feat(core): Moirai F1 common contracts (judgment, context projection, persona schema, catalog/policy, store)

[thisisjun786]

Summary

This PR defines the Moirai F1 common contracts in lina-core: judgment records and validation, context read projections, PersonaSchema derivation, the personal option catalog and arbitration policy, and an independent SQLite judgment ledger.

It is not an installed Moirai runtime or live-model qualification. The PR also contains separately scoped runtime test-fixture corrections; they are not evidence of F2 product integration.

Contract scope

• Canonical options bind normalized arguments, actor/scope, preconditions and effect semantics. Generated keys fit the shared identifier bound.
• Frozen snapshots identify objective profiles, policy, source revisions and scope. The judgment ledger verifies stored digests, indexed metadata and cross-record provenance.
• Each logical source kind/ID has exactly one frozen revision; contradictory revisions are rejected.
• One immutable objective ID/revision cannot occupy multiple module slots, even with different claimed digests. Snapshot and selection parsers enforce this independently of stored-profile lookup.
• Action arbitration preserves module recommendations, exclusions, ordering and concessions. Selection specifications are replayed from frozen candidate inputs rather than trusted independently.
• Context projections distinguish instruction revision from working-state revision and validate restored source references.
• Persona derivation preserves LIFE sharing allowlists, manual/axis locks and participant ownership. Its builder does not allocate authoritative current schema revisions.

Review resolution

The current review pass verifies prior fixes and addresses the remaining defects with failing-first regressions:

• Intention acceptance, deadline and transition timestamps use exact canonical UTC strings; invalid values are rejected without rewriting their evidence.
• Intention history cannot run backward before acceptance or an earlier transition; equal timestamps remain valid.
• Pure intention transitions validate the existing record before reading its status, revision, acceptance or history; malformed restored records cannot produce invalid next records.
• insufficient_evidence is the initial declared unavailability code. Arbitrary reason text cannot remove an objective from ordering; detailed diagnosis remains in existing text/evidence fields.
• Atropos commitment exclusions require nonempty attribution verified through the Host's intention owner. Unattributed claims retain ranked opposition; same-target suspend/cancel proposals keep their ranking exemption, not execution permission.
• Clotho's infeasible label is not a hard veto. Factual prerequisite failures use Host eligibility.
• An explicit Host rejection without a reason is distinguished from missing eligibility.
• Raw option arguments are length-bounded before normalization or omission, including whitespace-only values; bounded empty values still omit.
• New round/resolution writes check the owner's scoped intention revision; resolutions recheck active objective references. Historical receipts remain readable after legitimate later changes.
• Objective activation cannot restore the full references of a stale open round in the same agent/scope. This blocks A→B→A revival without freezing all profile changes or changing serialized snapshots; a fresh objective revision can start a new round.
• Per-agent/scope round sequences must increase; gaps are allowed.
• Candidate-free dialogue has an explicit schema-v2 record and reconstructible DialogueJudgmentRef. It preserves the three original assessments, frozen objective/snapshot/policy references and textual synthesis. Store write and reopen validation reject action candidates/specifications on dialogue and incomplete or mismatched provenance.
• Dialogue request/source hashes must match dialogueSource frozen by the Host in the round snapshot before assessment. Invented or rehashed record digests, and dialogue without those expected fingerprints, are rejected. Source owners retain the raw data; legacy action snapshots omit the optional field unchanged.
• An incomplete dialogue may remain held or end deferred after budget exhaustion, but cannot be recorded as resolved or produce an accepted-dialogue reference.
• Exact legacy action-v1 record bytes and digests remain unchanged. The physical schema is unchanged; older readers intentionally reject the new dialogue JSON capability rather than rewriting it.

Deliberate F1 boundaries

• Projection continuity (3995355370): the pure builder accepts a validated previous projection. F2 context/source assembly must persist and restore it per logical conversation/source stream; no production continuity owner is installed by this PR.
• Persona ownership (3995355387): AgentStore is the designated authoritative schema owner. F1 defines derivation/serialization; existing Behavior/projection migration is F4. The caller-supplied builder revision is not an authoritative allocator.
• Decisions, RNG and outbox (3995956610): F1 persists policy inputs/results and selection specifications. It does not persist a sampled decision, RNG advance or dispatchable outbox. Those atomic execution and recovery boundaries belong to F2.
• User cancellation authority (3999164940): F1 has no intention-bound, owner-verified cancellation approval path. New cancellation of adopted/active/suspended user commitments is therefore rejected; acceptance refs, catalog authority/confirmation strings and ranking results cannot authorize it. Historical v1 ledgers remain readable and terminal, and unadopted proposals may still be withdrawn. Real authority integration remains F2, not a caller-controlled approval flag.
• Schema additions (3995355563): the ledger intentionally rejects undeclared tables, indexes, views and triggers. Future owner-controlled additions require recognized, versioned schema evolution.
• Runtime fixture scope (3995355408): the archive-cap test seeds the first 256 terminal archive identities and exercises the real 257th boundary, reopening and replay. It does not claim to test 257 repeated archive writes.
• Restart fixture collection (3998853110): the owner is closed correctly. Pinned Bun 1.4.0 node:sqlite defers native statement cleanup until collection; in-process test replacement therefore clears the old connection before the private-copy audit. The call-through close spy triggers that audit boundary deterministically. An independent cold child-process exit and fresh protected HTTP read succeeds without GC or a close spy. This does not establish in-process hot-replacement support, and no production GC or relaxed file-audit check is introduced.
• Persona restart fixture: a hosted run exposed the default 5-second test timeout. The fixture now uses unchanged real prompt/preset bytes without unrelated shipped-avatar assets, removing repeated migration I/O. All three restarts and original assertions remain; no timeout increase or retry was added.

