spec + SQLite drivers: $like _ matches one code point on SQLite but one UTF-16 unit in formula / driver-memory — $like: '_' returns a stored emoji over REST; the spec never says which "character" is meant #20143

Activity

1. objectstack-fleet commented on Sep 25, 2026

   

   objectstack-fleetbot

   on Sep 25, 2026

   ContributorAuthor

   More actions

   Path: an API a customer can call | api-backend.query-contract-matrix | P2

   Triage: first grade — bug · priority:p3 · domain:spec · area:api · pm:queue — direction: _ is one code point

   Triage: lands in packages/spec/src/data/filter.zod.ts (likePatternToRegexSource, the regex built without the u flag, and the LIKE_DESCRIPTION / docblock wording) ⇒ domain:spec; rationale: the same REST filter answers differently by driver on characters outside the BMP (measured at GET /api/v1/data/...: $like: '_' returns a stored emoji on SQLite, not on the JS faces). Values outside the BMP with a _ pattern are a narrow shape ⇒ p3.

   Triage seat #6015 · session_01CRZSc7dU8oDStbTbSwhuZe · 2026-09-25T12:51Z. ⛔ Not a claim, ⛔ not a dispatch. Read: this card and its full thread, and origin/main.

   Why this is not sent to the decision box: the spec's own text already says "_ matches exactly one character", and the mainstream reading of one character in SQL LIKE (and SQLite GLOB) is one code point. Four of the five faces already answer by code points (better-sqlite3, sql.js, local and remote TursoDriver); the outlier is the one translation that omits the u flag. That is an implementation defect measured against the declared text, ⛔ not a choice between two contracts. If the claiming seat measures a real semantic reason for the UTF-16 reading (for example a CEL rule the formula face must keep), stop and report it here, and the card goes to the decision box then.

   Execution notes

   1. Give likePatternToRegexSource the u flag and check what that does to \ escapes and [\s\S]; reword the descriptions to say "one code point".
   2. Pin the family cells (_, __, a_b, a__b over values outside the BMP) on every face against the one code-point oracle.
   3. formula and driver-memory change their answer set: declare it (Clause-②) by the level rules.
2. objectstack-fleet

   added

   area:apiThe API a customer can call, and integrations — REST, connectors, webhooks, jobsThe API a customer can call, and integrations — REST, connectors, webhooks, jobs

   bugSomething isn't workingSomething isn't working

   domain:spec

   pm:queue

   priority:p3

   pm:dispatched

   and removed

   pm:queue

   on Sep 25, 2026
3. objectstack-fleet

   assigned

   os-zhuang

   on Sep 27, 2026
4. objectstack-fleet commented on Sep 27, 2026

   

   objectstack-fleetbot

   on Sep 27, 2026

   ContributorAuthor

   More actions

   Claim: PM loop round 1 · 2026-09-27T09:52Z
   Session: session_01Rjy9MeetSfq34PKn81CRiN
   Account: os-zhuang
   Branch: claude/issue-20143-like-underscore-code-point
   Worktree: objectstack-issue-20143
   Domain: domain:spec
   Seat: domain:spec#1
   File surface: packages/spec/src/data/filter.zod.ts likePatternToRegexSource (:1302 on e7f69dbb) and the LIKE_DESCRIPTION / docblock wording, with tests; formula / driver-memory tests only where their answer set flips; .changeset/. ⛔ Not PR #20204's filter.zod.ts hunks (:6, :276–:334, :1479, :1644); ⛔ not checkFilterConditionComparands (#20116); ⛔ no driver-sql / Turso source. (stop on breach; explain in the report)
   Container & model: S, mode:subagent, model: default judgment tier (the default slot is taken). The JS faces' answer set moves, so it is reviewed at CONTRACT_REVIEW_TIER before enqueue. Verify-lock queue empty at the gate read (09:48Z).
   Clause-②: yes
   Thread-read: 5832668207
   Serial constraints cleared: filter.zod.ts is shared with this seat's PR #20204 (#19886 stage 2b, in the merge queue); its hunks are disjoint from :1302, and the dev merges main after it lands. #20116's region (checkFilterConditionComparands) is untouched. No other open PR touches filter.zod.ts (all 14 open PRs' file lists read this act). Triage 5832668207 set the direction (one code point), with a stop if a real semantic reason for the UTF-16 reading is measured.
5. objectstack-fleet commented on Sep 27, 2026

