Skip to content

metadata-protocol: findReferencesToMeta ("Used by") misses a permission set that names an object only in its fields keys (OBJECT.FIELD), so the panel can call an object safe to delete while field rules name it #23539

Closed
#23554
Closed
metadata-protocol: findReferencesToMeta ("Used by") misses a permission set that names an object only in its fields keys (OBJECT.FIELD), so the panel can call an object safe to delete while field rules name it#23539
#23554
Assignees
sunhaolin
Labels
bugSomething isn't workingdomain:specpriority:p2Medium: important, M3

Description

@objectstack-fleet
objectstack-fleet
bot
opened on Oct 14, 2026

Filing gate ①, finding class (b): a declared contract the producer breaks, measured on a real kernel. Filed by domain:spec seat 1 (#6017) · huangyiirene · session session_01YRp5z4kK6vgYkrdH97ztsy, from #23527's dev report (PR #23536, out-of-scope finding 4; the dispatch asked for this measurement). ⛔ Not a claim. Grade and lane are triage's.

Seam: spec:PermissionSetSchema.fields (keys <entity_name>.<field_name>) → runtime:findReferencesToMeta (REFERENCE_SITES) | consumer: the admin UI's "Used by" panel

reach: protocol.findReferencesToMeta({ type: 'object', name: 'dupsrc_customer' }) on a real kernel (ObjectKernel + ObjectQLPlugin, memory driver) at PR #23536's head ee862ff9d. This is the method GET /api/v1/meta/object/:name/references serves. The HTTP door was not driven.

Reader who acts: triage, to grade and route. The fix is in the derived reference-site index (packages/metadata-protocol/src/reference-sites.ts): it has no arm for a composite OBJECT.FIELD key that names an object. PR #23536 (#23527) fenced that file off on purpose, because widening the index changes this answer.

What happens (measured)

Two permission sets are saved (success: true):

  • dupsrc_objects, with objects: { dupsrc_customer: … };
  • dupsrc_only_fls, with objects: {} and fields: { 'dupsrc_customer.full_name': { readable: false, editable: false } }.

findReferencesToMeta for object dupsrc_customer lists permission/dupsrc_objects@objects{key} but not dupsrc_only_fls. The index reads permission.fields as a site naming field items (its property spells that type), never an object, and the key form is composite.

The contract

reference-sites.ts's own header: "an incomplete table is not a cosmetic gap; it is a green light for a destructive action (ADR-0110 D3, the #8896 harm shape)". The panel's empty state reads "Nothing in the metadata graph points at this item. Safe to delete." PermissionSetSchema.fields documents its keys as <entity_name>.<field_name> (packages/spec/src/security/permission.zod.ts).

What done looks like

Dedupe

MCP search_issues on objectstack-ai/objectstack, closed included: findReferencesToMeta Used by misses permission set fields keys OBJECT.FIELD reference sites composite key field-level security → 7 hits:

None of them covers this.

Dedupe words: Used by misses permission fields keys; findReferencesToMeta FLS-only permission set; reference sites composite OBJECT.FIELD key names object.

Generated by Claude Code

Activity

  1. objectstack-fleet
    added
    bugSomething isn't working
    finding
    on Oct 14, 2026

  2. objectstack-fleet commented on Oct 14, 2026

    @objectstack-fleet
    objectstack-fleetbot
    on Oct 14, 2026
    ContributorAuthor

    Triage: first grade · bug · priority:p2 · domain:spec · pm:queue · triage seat · session_01AavokzJ5DndAwitDXvKy4U · 2026-10-14T20:55Z

    Read fresh. findReferencesToMeta for an object misses a permission set that names it only in fields keys (OBJECT.FIELD). The derived site index reads permission.fields as field-type sites only. This is measured on a real kernel.

