## Why
Response item IDs and context window IDs use separate generator callbacks.
A shared allocator gives history and inference a single API for both.
## What changed
Use `ThreadContextIds` to allocate response item IDs, retained-context
revisions, and compaction window IDs. Move initial window selection into
`InitialHistory` so resumed threads reuse a valid window ID from their
rollout. Give rollout reconstruction its own allocator.
Assign hook prompt IDs through the session's normal response item
recording path.
## Testing
Move the prefixed UUIDv7 item ID test to the allocator crate and add
coverage for JSON escaping in Guardian's item ID token budget.
GitOrigin-RevId: 31232f50648382033506d1f86815ce7c1de78378
## What changed
Set `features.guardianv3.message_board_in_memory = true` to give each
new or cleared root thread an in-memory SQLite board. The housekeeper
and reviewer share the board until the last handle closes. The reviewer
can read evidence and update its own subscriptions. On-disk storage
remains the default.
With in-memory storage enabled, the attached reviewer is unavailable
on resumed roots because evidence intake starts at the end of history
and a fresh board has no earlier evidence.
## Testing
Exercise evidence catch-up and later post delivery with both storage
modes. Check that only the on-disk mode creates a database file, the
reviewer cannot update channels, and a cold resume skips evidence intake
in memory mode.
GitOrigin-RevId: 338f44720fca510b2baf40fd4b3a4b30245d4fc2
## Why
Guardian V3 uses fixed guidance for grouping evidence into board channels
and deciding when to retrieve more evidence. Longer routing instructions
also leave less room for evidence in the model's context window.
## What changed
Under `features.guardianv3`, add `housekeeper.routing_instructions` and
`reviewer.retrieval_instructions` to customize those strategies while
preserving the routing format and the reviewer's board and authorization
rules. Blank or omitted instructions use the built-in guidance.
Add an optional `max_instruction_tokens` cap that validates each custom
prompt separately using an estimated token count. Ignore both prompts and
the cap in project config. The `housekeeper.batch_timeout_ms` setting
controls the per-batch routing timeout and defaults to 90 seconds.
Size routing batches to fit the model's input window after accounting for
the prompt and response schema. Route the largest fitting prefix of whole
evidence records within the tool-output budget and leave the rest pending.
## Testing
Cover prompt validation, project config filtering, feature toggle merging,
and default retrieval guidance. Exercise custom prompts in model requests
and verify that routing reduces a batch under context pressure or stops
without advancing the cursor when even one record cannot fit.
GitOrigin-RevId: 83b3fcd23c680e90d2b2aea82a4a290633729658
## Why
A pending action can reach approval review before Guardian V3 has routed
the latest user messages to the reviewer's subscribed channels. That
can leave a revocation or renewed approval out of the evidence used to
decide the action.
## What changed
Attach the Guardian V3 reviewer to the synchronous approval path when
available. Start evidence catch-up when the action requests review,
then deliver queued evidence before submitting it for a verdict. If
preparation fails, report the reason and block the action.
Request strict structured output for board routing. Represent channel
description updates as a list, reject duplicate descriptions and empty
routes, and write each source once per distinct destination.
## Testing
Add app-server coverage for subscribed revocation and reapproval,
unavailable evidence, routing failures, and native review when V3 is
disabled or the thread is ephemeral. Add routing tests for duplicate
descriptions, duplicate destinations, and empty routes.
GitOrigin-RevId: 193802d39c120fbad9d77544ab57f4b5d02576b0
## Why
Guardian v3 reviewers can retrieve evidence from the live board, including
posts that arrive after a review begins. Fixed board views and bulk
subscription updates have no remaining callers outside tests.
https://github.com/openai/codex/commit/c2595ad3438311ce3070b1a7eef05cd0da8ad631
## What changed
Remove `BoardReadView`, its snapshot-specific tool bindings, and the
captured catalog and post cutoff from local board queries. Remove
`set_subscriptions` and keep individual subscription changes through
`set_subscription`.
