v3.11.1

Changed

• Set the Claude plugin catalog display name to orq.ai.

v3.11.0

Added

• orq-analyze-traces/scripts/decide.py (RES-1629): one POST /v3/router/decisions call per trace over every trace in the window, to surface failing traces the analyst never read. Traces go whole to typesafe/jev-latest when they fit and to openai/gpt-6-luna otherwise, never chunked or trimmed. Subcommands probe, estimate, label (resumable, capped by a required --budget-usd, refuses windows over 5,000 traces without --max-traces), pick, agreement and rates. Never writes to the platform.
• orq-analyze-traces: a Phase 1 selection pass over the window, a Phase 4 calibration against the analyst's own labels with a fixed clearing rule (10+ agent positives, recall 0.7+), and a window pass that reports each cleared mode's flag rate, interval and flagged trace ids. Spend is approved through AskUserQuestion. Falls back to the sample-based path when decisions are unavailable or declined.

Changed

• Error-analysis artifact: rate is now the best available rate and rate_basis says which (window_flags or read); read_rate always holds the read-based rate. New decisions, traces.labelled, per-mode window_flags and flagged.
• orq-improve-agent, orq-recommend-evaluators: quote rate with its rate_basis, and treat flagged as unread examples, never as regression evidence.

Fixed

• Calibration counts missing analyst-positive labels as misses and rejects unreadable probabilities. Budget reservations halt the whole pass on refusal, retain uncertain charges and attempted overflow retries, and reject non-finite limits. Active-window metadata is written before labelling; reports and picks filter the cache to that population and refuse missing metadata. Add an offline regression suite and pinned API contract sources.
• Preserve complete API-returned trace IDs in analysis evidence, including platform ULIDs as well as OpenTelemetry IDs; do not substitute record or span IDs.

v3.10.1

Changed

• orq-setup-observability: added the first-party observability one-liners orq_ai_sdk.langchain.setup() and OpenAIAgentsInstrumentor().instrument() (RES-1221).
• orq-invoke-deployment: stream deployments with client.deployments.stream(...) (an EventStream) instead of invoke(stream=True), which returns a single response body and is not iterable (RES-1221).
• orq-setup-observability: the OpenAI Agents one-liner needs a configured TracerProvider (the instrumentor reads the global one and replaces the Agents SDK trace processors); documented, with the required packages.
• Agent text-extraction snippets (orq-invoke-deployment, orq-compare-agents) raise when the response has no output_text block instead of returning an empty answer.
• orq-invoke-deployment API reference: documented the conversation ({id}) and thread ({id, tags}) parameters of agent Responses.create, with factual rows.
• tests/scripts/run_factual_tests.py: SDK probes use SystemExit instead of assert, so python -O cannot turn a failure into a pass. Added a previous_response_id signature row.

v3.10.0

Added

• orq-review-agent skill (RES-887): a read-only health check of one orq agent. It checks run health, instructions, tools, {{skill.X}} references, evaluators and model choice against the suite's rubrics, gives each finding a high-level suggestion (never a rewritten prompt or diff), lists project evaluators the agent does not use and tools its instructions need but it lacks, writes one HTML report into a per-run directory (agent-review-<key>-<timestamp>/), posts a summary with the report link, shows the findings on request, then offers the skill that owns each fix. allowed-tools pre-approves no write verb. The model check (scripts/model_check.py, stdlib only) applies the RES-1338 A1 rule with the RES-887 guardrails, and lists a cheaper provider of the same model as its own finding. The rule text lives once in the script and is printed in every result. An unresolved model id lists the closest catalogue names, and a config contradiction certain to break every run is labelled predicted broken when there is no traffic to confirm it. Probed live on CLI 11.0.2; untested paths are marked [unverified].
• orq-shared/resources/config-context.md, prompting-guidelines.md and tool-description-guide.md: the Phase 0 relay list, the 11 prompting guidelines and the tool description guide, now shared by the skills that build, review and improve an agent.
• orq-shared/resources/model-slugs.md: which catalogue field names a model (id, refId, model_id) and the rule for resolving an agent's model id, cited by orq-cli, orq-evaluator-alignment, orq-recommend-evaluators and orq-review-agent.
• tests/factual/orq-review-agent.csv and tests/evals/orq-review-agent/ (three fires cases, three near-miss cases, a read-only behavioural case on support-bot and a cross-project behavioural case).