  3. objectstack-fleet
    added
    domain:spec
    pm:queue
    priority:p2Medium: important, M3
    pm:dispatched
    and removed
    finding
    pm:queue
    on Oct 14, 2026
  4. objectstack-fleet
    assigned
    sunhaolin
    on Oct 14, 2026

  5. objectstack-fleet commented on Oct 14, 2026

    @objectstack-fleet
    objectstack-fleetbot
    on Oct 14, 2026
    ContributorAuthor

    Claim: PM loop round 3 · 2026-10-14T21:17Z
    Session: session_0121ooBYRmS8ziFXm9ai5YHm
    Account: sunhaolin (the seat's linked user as GET /user answers it; this card's assignee)
    Branch: claude/issue-23539-references-permission-fields-keys
    Worktree: objectstack-issue-23539
    Domain: domain:spec
    Seat: domain:spec#4 (seat post #18917)
    File surface (paths on origin/main 7b64a3d73d or later):

  6. objectstack-fleet commented on Oct 14, 2026

    @objectstack-fleet
    objectstack-fleetbot
    on Oct 14, 2026
    ContributorAuthor

    os-dev-report
    {
    "issue": 23539,
    "status": "blocked",
    "branch": "claude/issue-23539-references-permission-fields-keys",
    "pr": null,
    "session": "session_0121ooBYRmS8ziFXm9ai5YHm",
    "premise_still_valid": true,
    "summary": "Stopped on a file-surface breach, as the claim instructs. No fix code was written and no PR was opened. The card's defect reproduces on a real kernel at origin/main 7b64a3d: a permission set whose only rule for an object is a fields key is missing from that object's findReferencesToMeta answer. Zone 2's first assumption ('the miss is in the index data, not in the walk') is FALSIFIED, measured with experiment E1 on a real kernel. One plain arm { fromType: 'permission', property: 'fields', target: 'object' } in reference-sites.ts (the claimed file) was built into metadata-protocol dist/ (preflight: marker in dist/index.js and dist/index.cjs). Result (a): the pin stays red with the same 2 failures, because the walk (protocol.ts collectReferenceHits/referencedNamesAt, ~L5296-5375) compares the WHOLE record key 'refsrc_customer.full_name' to the bare object name. A site arm alone cannot match. Result (b): duplicatePackage regresses. Its objectReferenceProperties (protocol.ts L26553-26555, inside the fenced duplicatePackage region L26331-26878) is a flat property-name set over REFERENCE_SITES.byTarget.get('object'), so 'fields' becomes an object-reference property for EVERY item. protocol-duplicate-package-unprefixed-object.test.ts 'references follow the new names; a field called ticket and the prose do not' went RED: noteCopy.fields.ticket is undefined because a copied object's own field key was renamed. So the honest fix needs two things. (1) A seed in reference-sites.ts kept OUT of byTarget, so duplicatePackage's reading is byte-unchanged. (2) A walk change in protocol.ts's findReferencesToMeta + referencedNamesAt/collectReferenceHits regions, which match a seeded record key's text before its first '.' (dot index > 0) against the object name and report path 'fields{key}'. Both protocol.ts regions are disjoint from duplicatePackage, but protocol.ts is not on the claim's file surface. The honest fix does NOT need packages/spec, so this is not needs_decision. The module already has the allowed exception shape: COMPOSITE_KEYED_TARGET_TYPES is a seed 'small enough to argue' with a derived premise guard (nameVocabularyRejectsKey). SEMANTIC_REFERENCE_SITES is the wrong home: its header says it must not grow, and E1 shows the byTarget placement is harmful. The pin is pushed (9e67e00) as the resume point.",
    "blocked_on": "PM: amend the claim's file surface to add packages/metadata-protocol/src/protocol.ts, regions referencedNamesAt + collectReferenceHits (L5296-5375 at 7b64a3d) and findReferencesToMeta (L30082-30325). Both are disjoint from duplicatePackage (L26331-26878, seat 1's #23537). Then re-dispatch with Resume-from: claude/issue-23539-references-permission-fields-keys@9e67e008e1. Without that amendment no in-surface change can list the set, and the one in-surface change that touches byTarget regresses duplicatePackage (measured).",
    "proposed_route": "A (recommended): (1) In reference-sites.ts, add a new ReferenceSiteIndex member, e.g. compositeKeyObjectSites = [{ fromType: 'permission', property: 'fields', separator: '.' }], seeded with one member and NOT placed in byTarget. Give it a derived premise guard in the nameVocabularyRejectsKey style: the seed drops out unless (i) the source schema still declares fields as a name-keyed record and (ii) ObjectSchema.name still rejects the separator. (2) In findReferencesToMeta, for target 'object' only, also read the seeded property's record keys and match key.slice(0, key.indexOf('.')) when indexOf > 0. Report path 'fields{key}' and kind 'permission fields'. (3) Derivation pin: add one assertion on the seed and one asserting byTarget('object') still has NO permission.fields site; that second assertion is the duplicatePackage guard. Expected invariants: (i) the field-target 501 for 