feat(mesh): upgrade to v0.74 and make shared agents reply reliably by micspiral · Pull Request #3467 · block/buzz

micspiral · 2026-07-29T04:55:39Z

Summary

Upgrade embedded Mesh to v0.74 and use canonical Gemma model IDs.
Keep shared compute serving through member joins, roster changes, app recovery, and community switching (no crashy!).
Wait for actual model readiness and avoid resuming incomplete downloads after quit.
mesh only scoped tweaks: improvements with buzz agent using mesh with thinking modes (corrected so they don't interfere) a reply tool (so it can talk back) and a failsafe to scoop up responses if that fails for small models (and scoped to mesh/shared compute only, no other providers affected).

Live validation

Tested across two Macs: a production Buzz 0.73.1 client joined and messaged an agent while this development build hosted Mesh v0.74. The host stayed in serving mode, Gemma completed both turns, and the agent log confirmed both replies were published back to the channel.

Also tested a build https://github.com/michaelneale/buzz-testing/releases/tag/v0.5.1-test1 across 2 unrelated machines and it sustained a connection, shared compute reliably.

Testing

Focused ACP, desktop Mesh, and harness checks pass locally. Full CI is running on this PR.

Pin mesh-llm to the v0.74.0 tag instead of a commit rev, and carry the desktop-side changes that upgrade needs: catalog/coordinator/recovery updates, the relay-mesh agent preset, and restart behaviour that survives community boundary changes and roster churn. Also promote reasoning to text in the OpenAI-compat parser when `content` is empty and there are no tool calls. Local models served through mesh sometimes return the entire answer as reasoning; with no tool calls there is nothing else to surface, so reporting an empty turn loses the reply. Signed-off-by: Michael Neale <michael.neale@gmail.com>

Publishing a reply required composing a `buzz messages send ...` string inside a `shell` call. That indirection is reliable for frontier models and unreliable for the small local models shared compute exists to serve: against the real ~3.2k-token Buzz system prompt and the real dev-mcp toolset, Gemma 4 E4B delivered a reply in only 2 of 8 samples, answering in prose the rest of the time — and sometimes emitting the literal command text as assistant content. Assistant text is never published, so those replies were silently dropped while the turn reported success. Exposing the publish action as its own typed tool removes the indirection: the model selects a tool by name and fills one string field. Measured on the same prompt and toolset, delivery went 2/8 -> 8/8 on gemma-4-E4B and 6/6 on gemma-4-26B-A4B, and the real ACP path (buzz-acp + buzz-agent + live relay) replied correctly 3 for 3. The tool duplicates no process machinery — it builds an argv, quotes it, and delegates to shell::run, inheriting its timeout, output capping, cancellation and process-group kill. Channel ids are normalized (the `[Context]` block renders `general (#uuid)`, so models pass `#uuid`) and validated to bare ids, and content is single-quoted, so model output cannot reach the shell as metacharacters. Signed-off-by: Michael Neale <michael.neale@gmail.com>

P5 could never pass: install_fake_shell returned the bin directory, but resolve_bash requires BUZZ_SHELL to name a file, so it silently fell back to real bash and the marker was never written. The fake shell also matched `-lc` while shell_flag() emits `-c` on Unix. Return the executable, accept both flags, and append rather than overwrite so a probing model does not erase the recorded send. P5 stops at the shell boundary, so it cannot show a reply ever lands. Add P6: the agent runs the genuine `buzz messages send` (dev-mcp's multicall shim, real buzz-cli) against a running relay, and the harness then queries the relay itself — the agent's prose is not evidence. Infra-gated on BUZZ_E2E_RELAY_URL / _PRIVATE_KEY / _CHANNEL; skipped when unset so P1-P5 still run without Postgres and Redis. Also drive the agent with the real base_prompt.md instead of a terse hand-written string. A short prompt is not a proxy for production: tool adherence in small models depends on it, so testing with 56 characters where Buzz injects ~3.2k tokens measures the harness, not Buzz. Signed-off-by: Michael Neale <michael.neale@gmail.com>

The relay-mesh preset defaulted BUZZ_AGENT_THINKING_EFFORT to "none" to spend the output budget on an answer instead of hidden reasoning. On local models that also switches tool calling off: llama.cpp's Gemma grammar is lazy under `tool_choice: auto`, and `reasoning_effort: "none"` keeps the model on its prose completion path so it never emits the tool-call token. The agent then answers in plain content, which is never published, and the turn reports success with nothing in the channel. Isolated by replaying a captured production request against a local mesh node and varying one field: with the real prompt and toolset, gemma-4-E4B emitted a tool call 0/8 times at "none", 3/4 at "minimal", and 4/4 at "low" (4/4 with the field omitted entirely). Default to "low", which buys enough deliberation to select a tool without spending the budget on hidden reasoning. Still a default, so explicit agent/persona/global values are preserved. The mesh e2e harness claims to supply exactly what apply_relay_mesh_env() does, so update it in lockstep — pinning "none" there would test a config the product no longer ships and hide the regression P5/P6 exist to catch. Signed-off-by: Michael Neale <michael.neale@gmail.com>

