Bump inspect-ai from 0.3.249 to 0.3.250

[dependabot]

Bumps inspect-ai from 0.3.249 to 0.3.250.

Changelog

Sourced from inspect-ai's changelog.

0.3.250 (28 July 2026)

• Sample Sources: Generate a task's samples dynamically while it runs by passing a SampleSource as the task's dataset (with enqueue_sample() for imperative additions) — the sample-level mirror of TaskSource, for RL loops and adaptive evals.
• Eval: Multi-task runs without task_retry_attempts now use the same task dispatcher as runs with retries (the separate no-retry dispatcher was removed).
• Control Channel: New inspect ctl task pause|resume and inspect ctl process pause|resume commands pause a running eval or eval-set (in-flight samples finish; nothing new starts) and resume it in place, with paused/quiesced reported by inspect ctl task list.
• Control Channel: New inspect ctl sample messages TASK SID [EPOCH] reads a running (or buffered-but-unlogged) sample's current conversation as a snapshot, with --tail/--all/--full.
• Control Channel: New inspect ctl process anomalies [PID] shows a running process's in-flight and anomalous trace actions by reading its trace file directly, so it works against busy, hung, or even exited processes.
• Control Channel: inspect ctl config changes are now recorded in each affected eval log (EvalLog.config_updates, with author, timestamp, optional --reason, and old → new values); effective_eval_config()/effective_generate_config() fold them over the launch config.
• Control Channel: inspect ctl sample events --full now pretty-prints the raw events instead of rendering a mostly-empty summary table.
• meridianlabs-ai/inspect_ai#163
• Anthropic: Support for Claude Opus 5 (released 2026-07-24), including model info, computer use, code execution, and disabling thinking via reasoning_effort="none".
• Anthropic: Support enabling model-generated citations for provided ContentDocument inputs.
• Google: Added model info for Gemini 3.6 Flash and Gemini 3.5 Flash-Lite (released 2026-07-21).
• OpenAI Compatible: A length-truncated streaming response (e.g. -M stream=true hitting max_tokens) now degrades gracefully to a max_tokens stop reason instead of raising LengthFinishReasonError and aborting the eval. (#4552)
• OpenRouter: Reasoning history replayed to Gemini models now contains only readable <think> text, rather than HTML-escaped signature JSON and encrypted payloads. (#4320)
• Agent Bridge: Support Anthropic clients that consume responses via with_raw_response, which previously failed with 'Message' object has no attribute 'parse'.
• meridianlabs-ai/inspect_ai#140
• Agent Bridge: Gemini tool results delivered as functionResponse parts in model-role turns are now converted to tool messages instead of dropped (fixes 400 "Requests ending with a model turn are not supported").
• Agent Bridge: Google streaming requests targeting an inspect model now raise the intended streaming-unsupported error instead of being sent to the real Google API.
• Sandbox: Sandbox.exec failure errors now include stdout as a fallback when stderr is empty, so commands that report diagnostics on stdout still produce a useful error message.
• Memory tool: Canonicalize /memories paths so equivalent spellings map to one file rather than several, and add an instance parameter to memory() for independent per-instance memory stores.
• Approval: Load policy configuration files from file:// URIs as well as plain local paths.
• Registry: Add a validation_predicate type for Scout extensions.
• Logging: Realtime streaming sample writes now normalize unserializable store values the same way as the standard log write path, instead of raising during serialization. (#4120)
• Tracing: trace-file reads (inspect trace commands and inspect ctl process anomalies) now skip truncated, corrupt, or unrecognized records — e.g. from a hard-killed process or a newer inspect version — instead of failing entirely.
• Windows: process liveness checks (used by inspect ctl discovery and pid targeting) no longer risk sending a console Ctrl+C to the probed process or misreporting live processes as dead.
• Inspect View: Added transcript event navigation — keyboard-driven turn and agent navigation with a focus view for drilling into agent subtrees (#354).
• Inspect View: Added connection limit history display to the Stats tab. (#447)
• Inspect View: Improved log parsing performance; added real-payload benchmarks. (#384)
• Inspect View: Timeline now requires a tool-calling loop for utility-agent classification and surfaces the hidden event count. (#425)
• Inspect View: Fixed timeline discarding a solver span containing a single agent span child when it also contains other displayed events. (#443)
• Inspect View: Fixed viewer to show the Action tab first for annotated browser actions. (#429)
• Inspect View: Fixed sanitization of state diff HTML, added alt text to images, and cleaned up leaked timers and scroll-to-top behavior (#428).
• Bugfix: Resolve file paths and resources passed to the chain_of_thought solver. (#4631)
• Bugfix: is_data_uri() now recognizes base64 data URIs that carry media-type parameters (e.g. ;charset=utf-8) or an empty media type (data:;base64,...), which were previously misclassified as non-data URIs.
• Bugfix: Crash recovery now preserves full sample metadata in resumed logs and score metrics instead of persisting summary-thinned placeholders.
• Bugfix: file_dataset() now recognizes JSON and CSV URLs with query parameters while preserving the complete URL passed to the selected dataset reader.
• Bugfix: Transcript timelines no longer hide workflow generate() calls with differing system prompts as utility agents; utility classification now requires a tool-calling agent loop.
• Bugfix: JSON, CSV, and Hugging Face dataset loaders now return an empty dataset when passed limit=0 instead of returning all samples.
• Bugfix: Parameterized score reducers (e.g. at_least(2), pass_at(2), pass_k(2)) now round-trip through the registry instead of raising LookupError when restored from a log.
• Bugfix: Fixed a memory leak where the result of any synchronous API call made from a context with no running event loop (e.g. each sample yielded by read_eval_log_samples()) was retained in memory forever after the caller released it.
• Bugfix: Cache traceback rendering so repeated identical errors no longer stall the event loop re-formatting the same traceback. (#4316)
• Bugfix: Anthropic requests no longer fail with 400 container_id is required when a turn mixes code-execution-backed server tools (e.g. web search with dynamic filtering) with client tool calls.