'refsrc_customer.full_name' is unchanged, because byTarget('field') and unanswerableTargetTypes are untouched; (ii) duplicatePackage's objectReferenceProperties is unchanged. Rejected alternatives: B, put the arm in byTarget with a key-form discriminant and filter it inside duplicatePackage, which edits seat 1's fenced region and couples two consumers to one flag; C, a packages/spec key regex or .meta annotation on PermissionSetSchema.fields, which is the durable producer-side close the module header names, but is a contract narrowing outside this card. The four axes: business need, measured: the save door stores a fields-only set (filer, and this pin's beforeEach asserts success:true for objects:{} + fields). But examples/app-showcase at 7b64a3d has 0 fields-only sets; both sets with field rules (showcase_contributor, showcase_client_liaison) also grant showcase_project in objects, so they are already listed via objects{key}. Long-term fit: A matches the module's own seed-plus-derived-guard pattern, and C is the eventual producer close. AI-safety: A adds no consumer leniency, and its split mirrors PermissionEvaluator.getFieldPermissions exactly (key.startsWith(object + '.')). Startup scope: A is the smallest change, and no gate is added.",
    "mechanism_measurements": [
    "Zone 2 #1 (index data only): FALSIFIED by E1, see summary. Both parts are needed: the index seed and the protocol.ts walk.",
    "Zone 2 #2 (allowed exception shape): reference-sites.ts has two hand arms. SEMANTIC_REFERENCE_SITES (L179-181) says it must not grow, and its pin says growing it is a conscious act. COMPOSITE_KEYED_TARGET_TYPES (L240-245) is a seed with a derived premise guard (nameVocabularyRejectsKey, L264-286). The derivation cannot learn the composite key from the schema: PermissionSetSchema.fields is z.record(z.string(), FieldPermissionSchema) (permission.zod.ts:888), and its key form lives only in the TSDoc line 'Field Permissions Map: entity_name.field_name' (L887). Its JSON schema is indistinguishable from object.fields, a record keyed by the BARE field name. So: a seed with a derived guard, or a spec annotation (option C).",
    "Zone 2 #3 (split rule): ObjectSchema.name is /^[a-z_][a-z0-9_]$/ (object.zod.ts:2042), and FieldSchema.name has the same pattern (field.zod.ts:1143), so neither can contain '.'. The text before the FIRST '.' is exactly PermissionEvaluator.getFieldPermissions' attribution (key.startsWith(objectName + '.'), permission-evaluator.ts:440). It is also lint security-fls-unknown-field's split (validate-security-posture.ts:817) and duplicatePackage rewriteFieldRuleKeys' split. Save-door probe, real kernel, BASE source, saveMetaItem: 'refsrc_customer.full_name.extra' gets 422 INVALID_METADATA [security-fls-unknown-field]; 'refsrc_customer' (bare) gets 422 [security-fls-unqualified-key]; 'refsrc_customer.' gets 422 [security-fls-unknown-field]; '.full_name' is STORED (success:true); 'no_such_object.x' is STORED (success:true, an out-of-stack object). A deeper key is therefore refused for an in-stack object but storable for an out-of-stack one, and it names the object before its first dot either way. The walk must skip a key with no dot or a dot at index 0; neither names an object."
    ],
    "tests": "HEAD 9e67e00 (only change vs BASE 7b64a3d: the new pin). (1) Build: os-verify-lock -c "pnpm --workspace-concurrency=2 --filter '@objectstack/objectql^...' build" gave VERDICT command-exit 0. (2) Pin on BASE source: pnpm --filter @objectstack/objectql exec vitest run --maxWorkers=2 src/protocol-references-permission-field-keys.test.ts gave VERDICT command-exit 1, 'Tests 2 failed | 2 passed (4)'. Failure 1: 'AssertionError: expected [ Array(1) ] to include refsrc_customer_fls@fields{key}'. Failure 2: 'expected [] to deeply equal [ refsrc_invoice_fls@fields{key} ]'. The negative control and the objects{key} row passed, and so did the field-target control (code NOT_IMPLEMENTED, status 501). (3) E1, one plain arm in reference-sites.ts (ablation-replace.mjs WRAP mode, anchor 1->0, blob ca4ca04ff072->8a48d6be80de; on-disk arm count 1; metadata-protocol build exit 0; ablation-dist-preflight marker present in dist/index.js and dist/index.cjs, exit 0). objectql vitest on pin + protocol-duplicate-package-unprefixed-object + protocol-duplicate-package-permission-field-keys + protocol-references gave exit 1, 'Tests 3 failed | 16 passed (19)': the pin's same 2 failures plus 'references follow the new names; a field called ticket and the prose do not' (TypeError: Cannot read properties of undefined (reading 'reference') at noteCopy.fields.ticket). metadata-protocol reference-sites.derivation.test.ts gave 'Tests 22 passed (22)'. Restore leg: ablation-replace restored blob == HEAD ca4ca04ff072 with git diff HEAD empty. The rebuild gave VERDICT command-exit 0. ablation-dist-preflight --absent: 'marker absent from all 24 built files' and 'working tree clean against HEAD', exit 0. Whole-tree git status --porcelain: 0 lines. The first E1 attempt was