, and pricing it is a lane-PM call, not a dev call."
    }
  ],
  "out_of_scope_findings": [
    "filed as #9748: two sibling `logger?.error?.(…)` SUMMARY reports (outbox-sweep.ts:227, permission-set-projection.ts:1067) that no catch guards, so no gate can see them — after this PR the per-row failure lands at `warn` on an error-less sink while the summary counting those same failures stays silent"
  ]
}

Generated by Claude Code

fix(plugin-email,plugin-security): the two durability SUMMARY reports…

feat(devx): a sink declaring an optional `error` must declare a guara…

fix(plugin-email,plugin-security): the two durability SUMMARY reports…

Require a guaranteed `warn` channel on any sink type that declares an…

fix(plugin-audit): key the lost auth-event row report per CAUSE, and …

fix(plugin-audit): read-audit reports once per CAUSE, not once per pr…

Navigation Menu

Repository navigation

check-durability-degradation-log-level: `loggerLevel` cannot see the `(logger.error ?? logger.warn)(…)` fallback, so a loud catch reads as `silent-swallow` — and the spelling it CAN see prints nothing #9657

The blind spot

Measured, with the reproduction

Why this direction of error is the dangerous one

The population

Options, none free

Related

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

check-durability-degradation-log-level: loggerLevel cannot see the (logger.error ?? logger.warn)(…) fallback, so a loud catch reads as silent-swallow — and the spelling it CAN see prints nothing #9657

Description

The blind spot

Measured, with the reproduction

Why this direction of error is the dangerous one

The population

Options, none free

Related

Activity

os-steve commented on Aug 18, 2026

os-steve commented on Aug 18, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions

check-durability-degradation-log-level: `loggerLevel` cannot see the `(logger.error ?? logger.warn)(…)` fallback, so a loud catch reads as `silent-swallow` — and the spelling it CAN see prints nothing #9657