[DenysMoskalenko]

Release audit of 12 October 2026, following maintenance/release-audit.md. Plugin 0.3.0 → 0.4.0.

Inventory

Item	Pinned	Current	Bucket	Outcome
rig / rig-core / rig-agent / rig-rmcp	0.43	0.44.0 (7 Oct)	b + c	migrated (below)
rmcp	2	3.5.1	hold	rig-rmcp 0.44 still pins rmcp = "2"
opentelemetry set	0.32 / 0.33 / 0.39	0.33.1 / 0.34 / 0.42.1	hold	reqwest-tracing 0.7.1 still has no opentelemetry_0_33 (TrueLayer/reqwest-middleware#288 open)
testcontainers	0.27	0.28.0	hold	testcontainers-modules is still 0.15.0
NATS server	2.14-alpine	2.15.1	hold	async-nats 0.50.0 has no server_2_15
otel collector contrib	0.161.0	0.162.0	a	tag; 0.162.0 now has a multi-arch image; no breaking entry touches the otlp receiver or debug exporter
cargo-nextest recommended	0.9.146	0.9.148	a	0.9.147 was never finalized
astral-sh/setup-uv (repo CI)	v10.2.0	v10.3.0	a	adds an input only
scaffold lockfile	-	-	a	cargo update: hyper 1.12, uuid 1.28, tokio-util 0.7.20, aws-lc-rs 1.18.2, pest 2.9.3 (adds stacker/psm, dev-only via insta); 588 → 590 packages
Rust toolchain, other crates, tools, images, actions	-	-	-	already current

rig 0.44 in building-rig-agents

A compile probe at 0.44 broke 8 snippets (16 errors); the 65 untagged fragments and the prose were checked by hand against the 0.44 sources and the upstream 0.43 → 0.44 migration section. Every edit is at the line the change invalidates:

• PromptResponse::output is a method; messages is a Vec.
• PromptError::{Provider, Memory, MaxTurns, Cancelled}, now #[non_exhaustive]; StreamingError is gone, so stream items fail with PromptError (agent_error, sse_events, frame in AXUM-INTEGRATION).
• Streaming: MultiTurnStreamItem::ToolResult, StreamEvent::Start { name }, tool-call arguments as fragments, stream_to_stdout returns the run's error, one chat span with gen_ai.request.stream.
• Tools: DynamicTool::try_from(mcp_tool)?, agent.into_tool()?.
• top_n returns named results; the in-memory store needs only Serialize.
• Structured output: ExtractorBuilder::tool_choice removed, retries(usize), StructuredOutputError::{Prompt, Deserialization { output, error }}.
• Media constructors audio_base64 / video_base64 / document_text / document_base64, and unsupported media becomes placeholder text.
• Malformed tool arguments reach on_invalid_tool_call on both run paths; the default answers with an error result; Repair fails the run; new DisallowedByToolChoice reason; the per-run table gains max_consecutive_malformed_tool_calls and accept_unknown_finish_reasons.
• Failed turns are a 500. A turn the provider ends in a content filter or an unknown finish reason, and malformed tool arguments that end the run, now fail as Provider(ProviderError::Response(..)) with no status. agent_error would have answered them 503 through its catch-all; its existing 500 arm now also matches ProviderError::Response and ProviderError::UnsupportedOption, since rig marks both not retryable and the house contract keeps 503 for a backend that is down. Provider outages still reach 503: Anthropic stream error events and OpenAI response.failed arrive as ProviderResponse. A test pins the content-filter case; the SKILL bullet and eval 1 say the same. Edge: an OpenAI Responses reply with status: failed in a 200 body (background mode) is a Response turn failure too, so it is now a 500.
• Typed generation options. AGENTS-CORE gains a Reasoning and Caching section with a compiled snippet: .reasoning(Effort::High) (Anthropic output_config.effort, adaptive thinking where the model takes it), .cache(CacheRetention::Long) (one-hour cache_control), .on_unsupported(..), the default efforts (Opus 5.5 and Haiku 5.5 medium, other Claude models high) and how a refused option surfaces. This closes the effort gap from the 25 Sep prompt audit.
• Description. ToolCallError, absent from rig 0.42–0.44, is replaced by "max turns limit errors (MaxTurnsError)", which matches 0.44's reached the max turns limit of N; triggering prompt 9 pastes that message.
• Thinking-binding models. Anthropic's docs list Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5; the skill named only the first two. rig 0.44's catalog has no Haiku 5.5, so rig sends no block_binding there; Hooks and Memory say so.
• Durable memory: rig 0.44 cannot read 0.43 assistant turns holding tool calls or reasoning, so the Memory reference says to clear or migrate those rows before upgrading.
• On Anthropic's own API rig 0.44 asks for context-bound thinking blocks to be dropped instead of failing the request (thinking.block_binding, adaptive thinking): the 400 warnings in Hooks and Memory say so.
• VERSION-DRIFT gains a 0.43 → 0.44 table and behaviour list (failed turns, argument fragments, eager_input_streaming, preserve_order, persisted history and AgentRun format).
• Eval file updated to the 0.44 names.

Follow-ups (not in this pull request)

• Behaviour evals for building-rig-agents were not run (owner decision); AGENTS.md now says to ask for eval approval at the start of a session.
• rig 0.44's catalog marks claude-opus-5 as binds_context, which Anthropic's docs do not list; rig sends block_binding there. Harmless as far as the skill teaches; worth an upstream issue.

Verification

• python3 scripts/validate_skills.py --strict --base main: 0 errors (9 pre-existing budget warnings).
• make check with Docker: every snippet compiles, clippy -D warnings clean, 70/70 snippet tests pass (including the new content-filter test).
• Scaffold: make check, make test (14/14), cargo deny check (advisories, bans, licenses, sources ok); root prek run --all-files ok.
• Reviews: three source-backed validators (API accuracy, migration completeness, repository conventions), a fresh-eye review, ponytail-review and /code-review xhigh; each finding checked against the rig 0.44 source before it was applied or rejected.
• Trigger-only eval for building-rig-agents (with-plugin arm, 2 runs per case, Opus 5.5): should-load 9/9 cases, 18/18 runs; should-not-load 18/18 runs kept building-rig-agents out (one skip-03 run missed firing rust-redis, a rust-redis recall miss). Cost $11.75.
• A second round of reviewers (rig-source accuracy, fresh-eye) on the post-review changes; each finding checked against the rig 0.44 source or Anthropic's docs before it was applied.
• Not run: the throwaway-service exercise (the scaffold changed only a comment, an image tag, nextest's recommended and the lockfile) and the behaviour evals.