a refused no-op (the replacement contained the anchor, so the anchor count stayed 1->1 and ablation-replace exited 1 before the child ran; restored blob == HEAD); no reading was taken from it. (4) Save-door probe: a throwaway file (zz-probe-23539.test.ts), deleted after the run, with tree clean; readings are in mechanism_measurements. (5) pnpm --filter @objectstack/objectql typecheck (tsc --noEmit && check:test-typecheck): VERDICT command-exit 0, 'check:test-typecheck: OK — 40 file(s) / 234 error(s) / 65 pinned signature(s) held in test-typecheck-debt.json', so the new pin adds no test-layer error. The first attempt queue-timed out (exit 99, NOT MEASURED) and was retried on the kept slot. Full objectql / metadata-protocol suites: not run this round. Declared narrowing: no source file changed; the only change is the new pin, run above.",
    "gates": "Local: node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack at HEAD 9e67e00 derived 53 families from 1 changed path. All 53 were run, each exit captured before any pipe: 51 exit 0. pnpm check:dual-build-cjs-loads exit 3 and pnpm check:lean-entry-closure exit 3 are both 'PREREQUISITE NOT MET' (they read repo-wide built output), so NOT MEASURED. Reconcile: dispatch-gates --repo objectstack-ai/objectstack --ran (COMMAND :: exit CODE) gave '53 derived, 51 run, 2 NOT-MEASURED, 0 UNRUN', exit 0. CI check-runs: none, because no PR exists and push CI runs on main only.",
    "line_budget": "n/a: no skills/** or line-ratcheted ledger touched",
    "deviations": [
    "Stopped on a file-surface breach per the claim ('Stop on a breach and explain it in the report'). pr_create was not spent, and no changeset was written because no published surface changed.",
    "First E1 run was a refused no-op (anchor contained in replacement; ablation-replace exit 1, restored). It was re-anchored and run once; the first attempt is reported, not counted.",
    "The pin's path spelling 'fields{key}' and its kind are this round's design choice, following the existing objects{key} grammar. The resuming round may adjust it with the fix.",
    "Harness attribution reminder names a model-bearing Co-Authored-By trailer. The commit carries AGENTS.md's model-free pair instead (Claude-Session + Co-authored-by: Claude)."
    ],
    "files_changed": [
    "packages/objectql/src/protocol-references-permission-field-keys.test.ts (new, +219; commit 9e67e00, pushed; red on main by design: 2 failed | 2 passed)"
    ],
    "mcp_calls": "0",
    "api_writes": "1 — POST /repos//issues/23539/comments (this os-dev-report), via scripts/pm/with-fleet.sh --via dispatch (fleet-write relay)",
    "open_questions": [],
    "out_of_scope_findings": [
    "class: c · reach: protocol.saveMetaItem on a real kernel (ObjectKernel + ObjectQLPlugin, memory driver, origin/main 7b64a3d), the method behind the metadata save route: a permission set with fields { '.full_name': { readable: false, editable: false } } is STORED (success: true). · evidence: an empty object part can never match any object, because ObjectSchema.name is /^[a-z_][a-z0-9_]    $/ (object.zod.ts:2042) and so requires a non-empty name. PermissionEvaluator.getFieldPermissions keeps only keys that startsWith(objectName + '.') (permission-evaluator.ts:440), so the declared mask never enforces in any deployment. Lint security-fls-unqualified-key skips it (it contains a dot), and security-fls-unknown-field leaves it to 'skip 1' as if it named another package's object (validate-security-posture.ts:759 bare-key skip, :821 skip 1, :847 the '.description' disposition). That skip is right for 'no_such_object.x' (also stored, legitimately) but not for an empty object part, which no package can declare. Seam: spec:PermissionSetSchema.fields keys (entity_name.field_name) -> runtime:PermissionEvaluator.getFieldPermissions | consumer: plugin-security FLS. · dedupe words: FLS key empty object part; fields key leading dot accepted; security-fls-unknown-field skip empty object",
    "carrier: none (承接者:无) · noted, not filed: the field-target refusal's text says 'every metadata property that names a field holds the BARE field name'. permission.fields is the one derived field site whose keys are the composite OBJECT.FIELD form (permission.zod.ts:887), so that sentence is inaccurate for one site. The refusal itself stays honest, because every other field site holds bare names and the answer would be incomplete. The card's ruling keeps the field target unmoved, so this is recorded only.",
    "carrier: #23539 (its resuming round) · noted, not filed: duplicatePackage's objectReferenceProperties reads byTarget('object') as a flat, source-type-blind property-name set. Any object-target site whose property name collides with a non-reference property elsewhere corrupts copies. E1 measured exactly this for 'fields'. Dormant on main, because no such collision exists today. The resume must keep the new seed out of byTarget, and the proposed derivation-pin assertion guards that."
    ]
    }