Shared compute pinned BUZZ_AGENT_THINKING_EFFORT. That is not a sampling hint: mesh translates `reasoning_effort` into the chat template's `enable_thinking` flag, so any value Buzz sends overrides the default the model authors shipped — and the right value is model-specific. Measured against a request captured from the real agent (full ~43k-token system prompt, real dev-mcp toolset, tool_choice=auto): effort gemma-4-E4B Qwen3-8B Qwen3-8B output tokens none 0/8 8/8 ~121 low 4/4 7/8 ~470 unset 6/6 8/8 ~500 `none` (the previous default, chosen when shared compute was Qwen3-only) stops gemma tool-calling outright, so its replies were never published. Enabling thinking instead makes Qwen3 spend ~4x the output budget, which risks the 4096 cap this preset sets. Omitting the field leaves `enable_thinking` as None, so mesh passes no override and each model uses its own template default: 3/3 delivery on gemma-4-E4B, gemma-4-26B-A4B, gemma-4-31B and Qwen3-8B, and 3/3 replies through the real buzz-acp + buzz-agent path against a live relay. Thinking levels are unaffected: an explicit agent/persona/global value is still preserved and applied. Signed-off-by: Michael Neale <michael.neale@gmail.com>

buzz-agent's contract is that output is tool calls: streamed assistant text is observability only and never published. Capable models honour that and call `send_message`. Small local models served over shared compute do not always — on multi-step turns they run their tools, then write the answer as prose and end the turn, so the turn reports success with nothing in the channel. Retain the turn's streamed text in AcpClient (bounded at 8 KiB) alongside a flag for whether a publish tool call was seen, and on EndTurn publish that text as a threaded channel reply when the agent published nothing itself. Gated on EndTurn only: MaxTokens / MaxTurnRequests mean the turn was truncated, so the text is not a deliberate answer. Bare acknowledgements ("OK", "Done") are filtered, and taking the buffer clears it so the same text can never post twice. Off by default via --deliver-plain-replies / BUZZ_ACP_DELIVER_PLAIN_REPLIES: for cloud models, which reliably publish their own replies, posting streamed text would double-post. The desktop shared-compute preset opts in, so this stays scoped to the case that needs it. Verified end to end against a live relay with stub ACP agents that pin each branch deterministically: a prose-only agent gets its answer delivered, and an agent that calls send_message while also streaming prose posts exactly once with the fallback silent. Signed-off-by: Michael Neale <michael.neale@gmail.com>

cargo fmt --all skips desktop/src-tauri (separate workspace), so this line only failed under just desktop-tauri-fmt-check in CI's Rust Lint job. Signed-off-by: Michael Neale <michael.neale@gmail.com>

micspiral · 2026-07-30T04:17:18Z

ok - yes this is working quite well now - even for a small model. The other fixes for sharing seem ok too. This did require a reply tool added, and a mesh scoped change to scoop up missed replies so it returns to buz, but that seems to work well. Shouldn't impact other models outside of mesh.

It also needed some tweaks to ensure thinking modes are used appropriately (very sensitive to those).

The typed `send_message` reply tool was registered unconditionally, so every ACP agent's toolset grew from seven tools to eight. That changes the prompt each agent sees, and the delivery benefit was only ever measured on a small local model served over shared compute (3/8 -> 8/8 on gemma-4-E4B). Cloud models already publish their own replies by composing a `buzz messages send` shell command, so they were absorbing a behavioural change for no gain. Gate the tool on BUZZ_ACP_SEND_MESSAGE_TOOL, off by default. buzz-acp forwards it to the MCP server as BUZZ_MCP_SEND_MESSAGE_TOOL — via the ServerSpec env, since spawn_one clears the environment and only passes a whitelist — and dev-mcp drops the route when unset. Agents without it publish replies through the shell tool exactly as they did before the tool existed. This gets its own flag rather than reusing --deliver-plain-replies: one controls which tools exist, the other whether the harness publishes on the agent's behalf, and conflating them would make both harder to reason about. The desktop shared-compute preset sets both. Verified over a real MCP handshake on two machines: tools/list returns seven tools with the variable unset and eight with it set, on both the build host and a second Mac running the bundled app. Signed-off-by: Michael Neale <michael.neale@gmail.com>

micspiral · 2026-07-30T05:14:31Z

finally! end to end.

tlongwell-block · 2026-07-30T13:58:39Z

Split requested by Tyler: the mesh-scoped commits from this PR (mesh v0.74 upgrade + shared-compute hardening + thinking-effort defaults) are now separately mergeable as #3741, with authorship preserved. The agent/ACP reply-behavior work (send_message tool, plain-reply delivery, e2e P5/P6, and the llm.rs reasoning→text parser change from 2cd640b) remains here as part 2 — details in #3741's description.