   

   objectstack-fleetbot

   on Sep 27, 2026

   ContributorAuthor

   More actions

   os-dev-report

   {
   "issue": 20143,
   "status": "done",
   "branch": "claude/issue-20143-like-underscore-code-point",
   "pr": "#20222",
   "session": "session_01Rjy9MeetSfq34PKn81CRiN — mode:subagent, the PM session id; identity = branch",
   "premise_still_valid": true,
   "summary": "Premise re-measured FIRST at e7f69db over 18 patterns x 9 astral/BMP values: every SQLite face (better-sqlite3, sql.js, TursoDriver local, TursoDriver remote on a real libSQL engine) equals the code-point oracle; the four JS faces (spec matchesLikePattern, formula, driver-memory $like door, driver-memory AST like door) answer by UTF-16 unit — 48 differing face-cells; at head 713d150: 0. Fix: new spec export likePatternToRegExp(pattern, foldAscii) compiles likePatternToRegexSource with the u flag (the flag now lives in the one shared definition); matchesLikePattern evaluates it; driver-memory both mingo doors switch from a local flagless new RegExp to it (measurement put the flag choice there, so a spec-only change would have left them on UTF-16). LIKE_DESCRIPTION, the translation docblock and the GLOB table now define a character as one Unicode code point. Stop condition measured NOT met: nothing in formula lowers CEL to $like, and formula CEL size() already counts code points (size of one emoji = 1). Merged origin/main 585c9af after #20204 landed (clean; its $ne hunks untouched).",
   "tests": "All at HEAD 713d150 (merge of origin/main 585c9af), one locked batch, os-verify-lock VERDICT command-exit 0 (held 1246s, waited 239s, shared box). Build closure turbo 10/10. spec check:generated exit 0 after merge (api-surface regenerated earlier: 0 breaking, 1 added likePatternToRegExp). Full suites: spec 542 files / 15943 passed + 2 todo; formula 38 / 1039; driver-memory 55 / 1339; driver-sqlite-wasm 35 / 656; driver-turso 73 / 1921 (16 skipped); driver-sql 189 files + 11 skipped / 3111 passed (170 skipped). Typecheck exit 0: spec, formula, driver-memory, driver-sql, driver-sqlite-wasm, driver-turso; new test files counted in compiled programs via --listFiles (tsconfig.json for the 4 drivers = 1 each; tsconfig.test.json via check:test-typecheck for spec and formula = 1 each). New suites: spec 23, formula 12, memory 11, sql 12, wasm 12, turso 12 — all pass. Cross-face probe from dist: 48 differing cells at e7f69db, 0 at b4d7c35 and 0 at 713d150. ABLATION (fix committed first; mutation via scripts/ablation-replace.mjs under its trap plus own trap restoring from HEAD): u flag replaced by an expression evaluating to empty flags carrying marker ABLATION20143; on-disk injected=1 anchor-remaining=0; pnpm --filter @objectstack/spec build exit 0; ablation-dist-preflight exit 0 (marker in dist). Direction predicted before running: JS faces red on family cells, SQLite faces red only on the cross-face leg. Observed: spec 13 failed/23, formula 11/12 (CEL size pin green), memory 10/11, sql 1/12, wasm 1/12, turso 1/12 (each the cross-face leg) = predicted. Restore: hash-object == HEAD blob 838a80720b3d, git diff HEAD empty, porcelain clean, rebuild exit 0, preflight --absent exit 0, all six suites green again.",
   "mcp_calls": "0 — no MCP GitHub tool called",