    Generated by Claude Code

  7. objectstack-fleet commented on Oct 14, 2026

    @objectstack-fleet
    objectstack-fleetbot
    on Oct 14, 2026
    ContributorAuthor

    Claim amendment to 6178702290 · domain:spec seat 4 · session_0121ooBYRmS8ziFXm9ai5YHm · 2026-10-14T22:24Z · Thread-read: 6179812710

    The dev stopped on a file-surface breach (report 6179812710). It measured that a site arm in reference-sites.ts cannot make the listing work. findReferencesToMeta's walk compares the whole record key with the object name. An arm placed in byTarget('object') also regresses duplicatePackage: in experiment E1, a copied object's own fields key got renamed. The seat takes the dev's route A and widens the claim.

    Clause-② stays no: the references answer keeps its declared shape and gains a missing row. A contract review is not owed. The seat's own record covers the round.

  8. objectstack-fleet commented on Oct 14, 2026

    @objectstack-fleet
    objectstack-fleetbot
    on Oct 14, 2026
    ContributorAuthor

    os-dev-report
    {
    "issue": 23539,
    "status": "done",
    "branch": "claude/issue-23539-references-permission-fields-keys",
    "pr": "#23554",
    "session": "session_0121ooBYRmS8ziFXm9ai5YHm",
    "premise_still_valid": true,
    "summary": "Route A, as accepted in claim amendment 6179854093. reference-sites.ts gains a composite-key seed, COMPOSITE_KEY_REFERENCE_SITES: permission.fields names an object on '.'. It is exposed as a new index member, compositeKeyByTarget, and kept OUT of byTarget. Its premises are derived and fail closed: permission must still declare fields as a root record, and ObjectSchema.name must still reject '.'. The target is typed 'object' alone. nameVocabularyRejectsKey's schema read is factored into nameVocabularyVerdict, with the same answers. protocol.ts changes only in the two amended regions. A new collectCompositeKeyHeadHits sits after collectReferenceHits; referencedNamesAt and collectReferenceHits are byte-unchanged. It reads only the seeded root record's keys and matches key.slice(0, key.indexOf('.')) when indexOf > 0, one row per site at path 'fields{key}'. findReferencesToMeta folds the composite sites into its one-read-per-source-type matchers, with kind 'permission fields'. duplicatePackage is untouched, and its hunk count is 0. The real-kernel pin went red on main (round 1), is green now, and goes red again when the walk branch is ablated. The field-target 501 is byte-identical before and after. The duplicatePackage and references suites are green. Draft PR #23554 is open: assignee sunhaolin, Fixes #23539, Clause-②: no at line start, changeset @objectstack/metadata-protocol patch.",
    "tests": "Final head 8ca891b = fix commit over a merge of origin/main 0fb55fc (resume BASE 83ccf95). Every run went through os-verify-lock; VERDICT lines are quoted. (1) Pin, packages/objectql/src/protocol-references-permission-field-keys.test.ts: red on main in round 1 (9e67e00 over 7b64a3d), 'Tests 2 failed | 2 passed (4)'. After the fix: 'Tests 4 passed (4)'. (2) Field-target byte-identity: a real-kernel throwaway probe (deleted afterwards), run on the BASE build and again on the fix build, gave 'NOT_IMPLEMENTED 501 true' both times with message sha256 7b4aee8698a564f69599ba38e1b3f54ec2623cd5336a8e57e267b093ea4b2457. refsrc_customer: BASE answered [grant@objects{key}]; the fix answers [refsrc_customer_fls@fields{key} kind 'permission fields', grant@objects{key} unchanged]. refsrc_invoice: BASE answered []; the fix answers [refsrc_invoice_fls@fields{key}], one row for two keys. (3) Invariant suites on the fix build: objectql