@michaelneale

…(split 1/2 of #3467) (#3741) ## Summary This is **part 1 of 2** split out from #3467 (per Tyler's request), carrying only the mesh-scoped changes. The agent/ACP response-behavior changes and the new `send_message` tool stay in #3467 as part 2. All commits are @michaelneale's work, cherry-picked with authorship preserved. - Upgrade embedded Mesh to v0.74.0 (tag-pinned instead of commit rev) and use canonical Gemma model IDs. - Keep shared compute serving through member joins, roster changes, app recovery, and community switching. - Wait for actual model readiness and avoid resuming incomplete downloads after quit. - Leave `BUZZ_AGENT_THINKING_EFFORT` unset by default so each model's chat template picks its own thinking default (`none` suppressed Gemma tool-calling entirely; pinning `low` made Qwen3 burn ~4x output budget). Explicit agent/persona/global values still win. ## Relationship to #3467 Contains the mesh commits from #3467 (`2cd640b23`, `0ad81c341`, `ad13ed841`) rebased onto current main, with one deliberate exclusion: the `crates/buzz-agent/src/llm.rs` reasoning→text parser change from `2cd640b23` is **not** here. That change unconditionally affects every OpenAI-compat/Responses provider, so it belongs with the reply-behavior work in part 2, where it can be reviewed as what it is. Not included (remaining in #3467 / part 2): - typed `send_message` tool in dev-mcp + `BUZZ_ACP_SEND_MESSAGE_TOOL` gating - plain-reply delivery fallback in buzz-acp (`BUZZ_ACP_DELIVER_PLAIN_REPLIES`) - the mesh_agent_e2e P5/P6 rewrite (exists to prove the reply path) - the two `env.insert` preset opt-ins in `relay_mesh.rs` for the flags above - the llm.rs parser change This PR is independently mergeable; part 2's flags are all off by default so it can land before or after. ## Testing - `cargo test -p buzz-relay --locked` — 780 passed (one telemetry test is order-sensitive under parallel default settings; passes in the pre-push suite and standalone, unrelated to this diff — files untouched here). - `just desktop-tauri-test` (default features) — 1877 passed. - `cargo test --locked --features mesh-llm` in `desktop/src-tauri` — 1961 passed, including the new relay-mesh preset and coordinator/recovery tests. - Both `Cargo.lock`s resolve with `--locked` against the v0.74.0 tag. - Full pre-push hook suite green (rust-tests, desktop-check/test, tauri checks). Live validation of the mesh v0.74 upgrade itself is documented on #3467 (two-Mac cross-version test). --------- Signed-off-by: Michael Neale <michael.neale@gmail.com> Signed-off-by: Tyler Longwell <tlongwell@block.xyz> Co-authored-by: Michael Neale <michael.neale@gmail.com> Co-authored-by: Tyler Longwell <tlongwell@block.xyz>

tlongwell-block · 2026-07-30T15:37:59Z

Closing per Tyler — this PR has been split, and the mesh half now lives in #3741.

What #3741 carries (rebased onto current main, Mic kept as author on all commits): the mesh v0.74 upgrade + shared-compute hardening (2cd640b23 minus the llm.rs parser change) and both thinking-effort commits (0ad81c341, ad13ed841). It has been independently verified and redteamed at head a360bf82a — CI fully green, live serve→join→inference proven on hardware — and is awaiting human approval to merge.

What still needs a new PR (part 2) — the agent/ACP reply-behavior work from this branch:

the typed send_message tool in dev-mcp (7b7b56f89) + its BUZZ_ACP_SEND_MESSAGE_TOOL gating (239542e83)
the plain-reply delivery fallback in buzz-acp behind BUZZ_ACP_DELIVER_PLAIN_REPLIES (bb4af4189, a11e7b85e)
the mesh_agent_e2e P5/P6 rewrite (bf735e235)
the crates/buzz-agent/src/llm.rs reasoning→text promotion from 2cd640b23 — deliberately excluded from feat(mesh): upgrade embedded mesh to v0.74 and harden shared compute (split 1/2 of #3467) #3741 because it unconditionally affects every OpenAI-compat/Responses provider, so it deserves review as the behavior change it is
the two relay_mesh.rs opt-in env lines for the flags above

None of that work is lost — it's all on micspiral/mesh-0-74-gemma and can be cherry-picked onto main once #3741 lands (the only overlap is relay_mesh.rs and the e2e example, both trivial).

Thanks Mic — the commit hygiene on this branch is what made the split clean.

…chive * origin/main: (22 commits) feat(catalog): resolve publisher display name in catalog detail pane (#3640) feat(mesh): upgrade embedded mesh to v0.74 and harden shared compute (split 1/2 of #3467) (#3741) docs(nips): specify kind:30621 multi-repo projects (NIP-MP) (#3163) Refine agent sharing dialog (#3699) desktop: enable getUserMedia in the Linux WebKitGTK webview (#3607) fix: align responsive agent views (#3688) Add macOS agent menu-bar menu (#3565) Fix pending message feedback (#3543) fix(desktop): remove remaining Projects panel fills (#3742) feat(mobile): desktop-parity emoji and thread experience (#3485) desktop: restore direct community member adds (#3634) fix(desktop): explain open agent access (#2561) fix(cli): resolve agents from owner records (#3178) fix(desktop): remove Projects overview card fills (#3416) feat(replica): portable heartbeat-token fence with snapshot-local reader routing (#3268) fix(git): channel binding tooling + author remediation for unbound repos (#3626) feat: configure S3 URL addressing style (#3400) feat: add first-class OpenRouter provider support (#1975) feat(agent,acp): wire provider total_tokens through NIP-AM publish chain (#3593) chore(release): release Buzz Desktop version 0.5.2 (#3624) ... Signed-off-by: Will Pfleger <pfleger.will@gmail.com>

…g-pipeline * origin/main: feat(catalog): resolve publisher display name in catalog detail pane (#3640) feat(mesh): upgrade embedded mesh to v0.74 and harden shared compute (split 1/2 of #3467) (#3741) docs(nips): specify kind:30621 multi-repo projects (NIP-MP) (#3163) Signed-off-by: npub1mn7jgtj4w2pd0g0zeuhxsa6jy6p0rewxz4kujt98my82ahfmp72sxjexk7 <dcfd242e557282d7a1e2cf2e6877522682f1e5c6156dc92ca7d90eaedd3b0f95@buzz.block.builderlab.xyz>