   "api_writes": "3 relay writes as objectstack-fleet[bot] (each one POST /repos/objectstack-ai/objectstack/dispatches executed by the fleet-write workflow): (1) pr_create -> POST /repos/objectstack-ai/objectstack/pulls (draft, PR 20222; run 36316413225 success); (2) label-write --assign os-zhuang -> POST /repos//issues/20222/assignees (run 36316461200 success; read back MATCHES); (3) post-stamped os-dev-report -> POST /repos//issues/20143/comments. Plus 6 git pushes of the branch (probe + 5 updates; git push is not a REST write). Reads only otherwise (GET issue, comments, pulls?head=).",
   "gates": {
   "derived_at": "713d1504 via node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack",
   "derived": 108,
   "ran_exit_0": 101,
   "unrun": 0,
   "reconcile": "dispatch-gates --ran: 108 derived famil(ies) accounted for — 101 run, 7 NOT-MEASURED (7 DERIVED from a recorded exit 3); exit 0",
   "not_measured": [
   "NOT MEASURED: pnpm --filter @objectstack/lint run check:doc-formula-expressions, reason: exit 3 PREREQUISITE NOT MET (@objectstack/lint not built)",
   "NOT MEASURED: pnpm --filter @objectstack/lint run check:doc-security-posture, reason: exit 3 PREREQUISITE NOT MET (@objectstack/lint not built)",
   "NOT MEASURED: pnpm check:docs-transcript-drift, reason: exit 3 PREREQUISITE NOT MET (packages/lint/dist absent)",
   "NOT MEASURED: pnpm check:lean-entry-closure, reason: exit 3 PREREQUISITE NOT MET (objectql dist absent)",
   "NOT MEASURED: pnpm --filter @objectstack/spec run check:skill-examples, reason: exit 3 PREREQUISITE NOT MET (client-react dist absent)",
   "NOT MEASURED: pnpm check:dual-build-cjs-loads, reason: exit 3 PREREQUISITE NOT MET (needs whole-workspace build)",
   "NOT MEASURED: pnpm check:type-check-debt, reason: exit 3 PREREQUISITE NOT MET (needs every package dist built)"
   ],
   "not_measured_cause": "the targeted prerequisite build (turbo build --filter=@objectstack/lint... --filter=@objectstack/objectql --filter=@objectstack/client-react) hit two os-verify-lock queue-timeouts (exit 99, 9m each, never acquired; holders issue-20085 batch and issue-20116 batch-b.sh, live); declared narrowing to CI Lint & Repo Gates",
   "roster_in_touched_dirs_exit_0": [
   "check:meta-url-spelling",
   "check:spec-changes",
   "check:authz-resolver",
   "check:error-code-casing",
   "check:filter-alias-parity",
   "check:object-def-param-keys",
   "check:tenant-chokepoint"
   ],
   "changeset_gates": "check-adr-0087-registration exit 0: [clause-②-narrowing] not-required (no-migration-prescription); check-changeset-no-major exit 0 (level axis reads the PR body in CI); check-empty-changeset exit 0",
   "locked_runs": [
   "run1 build+6 new suites: command-exit 0 (b4d7c35)",
   "two queue-timeouts exit 99 (ablation attempt, then reissued inside the batch)",
   "batch heavy.sh: command-exit 0 (713d150)",
   "prerequisite build x2: queue-timeout exit 99 = NOT MEASURED"
   ]
   },
   "line_budget": "skills/** untouched: no skill line budget applies. Diff vs merge base 585c9af: 12 files, +779 / -21 (source: filter.zod.ts +53/-8, memory-driver.ts +12/-8; the rest tests, changeset, 3 generated artifacts). Under the 5000-line Tier threshold; no governed surface touched.",
   "files_changed": [
   ".changeset/20143-like-underscore-code-point.md",
   "content/docs/references/data/filter.mdx",
   "packages/drivers/driver-memory/src/memory-20143-like-code-point.test.ts",
   "packages/drivers/driver-memory/src/memory-driver.ts",
   "packages/drivers/driver-sql/src/sql-driver-20143-like-code-point.test.ts",
   "packages/drivers/driver-sqlite-wasm/src/sqlite-wasm-20143-like-code-point.test.ts",
   "packages/drivers/driver-turso/src/turso-20143-like-code-point.test.ts",
   "packages/formula/src/matches-filter-like-code-point.test.ts",
   "packages/spec/api-surface/data.json",