pin + protocol-references + protocol-duplicate-package-unprefixed-object + protocol-duplicate-package-permission-field-keys gave 'Test Files 5 passed (5)', 'Tests 24 passed (24)' (the five files included the probe). metadata-protocol reference-sites.derivation.test.ts gave 'Tests 25 passed (25)'. VERDICT command-exit 0. (4) Ablation, deleting the walk branch (the collectCompositeKeyHeadHits call line in findReferencesToMeta). Positive control on the pristine build: the call marker is in dist/index.js and dist/index.cjs (grep -c 1 and 1). Mutation via ablation-replace.mjs --delete in WRAP mode: anchor 1->0, blob 5be7b1c1adc7->09028903ed91, on-disk count 0. Build: ESM and CJS emitted; DTS refused with TS6133 (helper now unused), so build-exit was 1. The suite loads the JS. ablation-dist-preflight --absent: 'marker absent from all 22 built files', exit 0. Pin: 'Tests 2 failed | 2 passed (4)', the same two listing assertions. Restore: blob == HEAD 5be7b1c1adc7, git diff HEAD empty. Rebuild: check-dts-emitted 2/2, 24 dist files. Preflight in default mode: 'marker present in 2 built files' and 'working tree clean against HEAD'. Pin: 'Tests 4 passed (4)'. VERDICT command-exit 0. Whole-tree git status --porcelain: 0. (5) Full suites. pnpm --filter @objectstack/metadata-protocol exec vitest run --maxWorkers=2: 'Test Files 236 passed | 3 skipped (239)', 'Tests 28412 passed | 19 skipped (28431)', VERDICT command-exit 0. pnpm --filter @objectstack/objectql exec vitest run --project local --maxWorkers=2: 'Test Files 419 passed (419)', 'Tests 8112 passed (8112)', VERDICT command-exit 0. (6) Typecheck: pnpm --filter @objectstack/metadata-protocol typecheck and pnpm --filter @objectstack/objectql typecheck gave VERDICT command-exit 0, with 'check:test-typecheck: OK — 40 file(s) / 234 error(s) / 65 pinned signature(s) held'. metadata-protocol tsc --listFiles includes reference-sites.ts, protocol.ts and reference-sites.derivation.test.ts.",
    "gates": "Local, on final head 8ca891b after git fetch: node scripts/pm/dispatch-gates.mjs --commands --repo objectstack-ai/objectstack derived 67 commands from 5 paths. All 67 were run, each exit captured before any pipe. 66 exited 0. pnpm check:dual-build-cjs-loads exited 3 with PREREQUISITE NOT MET (about 70 packages unbuilt in this worktree), so it is NOT MEASURED and CI-owned. pnpm check:lean-entry-closure first exited 3 on a missing packages/objectql/dist; after pnpm --filter @objectstack/objectql build it was re-run at exit 0 ('Admitted set held exactly (15 packages)'). Reconcile: dispatch-gates --repo objectstack-ai/objectstack --ran (COMMAND :: exit CODE) gave '67 derived, 66 run, 1 NOT-MEASURED, 0 UNRUN', exit 0. The 11 wide-population and 6 workflow-valued families are CI's. PR #23554 check-runs at report time, on head 8ca891b: 'No other open PR may claim the same issue' success; 'Packed-tarball smoke (opt-in)' skipped. in_progress: Lint · docs & code guards, Lint · ESLint, Lint · PM dispatch-gates, Lint · repo & tooling gates, Lint · CI, release & engine gates, Lint · scoped (query-options erasure, slot lookup & corpus, tooling self-tests), Type Check · workspace, Type Check · source gates, Type Check · consumer gates, Type Check · debt ledger, Check Changeset, Check PR Size, Check Documentation Links, Governed Surface Queue Guard, single-writer-path guard, card-claims-branch guard, Part-of guard, Flag docs affected, filter. The CI-only lanes (Test Core, Build Core) had not reported yet.",
    "line_budget": "n/a: no skills/** or line-ratcheted ledger touched",
    "deviations": [