…-style * origin/main: fix(acp): preserve truncated thread context (block#3340) feat(catalog): resolve publisher display name in catalog detail pane (block#3640) feat(mesh): upgrade embedded mesh to v0.74 and harden shared compute (split 1/2 of block#3467) (block#3741) docs(nips): specify kind:30621 multi-repo projects (NIP-MP) (block#3163) Refine agent sharing dialog (block#3699) desktop: enable getUserMedia in the Linux WebKitGTK webview (block#3607) fix: align responsive agent views (block#3688) Add macOS agent menu-bar menu (block#3565) Fix pending message feedback (block#3543) fix(desktop): remove remaining Projects panel fills (block#3742) feat(mobile): desktop-parity emoji and thread experience (block#3485) desktop: restore direct community member adds (block#3634) fix(desktop): explain open agent access (block#2561) fix(cli): resolve agents from owner records (block#3178) fix(desktop): remove Projects overview card fills (block#3416) feat(replica): portable heartbeat-token fence with snapshot-local reader routing (block#3268) Signed-off-by: Joah Gerstenberg <joah@squareup.com>

* origin/main: (59 commits) Fix video reviews in thread replies (#3719) feat(release): make desktop releases immutable (#3568) Make relay reconnect backoff authoritative (#3774) feat(desktop): add password-protected backups in settings (#3701) fix(desktop): reuse profiles when joining communities (#2155) Render mobile agent mention chips (#3702) fix(catalog): update Amp description (#3758) fix(acp): preserve truncated thread context (#3340) feat(catalog): resolve publisher display name in catalog detail pane (#3640) feat(mesh): upgrade embedded mesh to v0.74 and harden shared compute (split 1/2 of #3467) (#3741) docs(nips): specify kind:30621 multi-repo projects (NIP-MP) (#3163) Refine agent sharing dialog (#3699) desktop: enable getUserMedia in the Linux WebKitGTK webview (#3607) fix: align responsive agent views (#3688) Add macOS agent menu-bar menu (#3565) Fix pending message feedback (#3543) fix(desktop): remove remaining Projects panel fills (#3742) feat(mobile): desktop-parity emoji and thread experience (#3485) desktop: restore direct community member adds (#3634) fix(desktop): explain open agent access (#2561) ... # Conflicts: # desktop/scripts/check-file-sizes.mjs

* origin/main: (29 commits) feat(desktop): raise the install ceiling and make installs observable (#3368) fix(db): isolate usage metrics advisory-lock test on scratch DB (#3670) Add Devin as a preset ACP harness (#3225) feat(desktop): improve agent activity header ui (#3321) perf(presence): reduce heartbeat frequency (#3783) Tighten continuation message rows (#3724) Fix video reviews in thread replies (#3719) feat(release): make desktop releases immutable (#3568) Make relay reconnect backoff authoritative (#3774) feat(desktop): add password-protected backups in settings (#3701) fix(desktop): reuse profiles when joining communities (#2155) Render mobile agent mention chips (#3702) fix(catalog): update Amp description (#3758) fix(acp): preserve truncated thread context (#3340) feat(catalog): resolve publisher display name in catalog detail pane (#3640) feat(mesh): upgrade embedded mesh to v0.74 and harden shared compute (split 1/2 of #3467) (#3741) docs(nips): specify kind:30621 multi-repo projects (NIP-MP) (#3163) Refine agent sharing dialog (#3699) desktop: enable getUserMedia in the Linux WebKitGTK webview (#3607) fix: align responsive agent views (#3688) ... Signed-off-by: npub1g8493u0xfsjrvflg4n08ezd7vec99mnwzlv0qgwpr9d7gvjwhuzqx59rhw <41ea58f1e64c243627e8acde7c89be667052ee6e17d8f021c1195be4324ebf04@buzz.block.builderlab.xyz>