Commits

• 41adc99 Update CHANGELOG for version 0.3.250 release
• 3b7f744 Polish changelog order (#4661)
• 3209630 test(log): use async log APIs in trio-incompatible test bodies (#4622) (#4654)
• e188ba5 Fix mypy failure when httpx2 is installed alongside httpx (#4660)
• 66d7fa3 Merge pull request #4615 from UKGovernmentBEIS/large-samples-cd-layout
• d813199 Bump ts-mono to merged main (5d72b0d7)
• 80e5864 land-ts-mono skill: document paired-branch discovery from detached HEAD
• f42a2f2 Merge ts-mono origin/main into branch; bump gitlink, rebuild dist
• 1999216 Merge remote-tracking branch 'origin/main' into large-samples-cd-layout
• de36017 Restore control server under fastapi 0.140.7 (#4653)
• Additional commits viewable in compare view

Dependabot will resolve any conflicts with this PR as long as you don't alter it yourself. You can also trigger a rebase manually by commenting @dependabot rebase.

---

Dependabot commands and options

You can trigger Dependabot actions by commenting on this PR:

• @dependabot rebase will rebase this PR
• @dependabot recreate will recreate this PR, overwriting any edits that have been made to it
• @dependabot show <dependency name> ignore conditions will show all of the ignore conditions of the specified dependency
• @dependabot ignore this major version will close this PR and stop Dependabot creating any more for this major version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this minor version will close this PR and stop Dependabot creating any more for this minor version (unless you reopen the PR or upgrade to it yourself)
• @dependabot ignore this dependency will close this PR and stop Dependabot creating any more for this dependency (unless you reopen the PR or upgrade to it yourself)

[amrit110]

Automated fix applied and PR merged

The agentic fix loop successfully fixed this PR and merged it.

✓ Successfully fixed merge_only failures - Modified 0 files - Executed 71 agent actions - (30 info, 20 tool_call, 5 error, 14 tool_result, 2 reasoning)

View detailed trace on dashboard | Raw trace

AI Engineering Maintenance Bot