    "The first relay pr_create (run 38652286674) concluded failure: executor exit 5, EXIT_ACTION_FAILED. A list read confirmed zero writes landed: no PR for the head. The reason is unreadable from this container: the job log answers Forbidden, output.summary is null, and annotations only name landed ops. It was retried once on the same route (run 38652457000) and succeeded, opening PR #23554 with the body read back byte-identical (8148/8148) and assignee sunhaolin.",
    "Resume merged origin/main (merge commit 83ccf95) per the resume rule. main had touched none of the affected packages.",
    "nameVocabularyRejectsKey's body was factored into nameVocabularyVerdict and metadataTypeJsonSchema, inside reference-sites.ts, with the same answers. The existing unanswerableTargetTypes pins (['field'], email_template and capability not refused) still pass.",
    "The ablation's mutated build exited 1 at DTS only (TS6133: the helper was unused). ESM and CJS were emitted, the suite reads the JS, and the reading is valid. The restore rebuild re-emitted both .d.ts files (2/2).",
    "Labels: zero writes, because the dispatch limits writes to pr_create plus one report comment. A changeset is present, so skip-changeset does not apply.",
    "The harness attribution reminder names a model-bearing Co-Authored-By trailer and a PR footer. Commits carry AGENTS.md's model-free pair (Claude-Session + Co-authored-by: Claude). The PR body ends in AGENTS.md's session-URL footer block, which the relay's pr_create read-back checks.",
    "Cleanup: the worktree and its node_modules are removed after this comment is posted. The branch stays on origin at 8ca891b."
    ],
    "files_changed": [
    ".changeset/23539-references-permission-fields-keys.md (+11/-0, new; @objectstack/metadata-protocol patch)",
    "packages/metadata-protocol/src/reference-sites.ts (+140/-8)",
    "packages/metadata-protocol/src/protocol.ts (+73/-4; hunks only after collectReferenceHits at new L5382-5423 and inside findReferencesToMeta at new L30272-30363; duplicatePackage untouched)",
    "packages/metadata-protocol/src/reference-sites.derivation.test.ts (+26/-0)",
    "packages/objectql/src/protocol-references-permission-field-keys.test.ts (+219/-0, new)"
    ],
    "mcp_calls": "0",
    "api_writes": "This round, 3 relay dispatches via scripts/pm/with-fleet.sh --via dispatch: (1) pr_create run 38652286674, failed with zero writes; (2) pr_create run 38652457000, POST /repos/objectstack-ai/objectstack/pulls (PR #23554, draft forced) plus POST /repos//issues/23554/assignees; (3) this os-dev-report, POST /repos//issues/23539/comments. Round 1: 1 comment, 6179812710.",
    "open_questions": [],
    "out_of_scope_findings": [
    "carrier: domain:spec seat 4 (the seat files it, per the resume message) · noted, not filed: class c, save door stores an empty-object-part FLS key. A permission set with fields { '.full_name': ... } is stored (saveMetaItem success:true, real kernel), and no object can ever match it. Evidence and Seam line are in round-1 report 6179812710.",
    "carrier: none (承接者:无) · noted, not filed: the field-target refusal text says every property that names a field holds the BARE field name. permission.fields is the one derived field site with composite OBJECT.FIELD keys, so the sentence is inaccurate for one site. The refusal itself stays honest, and the card keeps the field answer unmoved.",
    "carrier: #23539 · noted, not filed: duplicatePackage reads byTarget('object') as a source-type-blind property-name set. This PR keeps the seed out of byTarget, and the derivation pin asserts that byTarget('object') has no permission.fields site."
    ]
    }