@wren

…3763) ## Why A Buzz agent's assistant text and reasoning are never shown to anyone — only what it posts through the CLI. A turn that runs fifteen tool calls and never publishes is a silent failure: the requester waits on a result that was produced and thrown away. This adds an optional reminder at the end-of-turn gate, off by default. Tyler asked for it in buzz-mesh; plan iterated to **9.5/10 with @wren** (Minimalness 9.7, Elegance 9.5, Correctness 9.3). ## What `BUZZ_AGENT_REQUIRE_REPLY=1` (default off, per-agent opt-in). A turn about to end with no recognized attempt to post gets a reminder and is rerolled. **At most two, then the turn ends regardless** — the guard catches accidental omission, it does not compel speech. The reminder text explicitly licenses silence so it cannot fight the base prompt's "silence is usually correct." **This is not a new MCP hook.** `RunCtx::run` *is* the turn, so the two per-turn locals need no plumbing, and every tool call already passes through it with arguments visible. The objection is appended at the existing `_Stop` gate and rides `push_hook_outputs_as_tool_results`, so the model receives it as a lower-trust tool result with `{hook, server, text}` attribution. No new trust path, no new lifecycle event, no dev-mcp or CLI protocol change. Earlier revisions of this plan needed four crates (a `_UserPromptSubmit` hook, a marker file, a `buzz-cli` change, dev-mcp state). Tyler pointed out the agent already knows both facts; that deleted all of it. Net runtime change is ~35 lines in `agent.rs` + ~4 in `config.rs`. ### Recognition contract A registered non-hook tool whose qualified name ends in `__shell`, whose `command` argument contains `messages send` or `reactions add`. - **The `__` separator is exact, not approximate.** Given `has()` + `!is_hook()`, `ends_with("__shell")` is *provably equivalent* to a bare name of `shell`: registration forbids `__` in server and bare names (`mcp.rs:227,268`) and qnames are `{server}__{bare}`, so a trailing `__shell` could only straddle the separator if the bare name began with `_` — which `is_hook` excludes. Without the separator, `powershell` and `noshell` would match. - **Reads the structured `command` field**, not serialized arguments, so a `description` that quotes a send cannot disarm the guard, and a non-string `command` is rejected rather than coerced. - **Detects an attempt, not a successful publish.** A failed send already returns non-zero exit and error JSON — louder than this reminder. The variable is named `buzz_reply_call_seen` so the code can't pretend otherwise. - **Checked after the per-turn tool-call cap**, since a discarded call never ran. - `messages send` also covers `messages send-diff`. Reactions count because the base prompt directs agents to react rather than post a bare acknowledgement. **Known limits, both deliberate and documented:** a command assembled at runtime (`$CMD`) or hidden in a wrapper script is missed; text that merely quotes a send (`echo "buzz messages send"`) matches. Missing a real post is the expensive direction and substring matching is the forgiving one there. Neither edge is pinned by a test, so the matcher stays free to improve. ### Budget Reminders share `BUZZ_AGENT_STOP_MAX_REJECTIONS`, the existing outer cap on every end-turn objection. Default 3 fits both; at 1 only one fits; at 0 the guard is off with the hooks. A round carrying both a hook objection and a reminder costs one rejection and delivers both texts. An independent budget would either violate that bound or need a second arbitration rule. ## Prior art - **#3467** (closed) built the same detector one layer up in `buzz-acp` for a different remedy. None of its symbols are on main — this borrows its permission to be coarse, but reads structured data that ACP didn't have. - **#3648** (open) detects turns with *no output at all*; a turn with fifteen tool calls and no post counts as output there, so it does not cover this case. - **#3741** (merged) is mesh-only. ## Testing **14 new tests.** 4 unit tests on the matcher; 10 integration tests through the ACP wire harness: off by default, `=0` still off, opted-in silent → exactly 2 reminders then `end_turn`, registered `fake__shell` send → 0 reminders, hallucinated `fake__shell` → still reminded, publish call truncated past the 64-call cap → still reminded, budget 1 → 1 reminder, budget 0 → off, combined `_Stop` hook objection + reminder → one round both texts and after 2 reminders the hook objection continues alone, unparseable `=true` → startup error naming the key. **10 mutation checks, each breaking a specific named test** — neutralize the nag cap, stop sharing the budget, neutralize `buzz_reply_call_seen`, drop `has`/`is_hook`, ignore the flag, drop the `__`, drop `reactions add`, read serialized args, move detection before truncation. `tests/bin/fake_mcp.rs` gains `FAKE_MCP_SHELL_TOOL=1`: it previously exposed no tool with a bare name of `shell`, so the satisfied-guard path was untestable. Full `cargo test -p buzz-agent` green at 9e0ae1f; clippy `-D warnings` and `cargo fmt --check` clean. **Unrelated flake found:** `cancelled_turn_with_usage_emits_notification_before_response` (`tests/fake_llm.rs`) is timing-sensitive. Under 10 loaded cores it fails **2/20 on this branch and 1/20 at unmodified `origin/main@02be413`** — pre-existing, not caused by this change (which is inert without the env var). Flagging so it isn't misattributed to the next PR that's open when CI hits it. ## Docs `crates/buzz-agent/README.md` is the primary home (env var, recognition contract, limits, budget interaction). `docs/MCP_DRIVEN_HOOKS.md` gets a short cross-reference explaining this is *not* a hook — otherwise readers hunt for a `_ReplyGuard` tool that doesn't exist. --------- Signed-off-by: tlongwell-block <109685178+tlongwell-block@users.noreply.github.com> Signed-off-by: npub1mprnacetjua2xx3p5eddmhxyk6wv929ymm5py8kd2xfxurxahspqqlgyta <d8473ee32b973aa31a21a65adddcc4b69cc2a8a4dee8121ecd51926e0cddbc02@buzz.block.builderlab.xyz> Co-authored-by: Dawn (sprout agent) <c6237ef84fa537c78dcee78efd2d4e59f728859c7f194da42ac51ededfa0be05@sprout-oss.stage.blox.sqprod.co> Co-authored-by: npub1mprnacetjua2xx3p5eddmhxyk6wv929ymm5py8kd2xfxurxahspqqlgyta <d8473ee32b973aa31a21a65adddcc4b69cc2a8a4dee8121ecd51926e0cddbc02@buzz.block.builderlab.xyz>