Phase authority: implementation roadmap and judgment contract.

Verification

• Full bun test at ff241e6: 4526 passed, 47 existing opt-in skips, 0 failed, 556 files. No skip was added by this review pass.
• Focused judgment scope: 606 tests passed; isolated Persona/runtime scope: 48 tests passed. Direct SQLite, real source-owner provenance and Fleet restart surfaces passed. Current-head aggregate results are shown in GitHub checks.
• bun run typecheck: root and browser TypeScript passed.
• bun run lint: passed. The one new optional-chain warning was fixed; 27 pre-existing warnings remain in untouched files.
• bun run ci:build: actual runtime assets built and read-only CLI paths smoke passed without creating application state.
• Independent public-API/SQLite runs proved unverified claims cannot veto, verified commitments remain protected, accepted target changes retain exemptions, stale writes and backward sequences reject, and resolved dialogue reopens with 3 assessments / 0 candidate rows / 0 selection specifications.
• Independent child-process cold restart returned protected HTTP 200 with the exact persisted V2 binding, without GC or close instrumentation.

Review-fix commits:

• 46b4b77 require canonical intention timestamps.
• 7eac3fe verify arbitration evidence and currentness.
• ebd7af9 persist candidate-free dialogue judgments.
• 5484813 validate source intentions before transitions.
• fb15519 close dialogue and snapshot boundary gaps.
• ee47070 enforce intention chronology and raw argument bounds.
• be29476 isolate Persona restart resources without increasing timeouts.
• cf3a46e reject cross-module objective aliases.
• fdce947 reject unverified live user cancellation while preserving historical records.
• 0785eda prevent objective reactivation from reviving stale open rounds.
• ff241e6 bind dialogue digests to Host-frozen source provenance.

Current-head GitHub results are available on the PR Checks tab.

No PR merge, release, deployment, paid model call or production-data operation is part of this review-resolution pass.

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 38bbb7a6-edc5-49dc-816e-82ccb60c65d9

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-09-13T09:27:01.087664Z	ff241e6	New commits
🔒 Security Review	✅ Completed 2026-09-12T04:34:35.148187Z	8cf0198	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

lina/packages/lina-core/src/agents/judgment-store.ts

Lines 75 to 77 in 0785eda

	if (!candidates || !set) {
	if (resolution.status !== "held" || selection !== null)
	throw Error(!set ? "incomplete assessment set" : "missing candidate set");

Persist budget-exhausted action rounds as deferred

In the action-v1 JudgmentStore path, when the evaluation budget expires before all three assessments or candidate closure, the Moirai contract requires a terminal deferred round without advancing selection RNG or outbox work, but this condition rejects every incomplete resolution unless its status is held. The caller therefore cannot durably record the required budget-exhaustion outcome and must either misclassify it or leave the round open; represent incomplete-assessment provenance and permit a validated deferred receipt.

AGENTS.md reference: AGENTS.md:L4-L4

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

💡 Codex Review

lina/packages/lina-core/src/agents/judgment-validation.ts

Line 519 in ff241e6

completeText: boundedText(row["completeText"], "complete text"),

Preserve complete module opinions beyond 1,000 characters

When a module produces an opinion longer than 1,000 UTF-16 characters, this default boundedText call rejects the entire assessment because MAX_TEXT is only 1,000. The Moirai contract explicitly defines completeText as the untruncated module opinion (docs/plans/platform/016_neural_preference_contract.md:73), and ordinary model judgments can readily exceed this limit, preventing the Host from recording or synthesizing an otherwise valid round; use a dedicated bound sized for complete model output rather than the legacy agent-field cap.

AGENTS.md reference: AGENTS.md:L4-L4

---

lina/packages/lina-core/src/agents/judgment-store.ts

Lines 75 to 77 in ff241e6

	if (!candidates || !set) {
	if (resolution.status !== "held" || selection !== null)
	throw Error(!set ? "incomplete assessment set" : "missing candidate set");

Allow exhausted action rounds to be recorded as deferred

For an action round whose deadline or evaluation budget expires before the candidate set or all three assessments are available, this branch accepts only held and rejects the required deferred terminal record. The declared contract says pre-selection budget exhaustion ends the round as deferred (docs/plans/platform/016_neural_preference_contract.md:89), so the Host cannot durably close such a round and instead leaves it open or misclassifies it as recoverably held; permit a deferred failure receipt when the missing evidence is attributed to exhausted budget.

AGENTS.md reference: AGENTS.md:L4-L4

---

lina/packages/lina-core/src/agents/persona-schema.ts

Lines 283 to 286 in ff241e6

	if (identity === null) return null;
	for (const profile of parseIdentityPolicy(identity).profiles)
	if (profile.agentId === agentId) return profile;
	return null;

Reject identity snapshots missing the selected agent

When a non-null identity snapshot contains policies only for other agents, this helper silently returns null, causing derivation to emit every projected dimension unlocked and omit sourceIdentity. The existing LIFE owner rejects exactly this mismatch as Missing LIFE identity policy in world/views.ts:251-252; accepting it here lets a stale or cross-agent snapshot erase manual/per-axis restrictions, so distinguish an intentionally absent identity input from a supplied snapshot that lacks the selected agent.

AGENTS.md reference: AGENTS.md:L4-L4

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".