Changed

• orq-analyze-traces Phase 0, orq-improve-agent Phase 3a and orq-build-agent link to the shared rubrics instead of keeping their own copies. orq-analyze-traces now routes setup health checks to orq-review-agent. orq-improve-agent Phase 1 also checks that the declared model is available in the project (orq models list, resolved per model-slugs.md). tool-description-guide.md moved from orq-build-agent/resources/ to orq-shared/resources/.

v3.9.3

Fixed

• Require evaluatorq 1.30.0 or newer for the Responses helper API, check the minimum in CI, and type-check SDK/HTTP examples directly from markdown to catch scope and schema regressions.
• Show paired orq/OpenAI Responses examples for models, agents and comparison jobs. Prefer orq for new code and existing orq imports; reuse OpenAI if it is already imported and orq is not. Document async and response-parsing differences.
• Replace the generated CLI command inventory, group listing and inventory-generation script with progressive --help discovery. Retain task-specific examples, behavior caveats and invocation regression checks.
• Use openai/gpt-6-luna in direct model invocation examples and model command defaults instead of older GPT models.
• Model and agent invocation guidance now defaults to POST /v3/router/responses and orq responses create, using provider/model-id for models and agent/<key> for managed agents. Prefer Responses over Chat Completions and avoid v2 inference. Other inference interfaces require an explicit user request and must never be automatic error fallbacks.
• Update invocation, observability, simulation and comparison examples to Responses request/output shapes, including streaming, structured output and previous_response_id continuation. Framework guidance explicitly selects Responses (including LangChain use_responses_api=True). Configuration, deployment, evaluator and telemetry APIs retain their own contracts.
• Route evaluator-alignment recommendation/rewrite helpers and reconstructed judge calls through Responses, preserving structured verdicts and pricing input/output tokens.
• Add a CI guard against non-Responses inference examples and runtime calls and local synthetic CLI routing probes for both models and agents. Document the CLI SSE decoding limitation and use SDK/curl for consuming streams on the same route. Contract checked against current Orq documentation, LangChain's Responses reference and installed CLI help on 2026-10-22; the reported EU-model failure has not been reproduced against a live agent.

v3.9.2

Changed

• orq-cli/resources/command-map.md "Building app URLs" is the single definition of trace and span links in the app: <host>/<workspace-key>/traces/(trace:<trace_id>) and …/(trace:<trace_id>//span:<span_id>), opened cold in a browser on prod on 2026-10-13. The snippet takes the host from ORQ_UI_BASE_URL or the session (urls.api_base_url, or server under ORQ_API_KEY) and fails loudly when there is none, or when the workspace key or an id is missing. It records that a link to a trace outside the user's selected Traces project opens on "No spans match your search" (switch to All projects), and that ids from an invocation's telemetry were not checked against the route (RES-1550).
• orq-invoke-deployment, orq-setup-observability, orq-analyze-traces, orq-compare-agents and orq-shared/resources/trace-queries.md hand the user a link to the trace instead of "go to my.orq.ai → Traces", and point at that section. orq-analyze-traces now requires full 32-hex ids in evidence and passing.
• orq-cli skill: with ORQ_API_KEY set, orq auth whoami on 11.0.2 reports active_workspace_key: null rather than the logged-in workspace; the constraint and the workspace-key snippet's error message say so.

Removed

• The traces?query=trace_id:is:<id> deep link from orq-cli; the app no longer routes it.