   "packages/spec/export-origins/data.json",
   "packages/spec/src/data/filter-like-code-point.test.ts",
   "packages/spec/src/data/filter.zod.ts"
   ],
   "deviations": [
   "File surface extended by measurement (order: 若实测表明真正的生产者在别包,按实测修): packages/drivers/driver-memory/src/memory-driver.ts source changed — its two mingo doors compiled the spec source with a local flagless new RegExp, so the flag choice lived there; they now call the spec likePatternToRegExp. Stated in the PR body.",
   "Test files added in driver-sql, driver-sqlite-wasm and driver-turso (tests only, no source) to execute Zone 1 execution note 2 (pin the family cells on every face); the claim listed formula/driver-memory tests only.",
   "7 derived gates NOT MEASURED (exit 3 prerequisites) after ~27 min of verify-lock queue with no acquisition; declared narrowing, CI runs them.",
   "PR body carries the claim line verbatim (Clause-②: yes) per os-dev; the changeset carries the measured arm (Clause-②: yes (narrowing)) as the order directs — both grade minor; the body names the arm in prose.",
   "Commit trailers use the AGENTS.md model-free pair (Claude-Session + Co-authored-by Claude) instead of the harness reminder model-named trailer, and the PR footer uses the AGENTS.md session-URL form; AGENTS.md (user instruction via CLAUDE.md) takes precedence per the reminder itself.",
   "Two long foreground commands passed the 600s tool cap and were moved to the background by the harness; each was then awaited in-turn with a foreground tail --pid block, never polled."
   ],
   "settled_choices": [
   {
   "choice": "Where the u flag lives: new spec export likePatternToRegExp (chosen) vs a spec constant LIKE_PATTERN_REGEX_FLAGS vs a literal u at each driver-memory call site vs a u-independent surrogate-aware source",
   "business_need": "all four fix the same measured defect for the same 5 measured consumers (spec, formula, 3 driver-memory doors); none adds a capability",
   "long_term": "the export keeps the translation and its compilation in ONE definition, the file own rule (a translation written per package agrees on the day it is typed and never again); per-site u literals re-create that drift; a surrogate-aware non-u source is a complex regex nobody can audit",
   "ai_error_proofing": "a RegExp-returning builder cannot be compiled without the flag; a constant or a literal is a second argument a future face forgets; likePatternToRegexSource stays exported (removal would be breaking) with a docblock prohibition",
   "startup_scope": "+1 export (api-surface: 0 breaking, 1 added), zero new gates, no new capability surface",
   "recommendation": "likePatternToRegExp"
   }
   ],
   "open_questions": [],
   "out_of_scope_findings": [
   "carrier: 承接者:无 · noted, not filed — hand-written docs restate _ is exactly one character (content/docs/api/client-sdk.mdx:553, content/docs/deployment/troubleshooting.mdx:148, packages/client/src/query-builder.ts:88 JSDoc); true under the spec now-defined reading; in PR Acceptance notes",
   "carrier: 承接者:无 · noted, not filed — likePatternToRegexSource remains exported with zero consumers outside the spec; a future face could compile it without u; docblock prohibition; retirement would be a removal (not this card); in PR Acceptance notes",
   "carrier: 承接者:无 · noted, not filed — unpaired surrogates in a value or pattern: unmeasured (UTF-8 SQLite cannot store one); Postgres/MySQL _ arms NOT MEASURED as on the card; in PR Acceptance notes"
   ]
   }
6. objectstack-fleet commented on Sep 27, 2026

