`looksLikeInternalErrorLeak` does not recognise Postgres `relation "…" does not exist` — measured false, so that dialect's line ships from every boundary that runs the predicate

[hotlong]

Found while implementing #8086 (PR #8130), which applies this predicate at the direct-mount package door. Measured, not inferred. Filed unassigned; nobody is on it.

The gap

looksLikeInternalErrorLeak (packages/types/src/error-leak.ts) is the shared "does this message leak server internals?" heuristic. Its keyword set is:

sqlite_ · sqlstate · startsWith("insert into ") · startsWith("update ")
startsWith("select ") · startsWith("delete from ") · constraint failed
unique constraint · foreign key

None of those matches Postgres's canonical missing-relation phrasing. Measured against the shipping predicate:

true   "SQLITE_ERROR: no such table: sys_metadata"
false  "relation \"sys_metadata\" does not exist"
true   "UNIQUE constraint failed: sys_metadata.name"
false  "Failed to delete customization overlay: relation \"sys_metadata\" does not exist"
true   "SQLSTATE 08006: connection failure"

So the SQLite phrasing of a missing table is withheld and the Postgres phrasing of the same condition is disclosed, naming a physical table, from every boundary that applies the predicate — HttpDispatcher.error (#3867) and the direct-mount package door (#8086) among them. The one seam that is NOT affected is rest-server.ts's resolveErrorResponse band, which withholds every 5xx regardless of the predicate (#5437) — which is why rest-5xx-message-sanitization.test.ts has a green "Postgres phrasing is withheld too" case while the predicate itself returns false for that string. That case pins the seam's behaviour, not the predicate's, and reading it as the latter is the trap this issue exists to name.

Pinned as a live case in packages/rest/src/package-door-5xx-message-sanitization.test.ts (the residual case, added by PR #8130), which goes red the day this is fixed.

Why not "just add the keyword"

That is one option, and it is the cheapest, but it is worth deciding rather than reflexing — the predicate's own module note already argues against growing it:

Deliberately a heuristic over the message, not a driver taxonomy … Teaching the heuristic to recognise [read-scope-sql] would have been more message sniffing, which is the mechanism #5352/#5367 exist to remove.

That was written when #5811 added declaresServerFault alongside it, keying on the producer's declaration rather than its phrasing. Each new dialect string is one more entry in a list that is unbounded by construction — Postgres alone has relation … does not exist, column … does not exist, permission denied for table …, and MySQL/MSSQL/Oracle each phrase all of it differently. Every one of them is a leak the current set misses, and the list can only ever be as complete as the dialects someone has met.

Options, briefly:

• A — add the phrasings. Cheap, immediate, and strictly better than today. Unbounded by construction, and each addition is a guess about dialects nobody in this repo has run.
• B — reach for declaresServerFault where the producer declares. Already exists (/analytics/query 仍把 RLS 策略字段名回显给调用方 —— read-scope 拒收的泄漏在姐妹面上没堵,#5367 只堵了 dataset 路由 #5811) for exactly the messages a phrasing test cannot see. Does not help a bare Error from a driver, which declares nothing — and that is precisely the case here.
• C — fix at the producer. The direction The direct-mount package door ships 5xx messages verbatim — no leak heuristic, unlike the dispatcher twin and unlike rest-server's own /data exit #8086's ruling recorded and did not fold in: a producer that does not interpolate driver text into a client-facing message needs no predicate to recognise its dialect. The only option whose correctness does not depend on having enumerated the world's SQL engines.

A and C are not exclusive; A is a reasonable interim while C is scoped, the same shape #8086 settled on for its own door.

Note for triage

I could not find an existing card for #8086's option C ("metadata-protocol stops interpolating driver text into client-facing messages; the raw text goes to logs only"). #8086's PM ruling records it as the direction and says it "needs a producer-by-producer sweep, which is its own card", but no such issue appears in a search. If it was never filed, this issue and #8131 are both arguments for it, and all three probably want to be read together.

[hotlong]

Triage: lands in domain:cli (packages/types/src/error-leak.ts), queued (pm:queue), type Bug, security.

Rationale and scope ruling: option A (add the missing dialect phrasings) as the immediate interim — strictly better than today — while option C (producer-side fix) is tracked separately as #8136. A and C are explicitly non-exclusive per the card, and #8086's ruling already recorded C as the direction, so this card must NOT grow into the producer sweep. Scope here: add the known Postgres phrasings (relation … does not exist, column … does not exist, permission denied for table …) and any equally canonical lines for drivers this repo actually ships (sqlite, postgres via driver-sql), ⛔ not a speculative census of dialects nobody runs — that is the unbounded-list trap the module note warns about.

The deliberately-red-in-future pin in packages/rest/src/package-door-5xx-message-sanitization.test.ts (added by PR #8130) goes red when this lands — invert/absorb it as the new green, per the pin-flip discipline. Also keep the trap the card names: rest-5xx-message-sanitization.test.ts's green "Postgres phrasing is withheld too" case pins the seam, not the predicate — do not cite it as predicate coverage.

Size/model suggestion: S-mechanical borderline M (judgment is in which phrasings qualify), sonnet acceptable, opus if the dev proposes reshaping the predicate.

---

Generated by Claude Code

[os-zhuang]

Claim: PM loop round 1
Session: session_016pY4Xb2iDecfDtT3CWoiTW
Branch: claude/issue-8132-error-leak-pg-phrasings
Worktree: objectstack-issue-8132
Domain: domain:cli
File surface: packages/types/src/error-leak.ts, packages/types/src/error-leak.test.ts (or the existing predicate pin), packages/rest/src/package-door-5xx-message-sanitization.test.ts (residual-case flip only) (stop on breach; explain in the report)
Container & model: M, mode:subagent, model: opus
Serial constraints cleared: packages/types — #8087 (dispatcher error-code ledger gate, packages/runtime + packages/types) is queued and ruled but deliberately NOT dispatched this round to keep this batch package-disjoint. packages/rest — #8039 (ruled 2026-08-12, rest-server.ts) likewise held to the next round; this card's only packages/rest touch is the residual pin at package-door-5xx-message-sanitization.test.ts:257. No open PR touches error-leak.ts (checked against the open-PR list at claim time).

Premise re-verified on origin/main @ 43f37e1 before dispatch:

• looksLikeInternalErrorLeak keyword set is unchanged and matches the card verbatim (sqlite_, sqlstate, the four startsWith statement forms, constraint failed, unique constraint, foreign key) — no Postgres phrasing among them.
• The deliberately-red-in-future pin is present: package-door-5xx-message-sanitization.test.ts:83 defines PG_NO_RELATION, and :257 asserts looksLikeInternalErrorLeak(PG_NO_RELATION) is false. That assertion is what inverts when this lands.

---

Generated by Claude Code