@michaelneale

…(split 1/2 of block#3467) (block#3741) ## Summary This is **part 1 of 2** split out from block#3467 (per Tyler's request), carrying only the mesh-scoped changes. The agent/ACP response-behavior changes and the new `send_message` tool stay in block#3467 as part 2. All commits are @michaelneale's work, cherry-picked with authorship preserved. - Upgrade embedded Mesh to v0.74.0 (tag-pinned instead of commit rev) and use canonical Gemma model IDs. - Keep shared compute serving through member joins, roster changes, app recovery, and community switching. - Wait for actual model readiness and avoid resuming incomplete downloads after quit. - Leave `BUZZ_AGENT_THINKING_EFFORT` unset by default so each model's chat template picks its own thinking default (`none` suppressed Gemma tool-calling entirely; pinning `low` made Qwen3 burn ~4x output budget). Explicit agent/persona/global values still win. ## Relationship to block#3467 Contains the mesh commits from block#3467 (`2cd640b23`, `0ad81c341`, `ad13ed841`) rebased onto current main, with one deliberate exclusion: the `crates/buzz-agent/src/llm.rs` reasoning→text parser change from `2cd640b23` is **not** here. That change unconditionally affects every OpenAI-compat/Responses provider, so it belongs with the reply-behavior work in part 2, where it can be reviewed as what it is. Not included (remaining in block#3467 / part 2): - typed `send_message` tool in dev-mcp + `BUZZ_ACP_SEND_MESSAGE_TOOL` gating - plain-reply delivery fallback in buzz-acp (`BUZZ_ACP_DELIVER_PLAIN_REPLIES`) - the mesh_agent_e2e P5/P6 rewrite (exists to prove the reply path) - the two `env.insert` preset opt-ins in `relay_mesh.rs` for the flags above - the llm.rs parser change This PR is independently mergeable; part 2's flags are all off by default so it can land before or after. ## Testing - `cargo test -p buzz-relay --locked` — 780 passed (one telemetry test is order-sensitive under parallel default settings; passes in the pre-push suite and standalone, unrelated to this diff — files untouched here). - `just desktop-tauri-test` (default features) — 1877 passed. - `cargo test --locked --features mesh-llm` in `desktop/src-tauri` — 1961 passed, including the new relay-mesh preset and coordinator/recovery tests. - Both `Cargo.lock`s resolve with `--locked` against the v0.74.0 tag. - Full pre-push hook suite green (rust-tests, desktop-check/test, tauri checks). Live validation of the mesh v0.74 upgrade itself is documented on block#3467 (two-Mac cross-version test). --------- Signed-off-by: Michael Neale <michael.neale@gmail.com> Signed-off-by: Tyler Longwell <tlongwell@block.xyz> Co-authored-by: Michael Neale <michael.neale@gmail.com> Co-authored-by: Tyler Longwell <tlongwell@block.xyz>

@wren

…lock#3763) ## Why A Buzz agent's assistant text and reasoning are never shown to anyone — only what it posts through the CLI. A turn that runs fifteen tool calls and never publishes is a silent failure: the requester waits on a result that was produced and thrown away. This adds an optional reminder at the end-of-turn gate, off by default. Tyler asked for it in buzz-mesh; plan iterated to **9.5/10 with @wren** (Minimalness 9.7, Elegance 9.5, Correctness 9.3). ## What `BUZZ_AGENT_REQUIRE_REPLY=1` (default off, per-agent opt-in). A turn about to end with no recognized attempt to post gets a reminder and is rerolled. **At most two, then the turn ends regardless** — the guard catches accidental omission, it does not compel speech. The reminder text explicitly licenses silence so it cannot fight the base prompt's "silence is usually correct." **This is not a new MCP hook.** `RunCtx::run` *is* the turn, so the two per-turn locals need no plumbing, and every tool call already passes through it with arguments visible. The objection is appended at the existing `_Stop` gate and rides `push_hook_outputs_as_tool_results`, so the model receives it as a lower-trust tool result with `{hook, server, text}` attribution. No new trust path, no new lifecycle event, no dev-mcp or CLI protocol change. Earlier revisions of this plan needed four crates (a `_UserPromptSubmit` hook, a marker file, a `buzz-cli` change, dev-mcp state). Tyler pointed out the agent already knows both facts; that deleted all of it. Net runtime change is ~35 lines in `agent.rs` + ~4 in `config.rs`. ### Recognition contract A registered non-hook tool whose qualified name ends in `__shell`, whose `command` argument contains `messages send` or `reactions add`. - **The `__` separator is exact, not approximate.** Given `has()` + `!is_hook()`, `ends_with("__shell")` is *provably equivalent* to a bare name of `shell`: registration forbids `__` in server and bare names (`mcp.rs:227,268`) and qnames are `{server}__{bare}`, so a trailing `__shell` could only straddle the separator if the bare name began with `_` — which `is_hook` excludes. Without the separator, `powershell` and `noshell` would match. - **Reads the structured `command` field**, not serialized arguments, so a `description` that quotes a send cannot disarm the guard, and a non-string `command` is rejected rather than coerced. - **Detects an attempt, not a successful publish.** A failed send already returns non-zero exit and error JSON — louder than this reminder. The variable is named `buzz_reply_call_seen` so the code can't pretend otherwise. - **Checked after the per-turn tool-call cap**, since a discarded call never ran. - `messages send` also covers `messages send-diff`. Reactions count because the base prompt directs agents to react rather than post a bare acknowledgement. **Known limits, both deliberate and documented:** a command assembled at runtime (`$CMD`) or hidden in a wrapper script is missed; text that merely quotes a send (`echo "buzz messages send"`) matches. Missing a real post is the expensive direction and substring matching is the forgiving one there. Neither edge is pinned by a test, so the matcher stays free to improve. ### Budget Reminders share `BUZZ_AGENT_STOP_MAX_REJECTIONS`, the existing outer cap on every end-turn objection. Default 3 fits both; at 1 only one fits; at 0 the guard is off with the hooks. A round carrying both a hook objection and a reminder costs one rejection and delivers both texts. An independent budget would either violate that bound or need a second arbitration rule. ## Prior art - **block#3467** (closed) built the same detector one layer up in `buzz-acp` for a different remedy. None of its symbols are on main — this borrows its permission to be coarse, but reads structured data that ACP didn't have. - **block#3648** (open) detects turns with *no output at all*; a turn with fifteen tool calls and no post counts as output there, so it does not cover this case. - **block#3741** (merged) is mesh-only. ## Testing **14 new tests.** 4 unit tests on the matcher; 10 integration tests through the ACP wire harness: off by default, `=0` still off, opted-in silent → exactly 2 reminders then `end_turn`, registered `fake__shell` send → 0 reminders, hallucinated `fake__shell` → still reminded, publish call truncated past the 64-call cap → still reminded, budget 1 → 1 reminder, budget 0 → off, combined `_Stop` hook objection + reminder → one round both texts and after 2 reminders the hook objection continues alone, unparseable `=true` → startup error naming the key. **10 mutation checks, each breaking a specific named test** — neutralize the nag cap, stop sharing the budget, neutralize `buzz_reply_call_seen`, drop `has`/`is_hook`, ignore the flag, drop the `__`, drop `reactions add`, read serialized args, move detection before truncation. `tests/bin/fake_mcp.rs` gains `FAKE_MCP_SHELL_TOOL=1`: it previously exposed no tool with a bare name of `shell`, so the satisfied-guard path was untestable. Full `cargo test -p buzz-agent` green at 9e0ae1f; clippy `-D warnings` and `cargo fmt --check` clean. **Unrelated flake found:** `cancelled_turn_with_usage_emits_notification_before_response` (`tests/fake_llm.rs`) is timing-sensitive. Under 10 loaded cores it fails **2/20 on this branch and 1/20 at unmodified `origin/main@02be413`** — pre-existing, not caused by this change (which is inert without the env var). Flagging so it isn't misattributed to the next PR that's open when CI hits it. ## Docs `crates/buzz-agent/README.md` is the primary home (env var, recognition contract, limits, budget interaction). `docs/MCP_DRIVEN_HOOKS.md` gets a short cross-reference explaining this is *not* a hook — otherwise readers hunt for a `_ReplyGuard` tool that doesn't exist. --------- Signed-off-by: tlongwell-block <109685178+tlongwell-block@users.noreply.github.com> Signed-off-by: npub1mprnacetjua2xx3p5eddmhxyk6wv929ymm5py8kd2xfxurxahspqqlgyta <d8473ee32b973aa31a21a65adddcc4b69cc2a8a4dee8121ecd51926e0cddbc02@buzz.block.builderlab.xyz> Co-authored-by: Dawn (sprout agent) <c6237ef84fa537c78dcee78efd2d4e59f728859c7f194da42ac51ededfa0be05@sprout-oss.stage.blox.sqprod.co> Co-authored-by: npub1mprnacetjua2xx3p5eddmhxyk6wv929ymm5py8kd2xfxurxahspqqlgyta <d8473ee32b973aa31a21a65adddcc4b69cc2a8a4dee8121ecd51926e0cddbc02@buzz.block.builderlab.xyz>