   

   objectstack-fleetbot

   on Sep 27, 2026

   ContributorAuthor

   More actions

   Review: ACCEPT · PR #20222 at head 713d1504d6e9c710c94dae8efaeda953e478b932 · 2026-09-27T11:52Z

   domain:spec seat 1 (session_01Rjy9MeetSfq34PKn81CRiN), reviewer of record, on claim 5854820140. Checked against GitHub, ⛔ not against the report (5855514577).

   Checklist

   • Shape: draft, base main, first line Fixes #20143, the body's only closing keyword. The card's one deliverable (_ = one code point on every face) is complete.

   • Scope: 12 files, +779/−21, no governed path (check-governed-merges.mjs --pr 20222: 0 of 12 paths, 800 lines). Source is filter.zod.ts (+53/−8) and memory-driver.ts (+12/−8). The rest is tests, the changeset and three generated files (api-surface/data.json, export-origins/data.json, references/data/filter.mdx).

   • filter.zod.ts, read by the seat:

     • the LIKE_DESCRIPTION sentence;
     • the likePatternToRegexSource docblock (the source means one code point ONLY under u);
     • a new export likePatternToRegExp(pattern, foldAscii?) = new RegExp(source, 'u');
     • matchesLikePattern routed through it;
     • one GLOB-table row.

     The translation itself is unchanged. The u-strict escape set the docblock names includes ] and }, so no pattern that compiled before throws under u.

   • Region fences: fix(spec)!: refuse an array under $ne at the shared comparand-shape face and at FieldOperatorsSchema.$ne #20204's hunks landed before the fork. spec: FilterConditionSchema (the save-time door) admits a non-boolean $null / $exists flag, so a stored dataset or widget filter carrying one saves clean and is refused 400 on every query face #20116's checkFilterConditionComparands (:1732) is untouched. The seat's git merge-tree of spec: FilterConditionSchema (the save-time door) admits a non-boolean $null / $exists flag, so a stored dataset or widget filter carrying one saves clean and is refused 400 on every query face #20116's current branch tip (60f57dad) against this head is clean.

   • Merge with current main (805af4f2): git merge-tree is clean. main moved by 8 commits since the merge base 585c9af5; the only overlap is sorted-list insertions in the two generated spec JSONs, which a three-way text merge takes as the union.

   • Changeset: minor on spec, formula and driver-memory; Clause-②: yes (narrowing); an ADR-0087 not-required (no-migration-prescription) marker with its reason. The answer-set table moves both ways, and only for values that hold a character outside the BMP.

   • Tests (report fields): 48 face cells differing before the change, 0 after. The ablation (flagless compile restored) turned the predicted cells red.

   • Gates (report): 108 derived, 101 exit 0, 7 NOT MEASURED. All 7 are exit-3 prerequisites (unbuilt lint / objectql dist) after ~27 minutes of lock queue. CI runs each of them.

   • CI at this head: 20 success / 2 skipped (Console Pin Gate, Packed-tarball smoke, both expected) / 10 in progress / 0 failing. This is an honest in-progress reading; nothing lands before the whole set is green.

   Claim-surface amendment (recorded by the seat)

   • The claim fenced driver-memory to tests. The dev changed packages/drivers/driver-memory/src/memory-driver.ts source: its two mingo doors ($like / $ilike and the AST like / ilike) compiled the spec source with a local flagless new RegExp, so the flag choice lived there, and they now call the spec's likePatternToRegExp. The order's producer-side clause covers this (「若实测表明真正的生产者在别包,直接按生产者侧修」), and the report states it.
   • The tests-only files in driver-sql, driver-sqlite-wasm, driver-turso and formula pin the family cells on every face, as Zone 1 execution note 2 asked. ⛔ There is no driver-sql or Turso source change.
   • memory-driver.ts belongs to domain:engine. The seat read every open pm:dispatched claim (29 cards): the one engine claim that names driver-memory (driver-sql + driver-memory: an epoch-millisecond NUMBER against a date field is read by neither driver's storage rule — where: { placed_on: { $gt: 1769940000000 } } returns 6 of 6 rows on SqlDriver and 0 on InMemoryDriver over REST #20203, session_01Bvd69VPa6puiNzzPUroDBx) fences it to tests and temporal-storage-form.ts, disjoint from the $like doors. No open PR other than this one touches memory-driver.ts. The seat follows this out-of-lane surface to landing (the claiming seat's cross-lane path).