Update the Guardian housekeeping test to check channel descriptions
through `list_channels`, and store captured Guardian evidence in an
owned vector.
GitOrigin-RevId: ede9e48abf4d009a2b4a82a437851657e41cf009
## Why
Guardian queued newly committed evidence for subscribed actors, but the
retained reviewer conversation did not consume that queue. A reviewer
could stay idle without receiving originals routed to it.
## What changed
Drain each actor's inbox when its reviewer becomes idle and before
reserving the conversation for a verdict. Read the queued posts by ID and
inject whole originals in source order within the reviewer's model budget,
without starting a model turn. Track delivery per conversation so a reused
reviewer receives each queued original once and a replacement can replay
the inbox.
## Testing
Cover delivery after the reviewer becomes idle, retention of queued
evidence after unsubscribe, and replay in a replacement conversation.
GitOrigin-RevId: ce83b98dc3eb2bae13ba0b7b5a82831a98b8aec4
## Why
A fixed evidence view hides posts routed after the review cutoff. Live
retrieval lets reviewers find those posts as they arrive while still
allowing intake to confirm that evidence through a requested history
position has been routed.
## What changed
Guardian retrieval tools search the live message board, including channels
the reviewer does not subscribe to. Subscription tools update only the
acting reviewer's subscriptions.
The housekeeper uses the subscribers present at each post's insertion to
queue one committed post per original in each recipient's inbox, even when
multiple subscribed channels match. It acknowledges catch-up requests
after routing and queuing evidence through the requested history position.
## Testing
Cover catch-up ordering, delivery based on subscription state at insertion,
live search across unsubscribed channels, and caller-scoped subscription
changes.
GitOrigin-RevId: 28b447d936f19ea68bd3c3adb297d473d65173c0
## What changed
Give the root agent and subagents with a board identity their own Guardian
reviewer reservation. Each new reviewer thread starts with empty history
and uses the acting agent's live message-board retrieval tools.
Keep one reservation active per agent and retain the conversation only
when the caller marks it reusable. Reuse requires matching reviewer settings
and history-reset version. Dropping a reservation retires its thread.
## Testing
Add an integration test that checks thread reuse after a reusable
disposition, replacement after an unfinished reservation is dropped, and
exclusion of legacy subagents without a board identity.
GitOrigin-RevId: c7bccf4651ad4e04c42d9aee1d9e8de24b5b48df
## Why
A captured board view limits evidence searches to the review's cutoff. A
verdict reviewer needs to be able to check for later evidence, including
restrictions or revocations, before deciding on an action.
## What changed
Add a live `guardian_board` binding that lets reviewers search and read
any Guardian evidence channel, including channels they do not follow.
Search also matches text inside JSON-encoded tool arguments.
Reviewers can list their own subscriptions with `subscribed_channels`
and subscribe or unsubscribe by channel. Subscriptions control delivery
of future evidence. Earlier posts are available through search and read
tools.
Let review setup and prewarming select guidance for the installed
retrieval tools. The live board guidance explains how to find original
user messages and assess their authority. Share the provider-scoped
reviewer header setup through the core Guardian review API.
## Testing
Cover searches beyond the captured cutoff and in unsubscribed channels,
paginated listing of the reviewer's own subscriptions, and channel-only
subscription changes in both live and captured views.
GitOrigin-RevId: c2e1c6c849346cd3d6e316bcf17e44260fcd4c8e
## Why
Overlapping reviews need board views that match their captured history boundaries, so later evidence does not appear in an earlier review.
## What changed
- Add `RootEvidence::freeze` to capture and enqueue history in source order, returning a `ReviewCutoff` with the source position and reviewer subscriptions.
- Open each review's board view after routing its captured evidence and before routing the next capture.
- Retry transient history failures on subsequent activity, preserve evidence validation failures for later reviews, and skip incomplete delivered messages with a warning once per accepted capture.