…ny — with real nodes (#3862) ## Summary CI now proves the full Buzz shared-compute join story end to end: a member can discover another member's served model **through the Buzz relay alone** and run inference over the mesh, while a non-member gets nothing — the relay rejects its auth, and the mesh refuses to route for it even holding a leaked endpoint address. This is deliberately different from mesh-llm's own CI smokes (which bootstrap two nodes with a hand-carried invite token / mdns): here the **relay is the control plane**, exactly like the desktop app: 1. **Membership** — identities A and B are added via `buzz-admin` (kind:13534 NIP-43 roster); C is not. 2. **Advertise** — each member publishes a client-signed kind:30003 discovery note carrying its MeshLLM owner binding and (for the serve node) `serveTargets[].endpointAddr`, covered by an endpoint-binding signature — the exact payload shape the desktop coordinator publishes. 3. **Trust** — the serve node derives its admission allowlist from the relay (statuses ∩ roster) and requires the **exact expected {A, B} owner-id set** before starting with `TrustPolicy::Allowlist`. 4. **Join** — the client verifies owner + endpoint bindings and membership, then dials the relay-discovered endpoint (the desktop join-watcher's `dial_endpoint_addr` step). No out-of-band token. 5. **Infer** — a chat completion against the client's local OpenAI endpoint routes over QUIC to the serve node's model (CPU, SmolLM2-135M, ~105MB). 6. **Deny (differential)** — the stranger's NIP-42 auth must fail with the relay's own membership rejection (`restricted: not a relay member` — successful auth or any unrelated connect error fails the run), and dialing the leaked endpoint must not produce a routed inference — **while the trusted client re-proves inference immediately afterwards**, so a dead serve node can't masquerade as an admission denial. ## What's in the PR - `crates/buzz-relay/examples/mesh_relay_lifecycle_smoke.rs` — the harness. One process per node (mesh-llm keeps process-global state under `~/.mesh-llm`), orchestrator + serve/client/stranger roles, byte-identical binding payloads to `desktop/src-tauri/src/mesh_llm/identity.rs` (called out with keep-in-sync comments). Child stdout is pumped through a reader thread so every wait has a hard deadline; timed-out children are killed; exit statuses are checked. - `scripts/ci-mesh-lifecycle-smoke.sh` — provisions a membership-gated relay (throwaway owner + signing identities via `buzz-admin generate-key`), runs the harness, cleans up. Fails fast if :3000 is already occupied (a stale open relay would mask gating). - `scripts/start-relay-for-tests.sh` — gains opt-in NIP-43 membership env passthrough (`BUZZ_REQUIRE_RELAY_MEMBERSHIP` + `RELAY_OWNER_PUBKEY` + `BUZZ_RELAY_PRIVATE_KEY`). Default behavior unchanged. - `.github/workflows/mesh-lifecycle.yml` — separate, path-filtered, non-required workflow (mesh paths, the harness's dependency crates, `Cargo.lock`, dispatch), pinned to `ubuntu-24.04`. Caches the mesh native runtime + HF model keyed on the lockfile hash, so a mesh pin bump rolls the runtime cache. Uploads relay + harness logs on failure. ## Scope This is an **independent protocol harness**: it speaks the same wire protocol and payload shapes as the desktop but re-implements the binding/verification logic (the desktop crate is outside the workspace). Regressions inside the desktop's own discovery filtering are the desktop unit tests' job; what this smoke proves is that the relay + mesh-llm SDK + admission stack support the lifecycle end to end. ## Relationship to mesh-llm's CI Follows the shape mesh-llm's own CI proved stable (tiny CPU model, one runner, multiple real mesh-llm processes over real QUIC — cf. their `ci-two-node-client-serving-smoke.sh`), but swaps the token bootstrap for the relay-driven lifecycle, which is the part only Buzz can test. ## Validation Green on GitHub Actions (ubuntu-24.04) across three runs, including after rebases onto the mesh v0.74 upgrade (#3467) and latest main: ``` PASS 1/6: relay-derived allowlist is exactly {A, B} PASS 2/6: serve member ready + advertised model: jc-builds/SmolLM2-135M-Instruct-Q4_K_M-GGUF:Q4_K_M PASS 3/6: client member discovered + joined via relay PASS 4/6: inference routed over the mesh: "PONG" PASS 5/6: relay rejected the stranger's NIP-42 auth (membership gate) PASS 6/6: stranger denied (gossip visible, inference rejected: 503 all tunnels failed) while trusted inference still routes PASS: full relay-driven mesh lifecycle verified ``` Also validated locally on macOS. `cargo fmt --all --check` and `cargo clippy -p buzz-relay --all-targets -- -D warnings` pass. ## Notes - The harness follows the repo's mesh `[dev-dependencies]` pin automatically, so it doubles as a canary for future mesh upgrades (it already caught the v0.73.1 → v0.74.0 bump during development). - The stranger "deny" accepts either shape mesh-llm exhibits: no model visibility at all, or gossip visibility with inference refused — mesh-llm applies the receiving node's owner policy after the gossip handshake, so admission gates *routing*, not gossip. The differential trusted-inference re-check (PASS 6/6) is what makes that a real denial rather than a dead server. - Model-visibility windows are tunable via `MESH_CLIENT_WINDOW_SECS` / `MESH_STRANGER_WINDOW_SECS` if shared runners prove slow — pin a longer window in the workflow env rather than re-running the job. --------- Signed-off-by: Michael Neale <michael.neale@gmail.com>