   Contract review: owed on Clause-②: yes (the JS faces' answer set moves in both directions). An isolated at-tier reviewer is running. needs:contract-review is hung on PR #20222 in the same act; ⛔ not readied before a same-shape PASS on this head.

   Findings, one line each (the PR's ## Acceptance notes)

   • Three hand-written surfaces restate 「_ is exactly one character」 (client-sdk.mdx:553, troubleshooting.mdx:148, the query-builder.ts:88 JSDoc) and remain true under the defined reading → Acceptance notes.
   • likePatternToRegexSource stays exported with no consumer outside the spec, and its docblock forbids a flagless compile → Acceptance notes. Retiring it is a removal, ⛔ not this card.
   • Unpaired surrogates and the Postgres / MySQL _ arms are NOT MEASURED, as on the card → Acceptance notes.
7. objectstack-fleet

   closed this as completedin #20222on Sep 27, 2026
8. objectstack-fleet commented on Sep 27, 2026

   

   objectstack-fleetbot

   on Sep 27, 2026

   ContributorAuthor

   More actions

   Landed: PR #20222 → main 93cfc3f593cbd794f244b6ce922e7410776cad8a · 2026-09-27T12:43Z

   domain:spec seat 1 (session_01Rjy9MeetSfq34PKn81CRiN), on claim 5854820140. Fixes #20143 closed this card completed.

   • Record chain: seat ACCEPT 5855574230 (with the claim-surface amendment for driver-memory source); at-tier PASS 5855732543 at head 713d1504, the head that landed; CI green at that head (34 success, 5 expected skips); merge queue enqueued 12:16:58Z, merged 12:43:03Z.
   • Verified on main, two readings:
     • The merge commit 93cfc3f5 is on origin/main, and no queue branch for fix(spec): $like / $ilike _ is one Unicode code point on every JS face #20222 remains.
     • Diff against its first parent: 12 files, +779/−21, the same as the PR.
     • export function likePatternToRegExp is present in filter.zod.ts at 93cfc3f5 and absent at its parent (the control). memory-driver.ts calls it twice.
   • Stray closures: none. This card is the only issue the merge closed; RLS: a policy's compiled filter skips the shared comparand-shape faces, so a null list member or null ordering bound reaches driver-sql, and the read and the write check disagree (the read hides a row its own check admits) #20212 (cross-referenced in the PR body) stays open.
   • pm:dispatched is lifted in this act; domain:* / area:* / type labels stay.

   Residuals, recorded here and not filed (the PASS record's ③):

   • MySQL arm: likePatternPredicate on mysql compares CAST(… AS BINARY), so _ would consume one UTF-8 byte there (read from sql-driver.ts:3956-3960), against the new sentence 「one code point」. It is pre-existing (it broke the old 「one character」 sentence for every non-ASCII value too) and unmeasured: there is no MySQL server here, and there is no named producer. ⇒ ⛔ Not a card under the filing gate (a self-declared unmeasured reach:). A future dispatch that can run MySQL files it with a measured cell.
   • likePatternToRegexSource stays exported with a docblock prohibition only; mingo would drop u if a future door passed $options. Both are recorded in the PASS record.
9. objectstack-fleet

   removed

   pm:dispatched

   on Sep 27, 2026
10. akarma-synetal

    added a commit that references this issue on Sep 28, 2026

    fix(spec): $like / $ilike `_` is one Unicode code point on every JS f…

    ...

    This issue will close once commit 93cfc3f is merged into the 'main' branch.

    93cfc3f
11. akarma-synetal

    added a commit that references this issue on Oct 7, 2026

    test(spec): the third data/ file group's test titles state each cited…

    ...

    969ffba