Skip unrecognized stream output, add list_models and default headers, and prepare 0.4.5

[hiyouga]

Four changes, one release folder (changelog/0.4.5/).

1. Unknown stream output is skipped unless AGENTHUB_DEBUG is set (fix)

A relay in front of a Responses-compatible API injected {"type":"ping","cost":"@"} into the SSE stream. ping is the Anthropic Messages heartbeat spelling, so the Responses clients' unknown-event guard raised Unknown output: {"type":"ping","cost":"@"} and the stream died; the harness in front classified the throw as a retryable request failure and every retry hit the next heartbeat. 0.4.3 (#174, for Prism-Shadow/penguin-harness#286) had taught each family only its own spelling — keepalive for Responses, ping for Messages — so either spelling crossing into the other protocol still raised.

The clients no longer end a stream over output they do not recognize: claude5, ant_messages, openai_responses, gpt5_6 and minimax_m3 mark an unknown event unused, and gemini3_7 skips an unknown content part. is_debug_enabled() / isDebugEnabled() in utils reads AGENTHUB_DEBUG per event; set it to anything other than 0/false/no/off and every unrecognized event or part raises again with the same message.

Two notes on scope:

• gemini3_7's part guard got the same treatment. It reports an unknown content part, not an event, so it sits outside the letter of the change — but it is the same failure mode (one unrecognized thing kills the stream), and leaving it would have left the policy with a hole. Easy to revert if content parts should stay loud.
• The per-protocol lists of known no-op events are unchanged, so a heartbeat a client already knew is still recognized by name rather than by falling through.

The heartbeat tests were rewritten around the switch and the files renamed to test_unknown_events.py / unknown-events.test.ts: heartbeats, cross-protocol spellings, in-protocol unknown events, gateway error frames and unknown Gemini parts all pass through silently by default, and each raises once AGENTHUB_DEBUG is set.

2. AutoLLMClient.list_models() (feature)

Returns the model ids the configured endpoint actually serves, as list[str], in the endpoint's own order. Every client implements it against its own SDK (models.list() for the OpenAI-compatible and Anthropic families, the Gemini model list for gemini3_7); paging belongs to the SDKs, so a result spans every page.

The playground exposes it through a List models control on the Base URL label row, backed by POST /api/models; the result renders in a panel that overlays the field, so the control costs no vertical space while idle. A rejected listing returns 400 with the provider's message — verified unmocked against an unreachable base URL (Connection error.) and a bedrock:// URL (Bedrock does not support listing models.).

Two judgment calls:

• Gemini names are path-qualified (models/gemini-3.7-flash, publishers/google/models/…), so the client returns the last path segment: an id from list_models() is directly usable as AutoLLMClient's model.
• Bedrock has no models resource in either SDK, so claude5 raises the new UnsupportedOperationError. A dedicated class rather than UnsupportedParameterError because penguin-core reads .parameter to classify errors (err.parameter === "fast_mode"), and a method name has no business in a field consumers branch on.

LLMClient declares the method abstract. Not a public break: neither package exports LLMClient.

3. Clients accept default headers (feature)

AutoLLMClient(..., default_headers=…) / { defaultHeaders }, sent with every request to that endpoint — for gateways that reject a client they do not recognize (unauthorized_client_error, a required HTTP-Referer / X-Title / Anthropic-Beta). A connection property, not a request one: passed once at construction, never entering UniConfig. OpenAI and Anthropic take default_headers directly; Gemini carries them inside http_options.headers.

Verified over the wire, not by poking SDK internals: both suites start a local HTTP server, list models through the OpenAI, Anthropic and Gemini clients, and assert the declared headers arrived — plus that a client declaring none adds none. (The OpenAI SDK lowercases custom header names while the Anthropic one sends them as given, so the comparison is case-insensitive.)

The playground gained an Extra Headers field; text that is not a JSON object marks the field instead of being sent.

4. The Responses input transforms match the GPT client (refactor)

openai_responses and minimax_m3 dropped the early-continue guards for one if/elif chain, and now flush buffered text once at the end of a message, exactly as gpt5_6 does.

That second half changes replay order, deliberately. Extracting the pre-refactor transform from HEAD and running it beside the new one over 7 message shapes: 4 differ, all in the same way — text that preceded a top-level item now follows it, and two text spans separated by such an item arrive merged. [text, tool_call] replays as [function_call, message], which is what gpt5_6 has always sent. Messages carrying only text and images are untouched.

Verification

• src_py: make lint clean (ruff check + format); uv run pytest tests --ignore=tests/test_client.py → 199 passed.
• src_ts: npm run lint, npm run build clean; the offline suites → 184 passed.
• git merge-tree against glm-5.3-ga-and-model-list, the other open branch touching both playground files, reports no conflict.
• The e2e suites need API keys, so they are CI's.

🤖 Generated with Claude Code

https://claude.ai/code/session_013STCF9fADwqT8jJH9C3g5h

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.