michaelneale force-pushed the micspiral/mesh-0-74-gemma branch from f79a72c to 59eb49c Compare July 29, 2026 13:11

micspiral changed the title ~~feat(mesh): upgrade to v0.74 and validate Buzz replies~~ feat(mesh): upgrade to v0.74 and make shared agents reply reliably Jul 29, 2026

micspiral marked this pull request as ready for review July 29, 2026 13:12

micspiral requested a review from a team as a code owner July 29, 2026 13:12

micspiral requested a review from tlongwell-block July 29, 2026 13:17

michaelneale added 4 commits July 30, 2026 05:55

michaelneale force-pushed the micspiral/mesh-0-74-gemma branch from 0136bef to 0ad81c3 Compare July 29, 2026 20:03

michaelneale added 3 commits July 30, 2026 10:47

style(mesh): rustfmt the shared-compute delivery assertion

a11e7b8

cargo fmt --all skips desktop/src-tauri (separate workspace), so this line only failed under just desktop-tauri-fmt-check in CI's Rust Lint job. Signed-off-by: Michael Neale <michael.neale@gmail.com>

micspiral enabled auto-merge (squash) July 30, 2026 05:16

tlongwell-block mentioned this pull request Jul 30, 2026

feat(mesh): upgrade embedded mesh to v0.74 and harden shared compute (split 1/2 of #3467) #3741

Merged

tlongwell-block closed this Jul 30, 2026

auto-merge was automatically disabled July 30, 2026 15:38
Pull request was closed

tlongwell-block mentioned this pull request Jul 30, 2026

feat(agent): optional reply guard reminds a silent turn to publish #3763

Merged

github-actions Bot mentioned this pull request Jul 30, 2026

[upstream-sync] upstream-sync 2026-07-30: 16 commits from block/buzz adrienlacombe/buzz#5

Closed

michaelneale mentioned this pull request Jul 31, 2026

ci: prove the relay-driven mesh lifecycle — discover, join, infer, deny — with real nodes #3862

Merged

wpfleger96 mentioned this pull request Jul 31, 2026

chore(release): release Buzz Relay version 0.2.1 #2856

Open

github-actions Bot mentioned this pull request Aug 2, 2026

[Buzz update] desktop-v0.5.3 available BrianInAz/buzz#9

Open

Navigation Menu

Uh oh!

feat(mesh): upgrade to v0.74 and make shared agents reply reliably - #3467

feat(mesh): upgrade to v0.74 and make shared agents reply reliably#3467
micspiral wants to merge 8 commits into
mainfrom
micspiral/mesh-0-74-gemma

micspiral commented Jul 29, 2026 •

edited

Loading

Uh oh!

micspiral commented Jul 30, 2026

Uh oh!

micspiral commented Jul 30, 2026

Uh oh!

tlongwell-block commented Jul 30, 2026

Uh oh!

tlongwell-block commented Jul 30, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Uh oh!

Conversation

micspiral commented Jul 29, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Live validation

Testing

Uh oh!

micspiral commented Jul 30, 2026

Uh oh!

micspiral commented Jul 30, 2026

Uh oh!

tlongwell-block commented Jul 30, 2026

Uh oh!

tlongwell-block commented Jul 30, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

micspiral commented Jul 29, 2026 •

edited

Loading