fix(acp): deliver plain-text replies for self-hosted OpenAI-compatible agents

[repudi8or]

Summary

• Small local models served over a self-hosted OpenAI-compatible endpoint (e.g. llama.cpp/vLLM/mesh-llm) routinely finish a multi-step turn by writing the answer as prose instead of calling send_message — the turn reports EndTurn successfully, the ACP activity log shows the correct reply, but nothing is ever published to the channel. This is what we observed running a self-hosted local-model agent against a mesh-llm deployment.
• crates/buzz-acp: retain each turn's streamed text (bounded at 8 KiB) plus whether a publish tool call was seen; on EndTurn, if nothing was published, deliver that text as a threaded channel reply instead. Gated strictly on EndTurn (not MaxTokens/MaxTurnRequests, which mean the turn was truncated) and filters bare acknowledgements ("OK", "Done"). Off by default via --deliver-plain-replies / BUZZ_ACP_DELIVER_PLAIN_REPLIES, since cloud models reliably call send_message themselves and enabling it there would risk double-posting.
• desktop/src-tauri: opt the built-in "Buzz shared compute" (relay-mesh) preset into this fallback, and separately default it on for the general openai-compat provider too (a self-hosted endpoint configured directly, not through the built-in preset) — scoped narrowly so plain openai (the real OpenAI cloud API) is untouched. An explicit user override (including an intentional false) always survives.
• crates/buzz-agent: while validating the openai-compat fallback above against a real self-hosted setup (a proxy fanning a single model alias out across several backend providers), we also hit a second, unrelated failure mode worth fixing alongside it — a 400 from a backend routed through OpenRouter: "auto tool schema uses unsupported assertions or reserved metadata". The root cause: buzz-dev-mcp's tool schemas are auto-derived via rmcp/schemars, which routinely emits JSON Schema metadata and assertion keywords ($schema, title, format, additionalProperties, range/length constraints, ...) that Anthropic and the official OpenAI API happily ignore, but that some openai-compat backends validate strictly and reject outright — breaking every tool call for that agent, not just the reply-delivery path. Added sanitize_tool_schema, applied once at tool registration alongside the existing schema-size cap (cap_schema), which strips a denylist of those keywords from the whole schema tree before it's ever sent to a provider. Every stripped key only narrows or documents what a schema accepts, so this can only loosen a backend's validation — it can't change what a tool call actually does, and structural/semantic keywords (type, properties, required, items, description, enum) are left untouched. This is the same underlying theme as the reply-delivery fix above (self-hosted openai-compat backends don't share the reference providers' tolerances), so it seemed worth landing together — useful for anyone pointing buzz-agent at vLLM, llama.cpp, a litellm/multi-provider proxy, or any OpenRouter-routed model with a stricter function-calling schema validator than Claude/GPT.
• desktop/src-tauri: merged main in to pick up two commits that had landed on crates/buzz-acp since this branch started, and split readiness.rs's test module (readiness_tests.rs / readiness_provider_tests.rs) to stay under the desktop file-size ratchet after that merge — no behavior change, same tests, same assertions.

Test plan

• cargo test --manifest-path desktop/src-tauri/Cargo.toml --lib — 2250 passed, 0 failed, 14 ignored (full crate, not scoped to touched files), run independently by two people
• New targeted tests: publish_detection_matches_tool_and_shell_shapes, bare_acknowledgements_are_not_worth_publishing, non_mesh_provider_does_not_opt_into_plain_reply_delivery, openai_compat_provider_opts_into_plain_reply_delivery, real_openai_provider_does_not_opt_into_plain_reply_delivery, explicit_deliver_plain_replies_override_is_preserved_for_openai_compat
• cargo test -p buzz-agent --lib — 401 passed, 0 failed, including 4 new schema-sanitization tests (sanitize_tool_schema_strips_root_metadata, sanitize_tool_schema_strips_nested_property_assertions, sanitize_tool_schema_preserves_structural_and_semantic_keywords, sanitize_tool_schema_recurses_into_arrays)
• cargo clippy -p buzz-agent -p buzz-acp --lib -- -D warnings and cargo fmt --manifest-path desktop/src-tauri/Cargo.toml --all -- --check clean
• Full pre-push suite green after the main merge: branch-skew, rust-tests, desktop-check (file-size ratchet), desktop-typecheck, desktop-test (4480 JS tests), mobile-test, desktop-tauri-checks (2250 Rust tests)
• Independent code review of the diff
• Live end-to-end repro against the original self-hosted agent — not done pre-merge: the fix spans buzz-acp (bundled inside the Buzz Desktop app) and desktop/src-tauri, and there's no way to swap just one agent's binary without patching the signed Desktop app bundle used by every other agent on that machine. Recommend validating live once this ships through the normal Desktop build/update path.

Diagnosed and coordinated by the buzz-acp team (architect, dev, tester, conductor) in the team's working channel; core mechanism forward-ported from an unmerged branch (micspiral/mesh-0-74-gemma, commit bb4af4189) by Michael Neale.

with a helping hand from Claude Code

[repudi8or]

@wesbillman this fixes a bug affecting self-hosted OpenAI-compatible agents (e.g. local llama.cpp/mesh-llm endpoints) silently dropping their replies instead of publishing them — full context and test results in the PR description. Would appreciate a review/merge when you have a moment. I only have read access to block/buzz directly (opened from a fork), so can't self-merge.

[repudi8or]

Live end-to-end repro: confirmed ✅

Independently verified the events directly on the relay (not just trusting the report):

• Hello: 92df0f20db6cea23f2744e6b24dc476e09aae3a66299168e302fe2edf97abd0a — @Local LLM (llm1) hello, tagged to the dev-branch test agent's pubkey 5299bb38ce49b5090c08022359842c396398a6b36efd4cbcd8484d1aa56b22eb.
• Reply: 2351d7ad8b56f94ac457be5e880f5714dd3c59a45bf201a5f9343091f2339de8 — kind 9, from 5299bb38…, with an e-tag reply reference back to the hello event. Correctly threaded.

This closes the one unchecked item in the test plan: a self-hosted openai-compat agent's plain-text turn (no send_message tool call — the model never called it) landed as a real threaded channel reply instead of vanishing after generation. Setup: repudi8or's just dev isolated Desktop build (already on fix/deliver-plain-text-replies) spun up its own test copy of the "Local LLM (llm1)" agent against the real production relay, so this exercised the actual shipped code path end-to-end — not a mock.

Non-blocking follow-up spotted in the reply content (not the delivery mechanism, which is what this PR fixes): the model looped through many near-duplicate self-summaries for the full turn, and the delivered text is hard-cut mid-word — consistent with hitting the 8KiB MAX_TURN_TEXT_BYTES per-turn buffer cap in acp.rs. Delivery/threading itself is correct; this is llm1's own response quality (repetition / no natural stop) and worth a separate issue (better stop behavior or a repetition guard for small local models), not a blocker here.