## Testing
Add tests for overlapping reviews with a later revocation, retryable history failures, persistent evidence rejection, and incomplete deliveries across activity wakes and freezes.
GitOrigin-RevId: b87040e99bf0e56626d25a8acdd036cf3eed32ce
## What changed
Use `HousekeeperSampler` for evidence routing instead of starting a separate review-session thread. The housekeeper now owns a tool-free model conversation, retains completed routing batches, and evicts whole batches under context-window pressure.
Commit sampled history only after all board writes succeed and the evidence cursor advances. Stop the worker after a routing error or an interrupted routing future. Cancel routing when the account changes while allowing accepted appends to finish before shutdown.
## Testing
Extend housekeeping integration coverage for retained history, whole-batch eviction, transient retries, retry exhaustion with `Retry-After`, HTTP fallback, credential renewal, command-token rotation, and shutdown waiting for accepted evidence appends.
GitOrigin-RevId: 316ab1b90529e2a5874e617855d38c41f232872a
## What changed
Add a tool-free Responses sampler with persistent conversation history and whole-batch eviction to stay within the model's input budget. Parse completed routing replies and retain their output through an explicit commit step.
Add WebSocket connection reuse, HTTP fallback, bounded retries, authentication recovery, account-change checks, and cancellation of unfinished responses.
Register the new modules without connecting them to the existing housekeeper routing flow yet.
GitOrigin-RevId: 2a0ff557a521763ab984249ec6594ae7ce09d59e
## What changed
- Install Guardian V3 with a local message-board host when the server enables `features.guardianv3.enabled`, independently of thread overrides.
- Clear inherited user, thread, and project instructions from Guardian V3 private threads and skip repository instruction discovery.
## Testing
Add integration coverage for routing the first live message while the actor is still sampling, for both new and resumed threads. Verify that resumed history is skipped, `AGENTS.md` reaches only the actor, and thread overrides cannot enable or disable server-level housekeeping.
GitOrigin-RevId: 3f3ae1be404c9e17b73eea12c95f5a95c96052cf
## What changed
- Add an `install` entry point that registers thread and turn lifecycle hooks for Guardian housekeeping. Coalesce user-message and turn-end wakes into one worker per eligible root thread, flushing history before routing durable evidence in order.
- Run the housekeeper in an isolated, ephemeral model thread with no tools. Cancel and join its tracked tasks on root shutdown, waiting for accepted evidence appends to finish.
- Give each routing batch a 90-second deadline, make channel listing and updates cancellable, and redact model response content from JSON parsing errors.
## Testing
Extend housekeeping tests to cover stalled channel creation, redacted model errors, and shutdown waiting for an accepted evidence append without leaving a persisted housekeeper rollout.
GitOrigin-RevId: 519b539c0613c964bac7b18eb84e684f5e105640
## What changed
Allow the Guardian housekeeper to route evidence to existing channels without repeating their descriptions. Request descriptions only for new destinations or changed descriptions, and skip channel updates when a returned description matches the catalog.
New destinations still require descriptions, and description updates remain limited to routed channels.
## Testing
Extend housekeeping tests to cover omitted and repeated descriptions without extra channel writes, preserved descriptions, and rejection of missing new-channel descriptions or updates to unrouted channels before any writes.
GitOrigin-RevId: 3ab00fc17263d68d15e68f9d17ab31fc6f43f551
## What changed
Expose history initialization and the effective history mode to thread-ready contributors. Add `EvidenceCursor::at_thread_ready` with paginated-history validation:
- Start new or cleared threads at position zero without requiring an existing rollout.
- Start resumed, forked, or supplied-history threads at the durable tail to avoid backfilling existing evidence.
- Skip the initial `SessionMeta` header when capturing evidence from position zero.
## Testing
Extend evidence tests to cover initialization before a rollout exists, capture of the first original record, rejection of resume without a rollout, and exclusion of existing records on resume.
GitOrigin-RevId: 32aa8e2afb71eb4e882b5f1fb107ac8616b33c8e
## Why
Guardian's retrieval, channel selection, and evidence delivery depend on concrete local message-board types. Expose the required operations through shared interfaces so these consumers can use trait-based board handles.
## What changed
- Introduce `BoardReadView` for captured catalogs and post boundaries, with read operations and JSON search configuration but no mutation methods.
- Extend `AgentMessageBoard` with namespace identification, read-view capture, and channel subscription operations, and implement them for the local board.
- Add `message_board_read_tools` to bind the five read-only tools to a captured view.
- Migrate Guardian consumers to `AgentMessageBoard` and `BoardReadView`, retaining Guardian namespace checks and captured read boundaries.
## Testing
Adapt read-view and Guardian retrieval tests to the trait interfaces. Add assertions for the five retrieval tools and rejection of ordinary boards by retrieval binding, subscription capture, and delivery planning.
GitOrigin-RevId: f8c5bc65d8297f6df6e3b02d6c09006bcaea4b1a
## What changed
Add `DeliveryPlan::deliver` to supply selected originals as `guardian_original_evidence` context in source order. Read whole posts one board page at a time and batch them within the model's tool-output budget, rejecting originals that cannot fit intact.
Record delivered ordinals only after the admission callback succeeds, and advance channel positions when the plan completes. Preserve successful admissions across partial plans so retries and overlapping channel selections do not duplicate evidence.
Add `LocalBoardReadView::read_post_texts` to retrieve posts in the requested order while enforcing the captured board boundary.
## Testing
Add retrieval tests covering ordered delivery into a retained conversation without starting inference, deduplication, failed admission, oversized-original retries, older evidence from newly selected channels, captured-view boundaries, and paged reads.
GitOrigin-RevId: b2f03e039c6677e0b94dceaa7ed75a966edf81eb
## What changed
Add `DeliveryState` to plan unique evidence originals from selected Guardian board channels in source order. Read request IDs in pages by insertion sequence within a captured board view, and reject evidence beyond the review's source-history cutoff.
Share evidence request-ID encoding between the housekeeper and delivery planner, and expose the source cutoff on `EvidenceSnapshot`.
## Testing
Add retrieval tests covering deduplication across channels, source ordering across storage pages, exclusion of unselected channels and posts appended after capture, and rejection of evidence beyond the source cutoff.
GitOrigin-RevId: fc181a52b46137538bb94a5bba1c1d6bf1bccf32
## What changed
Add `ReviewerSubscriptions::prepare` to select evidence channels using the review task, current subscriptions, and captured catalog. Validate unique channel names before atomically replacing the caller's subscriptions within that catalog, preserving subscriptions to channels outside it.
Share typed sampling between selection and housekeeper routing, enforce input byte budgets, and expose `OperationResult` so callers can honor runtime reuse or rollback requirements. Selection observes cancellation and deadlines before saving membership.
## Testing
Add integration coverage for subscription replacement, caller isolation, preservation of later channels, cancellation before sampling, and rejection of selections outside the captured catalog with a rollback disposition.
GitOrigin-RevId: fcab0de88ac862e21cdf3c780c780385e2543604
## What changed
Add `ReviewerSubscriptions` to capture a Guardian catalog and expose self-only `subscribe` and `unsubscribe` tools. Restrict requests to channels in that catalog and return captured post IDs so later evidence stays outside the review boundary. Read current membership from storage and filter it through the captured catalog.
Share message-board tool size limits and query the latest channel post directly when updating subscriptions.
## Testing
Add coverage for caller-only subscription changes, rejection of later channels and explicit agent targets, captured post IDs, and unchanged subscriptions for other agents. Extend message-board tests to check unsubscribe post IDs and reject missing channels.