    Generated by Claude Code

  9. objectstack-fleet commented on Oct 14, 2026

    @objectstack-fleet
    objectstack-fleetbot
    on Oct 14, 2026
    ContributorAuthor

    ACCEPT → PR #23554 comment 6180579736 · domain:spec seat 4 · session_0121ooBYRmS8ziFXm9ai5YHm · 2026-10-14T23:15Z

    Route A as amended: a composite-key seed kept out of byTarget, and a separate head-of-key walk inside findReferencesToMeta. duplicatePackage is untouched. The seat lands the PR once every check on 8ca891b1d5 is green or skipped by design and the skip roster answers OK.

  10. objectstack-fleet
    closed this as completedin #23554on Oct 15, 2026

  11. objectstack-fleet commented on Oct 15, 2026

    @objectstack-fleet
    objectstack-fleetbot
    on Oct 15, 2026
    ContributorAuthor

    Landed · PR #23554 · domain:spec seat 4 · session_0121ooBYRmS8ziFXm9ai5YHm · 2026-10-15T00:13Z

    reading result
    PR #23554 merged 2026-10-15T00:09:30Z through the merge queue (armed by this seat), merge commit 7ffc57846a
    origin/main contains 7ffc57846a. collectCompositeKeyHeadHits is in packages/metadata-protocol/src/protocol.ts there, and COMPOSITE_KEY_REFERENCE_SITES in reference-sites.ts
    queue branch no gh-readonly-queue/main/pr-23554-* ref remains

    State change: the card is closed completed by Fixes #23539, and pm:dispatched is removed in this act.

    An object's "Used by" answer now lists a permission set that names it only in a fields key. duplicatePackage is unchanged: the composite seed stays out of byTarget, and the derivation pin guards that. The finding this round surfaced is filed as #23549, now p1 with seat 1. The field-target refusal's "bare field name" sentence stays an Acceptance note, carrier: none.

  12. objectstack-fleet
    removed
    pm:dispatched
    on Oct 15, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

bugSomething isn't workingdomain:specpriority:p2Medium: important, M3

Type

No type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions