## Why
Guardian stores a `SecurityRiskScore` before advancing its scored-tool-call
counter. Waiting only for the score can therefore let the test continue while
the asynchronous sampler is still updating its progress state.
## What changed
Wait for `latest_scored_tool_call` to advance before exercising Guardian v2
approval decisions, and reuse the same progress tracker for the subsequent
assertions.
GitOrigin-RevId: 3d92c9a3ccf52f4f43a2454aac60d2479258c1c8
## What changed
- Track the latest tool call and latest scored tool call for each Guardian v2 thread.
- Skip approval review when the score lags by more than `max_tool_call_lag` tool calls.
- Add `max_tool_call_lag` to the Guardian v2 configuration and schema, with a default of three.
## Testing
- Cover approval review at, above, and after recovering from the configured lag limit.
- Cover deserialization of `max_tool_call_lag`.
GitOrigin-RevId: fd870e143d95ee76491180e2aed742717248ff78
## Why
Guardian sampling can encounter temporary connection or service failures while
opening a WebSocket, starting a request, or consuming its response stream.
## What changed
- Retry transient transport, streaming, rate-limit, and server failures up to
two times across the full sampling flow.
- Recover from unauthorized responses through the configured auth manager and
discard pooled connections that may contain stale credentials.
- Allow sampler initialization to continue when its initial WebSocket prewarm
fails, so the first sample can establish a fresh connection.
## Testing
Added coverage for recovery after prewarm failures and `503` responses, and for
stopping once the transient retry budget is exhausted.
GitOrigin-RevId: af74c2df95c4241184ca641ef4418db6aa4ab695
## What changed
- Add optional `guardian_v2` model-message defaults for classifier instructions,
review thresholds, reasoning effort, transcript selection and limits, and
action and compaction token limits.
- Apply those defaults when Guardian v2 samples a tool call, while preserving
explicit `[features.guardianv2]` settings over catalog values.
- Preserve the new catalog configuration when applying model overrides.
## Testing
- Cover model-catalog serialization and model override behavior.
- Verify that sampling uses catalog defaults while retaining local overrides.
GitOrigin-RevId: 4ae9ec235c647acbca05a20b7c4b87a4e1331161
## What changed
- Add the `features.guardianv2.transcript.include_images` opt-in setting.
- Include recent images from user and assistant messages and configured tool outputs in Guardian v2 sampling requests.
- Bound image context to the four most recent images and 8 MiB total, and omit image detail hints when constructing the sampling request.
## Testing
- Cover configuration parsing and image forwarding from messages and tool outputs.
GitOrigin-RevId: b0ae6b6182535fd084ea103ed01dc3c803d7d1f7
## What changed
- Add `max_parent_compaction_tokens`, defaulting to 25,000 tokens, to the Guardian v2 feature configuration.
- Reuse the latest encrypted parent compaction only when its complete serialized item fits within the configured limit.
- Fail closed with the maximum action-risk score when the latest encrypted compaction is oversized, instead of sampling with older context.
## Testing
- Cover exact-boundary and oversized compaction items, including oversized passthrough metadata.
- Verify that an oversized latest compaction bypasses sampling and prevents approval.
GitOrigin-RevId: 0b308e7bec0f41d46433dfdf656948cee5cba667
## What changed
- Allow `features.guardianv2` to remain a boolean toggle or specify classifier instructions, the review threshold, reasoning effort, and action and instruction token limits.
- Add transcript controls for included sources, per-entry and total token budgets, and the number of recent non-user entries.
- Validate configured ranges and relationships, expose them in the generated config schema, and apply the resolved settings throughout Guardian v2 classification and approval review.
## Testing
- Cover boolean compatibility, configuration parsing and boundaries, config-manager rejection, and the resulting classifier request and approval decision.
GitOrigin-RevId: dc018b53b782e2d6b4d6a795cca4945130a71678
## What changed
- Include the resolved Guardian security policy in the v2 classifier's developer instructions.
- Prefer an explicit `guardian_policy_config`, then the reviewer model's catalog policy, and finally the bundled policy.
- Limit the combined classifier instructions and policy to 10,000 tokens.
## Testing
- Cover configured and catalog policies, policy truncation, and the bundled-policy fallback.
GitOrigin-RevId: 1b84d70081dd924e214636b4b470d57a63d7cb88
## Why
Concurrent Guardian samples can finish out of order. A stale result rejected by
the thread's latest-score update could still be appended to the rollout.
## What changed
Return without persisting a risk score when `insert_if` rejects it as older than
the score already stored on the thread.
GitOrigin-RevId: 7b2c895742753a19dfe48fd7f8ecd912d3cd3af6
## What changed
- Expand the Guardian sampling pool from 8 to 16 WebSocket connections.
- When the pool is full, supersede the oldest request that has already produced a score before replacing an unfinished classification. Treat superseded classifications as a no-op in the extension.
- Stop response WebSocket work when its event consumer is dropped, including while waiting for the connection lock or draining a completed sample.
- Allow retries across both initially warmed connections for retryable stream failures.
## Testing
- Add a concurrent sampler test that fills the pool and verifies scored drains are replaced before an unfinished classification.
GitOrigin-RevId: 9e4df7b01516094726ff3e11878f8ca742d21b03
## What changed
- Give approval review contributors the first opportunity to resolve an action when automatic review is not required for the model.
- Return an extension's decision directly and fall back to Guardian when no extension claims the review.
- Update Guardian V2 to approve actions below its risk threshold through the review contributor API, and remove the extension approval-policy API.
## Testing
- Verify that an extension approval bypasses both the Guardian model call and the user approval prompt.
- Cover Guardian V2 review decisions for low-risk, high-risk, and disabled threads.
GitOrigin-RevId: 5be56941143aff6f358b75b5f8101b7c047147ae
## Why
Oversized tool arguments could make Guardian V2's model-visible action unbounded.
## What changed
- Cap serialized actions at 10,000 tokens while preserving the trusted tool name and `call_id`.
- Truncate nested string values evenly across the available budget.
- Omit lower-priority fields when the JSON structure alone exceeds the limit, while retaining useful action fields when they fit.
## Testing
Add coverage for oversized nested strings, arrays, and object keys, including spoofed identity fields and omission-marker collisions.
GitOrigin-RevId: 13f7dc93da3761db18a34087225196318a3b8098
## Why
Concurrent Guardian samples can finish out of order, allowing an older sample to
replace the thread's newer in-memory security risk score.
## What changed
- Record when each `SecurityRiskScore` sample starts while preserving compatibility
with records that have no timestamp.
- Add `ExtensionData::insert_if` so checking and replacing a typed attachment happen
under the same lock.
- Update Guardian V2 to replace the thread's risk score only when the completed sample
is newer than the stored one.
## Testing
- Verify concurrent conditional inserts retain the newest value.
- Verify Guardian-generated risk scores include a sampling timestamp.
GitOrigin-RevId: 6543a4e5497da4c623331050789e91613f931c4b
## What changed
- Add a thread-scoped extension API for requesting automatic review, resolving
multiple contributors to the strongest requirement.
- Have Guardian v2 request automatic review when it is enabled and the latest
`action_risk` score is at least `0.8`.
- Honor extension review requirements across tool approvals, MCP elicitation,
and remote execution while preserving host configuration and managed-policy
constraints.
## Testing
- Cover contributor registration and requirement resolution.
- Verify Guardian v2 requests review at the threshold, returns to the default
below it, and does not affect threads where the extension is disabled.
- Exercise extension-requested strict review with a remote model override.
GitOrigin-RevId: 6329fc23585d5cc7173c39bb9952c3109ec3884b
## What changed
- Replace the single `category` and `score` fields in `SecurityRiskScore` with a
deterministic map of category names to scores.
- Validate the full Guardian V2 classifier response before storing it as one
thread extension value and, for non-ephemeral threads, one rollout item.
- Cover serialization and rollout loading with snapshots containing multiple
score categories.
GitOrigin-RevId: 24525de691f96bd34ccfe04acab2f01aca309837
## Why
Guardian V2 classification requests should retain compacted conversation context only when the parent and classifier model configurations can share the encrypted compaction.
## What changed
- Resolve the classifier model's compaction compatibility hash when initializing its sampler.
- Add the latest encrypted parent compaction to classification input when both models have the same non-empty compatibility hash.
- Ignore missing, empty, or unidentified compactions, and omit parent context when compatibility cannot be established.
## Testing
- Cover latest-compaction selection and invalid compaction data.
- Cover matching, mismatched, missing, empty, and changed compatibility hashes.
GitOrigin-RevId: f0fcb09fd028cd2f397fdbc3d541b4bc464b0beb
## What changed
- Replace the single byte limit with token-derived per-entry and separate message and tool transcript budgets.
- Preserve the first and latest user messages, then fill the remaining budgets with recent user, assistant, and tool context while limiting retained non-user entries.
- Truncate oversized entries in the middle without splitting UTF-8 characters, and report the approximate omitted token count.
- Exclude reasoning from the default transcript sources while retaining support when it is explicitly configured.
## Testing
- Cover oversized message and tool-result truncation, user-message retention, separate tool budgeting, configured reasoning, and the resulting Guardian sampling request.
GitOrigin-RevId: aaeaa5fb5d902b5a1581a8a819bf34724bde794b
## What changed
- Represent Guardian sampling input as an ordered list of text entries.
- Send each transcript entry and approval-request delimiter as a separate input
content item while preserving the existing transcript byte limit.
## Testing
- Update transcript and sampler tests to verify entry boundaries, multi-item
request serialization, and bounded transcript behavior.
GitOrigin-RevId: 433e2e6ca18ce9288025715f5956295bfe852e98
## What changed
- Register Guardian V2 for thread startup as well as tool-call lifecycle events.
- When `guardianv2` is enabled, create and store a thread-local Luna sampler using the thread's model provider, authentication policy, session metadata, originator, and service tier.
- Emit a warning if sampler initialization fails, and skip tool classification when the thread has no sampler.
## Testing
- Update the extension test to enable `guardianv2`, run thread startup, and verify tool-call sampling through the initialized connection.
GitOrigin-RevId: 5ab732e85c9827495c67f57117eb5be450a7f90e
## What changed
- Parse and validate the structured risk scores returned for each classified tool call.
- Publish each score through thread extension data and persist it in rollout history for non-ephemeral threads.
- Report parsing, thread lookup, validation, and persistence failures through the extension warning sink.
## Testing
- Extend the Guardian V2 tool-classification test to verify that the sampled `action_risk` score is available from thread extension data and written to rollout history.
GitOrigin-RevId: 46614fe0d686736dfb1f09398c6fcd848c7d8b2c
## What changed
- Opt Guardian V2 Luna sampler WebSocket connections into responses lite.
- Send `reasoning.context` as `all_turns` on sampling requests.
## Testing
- Extend sampler and extension tests to verify the connection header and reasoning context.
GitOrigin-RevId: 4b0263c01b8259ed5268c222dcf7559912baba8c
## Why
Guardian V2 needs the requested action and its conversation context to assess
risk; a tool name and call ID alone do not describe what the tool will do.
## What changed
- Expose the original, pre-hook `ToolPayload` to tool lifecycle contributors.
- Build Guardian V2 classifier input from the bounded conversation transcript
and a structured planned-action object containing the tool name and arguments.
- Warn and skip classification if the planned action cannot be serialized.
## Testing
- Verify lifecycle contributors receive tool arguments.
- Verify Guardian V2 sends transcript history and planned-action JSON to the
classifier.
GitOrigin-RevId: 28266ef7c1d2267ba705fcfbd9be4ef48e64e104
## What changed
- Add a configurable renderer that converts conversation messages, agent
messages, tool calls and outputs, and reasoning into a numbered plaintext
transcript.
- Keep only relevant readable content, including manual-approval developer
messages, while omitting media payloads, encrypted content, and unsupported
response items.
- Limit transcripts to the most recent 320 KiB without splitting UTF-8
characters.
## Testing
- Cover source selection, content filtering, tool-name correlation, and bounded
UTF-8 truncation with unit tests.
GitOrigin-RevId: e2c6535f4a895862ca744258399fb36fc92e42c6
## What changed
- Register a tool lifecycle contributor when Guardian V2 is installed.
- On tool start, asynchronously ask the existing Luna sampler for a low-effort
`action_risk` score constrained to the range from `0.0` to `1.0`.
- Bound the tool metadata supplied to the classifier and treat it as untrusted
input.
- Emit an extension warning if classification sampling fails.
## Testing
Add an integration test that verifies the contributor reuses the configured
Luna pool and sends the expected model, attribution metadata, reasoning effort,
strict output schema, and tool-call context.
GitOrigin-RevId: 6b69fcc0443086544086849c38bcdd90560a291f
## What changed
- Prewarm two Responses WebSocket connections and grow the pool on demand, up to eight connections, so overlapping samples can run concurrently.
- Lease each connection to a single sample and return healthy connections to the idle pool after completion.
- Replace closed or 55-minute-old connections and retry a sample once after a retryable or stream error.
- Keep the sampler available when the second prewarm connection cannot be opened.
## Testing
- Cover connection reuse, partial prewarm success, concurrent pool growth, and retrying an expired connection on another warm connection.
GitOrigin-RevId: 3ee7b05adc00a2135ba7cb35f6e4455199bf3784
## Why
Structured output can be complete before the Responses stream emits its terminal
events. Waiting for those events unnecessarily delays the sampler and can leave it
waiting after usable output has arrived.
## What changed
- Return a Luna sample as soon as accumulated text deltas parse as a complete JSON
object.
- Continue draining the remaining response events in the background so the
authenticated WebSocket stays reusable for subsequent samples.
- Enforce the output-size limit while accumulating deltas on the early-return path.
## Testing
- Cover returning complete JSON without terminal response events.
- Verify that early return still allows the authenticated WebSocket to be reused.
GitOrigin-RevId: 80c9963f1540f7862356aaad3503e3071a8fb92a
## What changed
- Add a `LunaSampler` that opens an authenticated Responses WebSocket and reuses it for structured, tool-free `gpt-5.6-luna` requests.
- Carry the host's provider, authentication, proxy, attribution, and service-tier configuration into the connection and requests.
- Require strict JSON schemas, preserve per-request reasoning effort and turn metadata, and reject missing or oversized output.
## Testing
- Add a WebSocket test covering authentication and attribution headers, connection reuse, structured request fields, per-turn metadata, reasoning effort, and returned output.
GitOrigin-RevId: aef85fd8ddcfcb7c76a9ede0d3191e3a0321656b
## What changed
- Add the `codex-guardian-v2` crate with an extension install entry point that does not register contributors yet.
- Register the crate in the Cargo workspace and add its Bazel target.
GitOrigin-RevId: 7c0c9479c6722d65f0f0dad9b0fd4a756e4649b3
jif