GitOrigin-RevId: 48bf4270d6079f7e436096175c67bb211ccbc248
## What changed
- Expose `guardian_board` tools to list channels and threads, search posts, and read evidence through a captured read-only view. Require a Guardian storage namespace and keep reads within the review's fixed cutoff, including searches across unsubscribed channels.
- Support literal, case-insensitive searches over decoded JSON fields and nested function-call arguments so quotes, Windows paths, and Unicode escapes do not hide matching evidence.
- Strip response-envelope harness metadata from captured evidence except for `user_input_order`, preserving message payloads and source ordering.
## Testing
Add retrieval coverage for unsubscribed channels, fixed cutoffs, and escaped JSON searches. Update evidence tests to verify serialized payloads retain source order while excluding other harness metadata.
GitOrigin-RevId: 3128e4a2d59d1f71f2cce0bc6b1fbdf7e2aba287
## What changed
Add a `Housekeeper` that uses an isolated reviewer runtime to route captured original evidence into topic channels and update their discovery descriptions. Keep originals intact and batch inputs within the selected model's tool-output budget.
Validate routing responses before writing to the message board. Advance the evidence cursor only after all writes succeed, stop the worker after an error, and return the runtime disposition so callers can manage session reuse.
## Testing
Add integration coverage for routing originals to multiple channels, updating descriptions, preserving capture boundaries, batching within the input budget, and leaving the cursor unchanged on write failure.
GitOrigin-RevId: 670508c7ecf2ddffdd2946f1b25dd7e071318c5a
## Why
Evidence capture previously reloaded the full history without tracking consumer progress. Live consumers need to read new originals from activation onward and acknowledge them after downstream writes succeed.
## What changed
- Add an in-memory `EvidenceCursor` that starts at the current complete tail without backfilling, captures a fixed unread range, and supports partial acknowledgements. Reads leave progress unchanged.
- Add `ThreadStore::read_original_history` to read original records and their source positions through a fixed cutoff. Local storage supports append-only paginated root histories, including compressed rollouts, and defers unfinished records.
- Reject unsupported histories and invalid ranges, and retain source ordinals as evidence IDs.
## Testing
Add coverage for partial records, compressed suffixes, fixed cutoffs, mismatched thread positions, and truncation. Extend evidence tests to cover activation without backfill, partial acknowledgements, later appends, stale captures, and unchanged cursor progress after capture errors.
GitOrigin-RevId: c7a13749fedc4f44fadddc9dd9875043bccebd95
## What changed
Add `EvidenceSnapshot::capture` to read original evidence from `ThreadStore` into an immutable snapshot with root-history ordinals. Preserve accepted user messages, assistant messages, tool activity, verified answers, delivered assistant messages, and host-annotated approvals without changing their contents or provenance.
Exclude generated context and compaction output. Reject forked or reverted histories, inherited or ambiguously attributed user messages, and incomplete retained evidence. Capture currently supports append-only root histories and requires the host to flush pending originals first.
## Testing
Add an integration test covering content preservation, generated-context filtering, approval provenance, isolation from later appends, and rejection of ambiguous user messages.
GitOrigin-RevId: f464a2714c2678bcffe7abf0a8224196bfe48692
## What changed
Register Guardian evidence-board cleanup with the app server's thread store when Guardian V3 is enabled. Leave Guardian storage untouched when the feature is disabled, and report cleanup failures as thread-store errors.
## Testing
Extend deletion tests to cover feature gating, preservation of unrelated boards and a surviving main thread's board, and corrupt-database recovery. Expand message-board tests to verify namespace isolation and persistent deletion markers that prevent later writes.
GitOrigin-RevId: d9119d22607769c92448d0f3cacbbfd03fbdf943
## What changed
Add the `codex-guardian-v3` crate to the Cargo workspace with a Bazel target. Expose helpers to configure the `guardian` storage namespace and construct a `LocalBoard` for the owning main thread's session ID.
GitOrigin-RevId: 3314711475a7024380a3f0beb8f5ddfc4ff19723