mirror of
https://github.com/block/buzz.git
synced 2026-10-29 13:31:44 +00:00
wpfleger/deterministic-agent-commit-identity
143 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Duncan
|
88f9341b2a |
feat(acp): enforce deterministic agent git commit identity
Managed agents commit through a Git wrapper that pins author/committer to the agent's Nostr identity, signs with the agent key, and refuses identity overrides, so commit attribution no longer depends on host Git config. Co-authored-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Will Pfleger <pfleger.will@gmail.com> |
||
|
Logan Johnson
|
ea1e97e65f |
feat(acp): run one prepared task from a file or stdin (#7851)
🤖 ## Summary Scripts, scheduled jobs, and other automation need to give an agent one prepared task and wait for a result without starting a long-running chat listener. This adds `buzz-acp run`: it reads a task from a local file or stdin, runs one fresh agent session with the configured tools and instructions, cleans up, and exits with a machine-readable status. ```sh buzz-acp run --task ./task.json cat task.json | buzz-acp run --task - ``` - The versioned JSON task specifies a task ID, intended agent public key, prepared prompt, and execution time limit. Invalid input or an identity mismatch is rejected before agent launch. Credentials, executable, tools, and permissions remain controlled by trusted launch settings, not the task document. - Each invocation uses the normal agent equipment: base/persona/team instructions, enabled memory, tools, model and effort settings, permissions, and Git identity, credentials, and signing. Both task and conversational modes obtain these capabilities from one shared runtime owner; new shared setup belongs there rather than in separate entry points. - Launch configuration reads have a separate ten-second bound and can be cancelled even when a prompt file blocks. A deadline covers adapter startup through the task turn. Interrupts and termination signals trigger bounded cleanup. Stdout contains one terminal JSON record; diagnostics stay on stderr. A normal turn ending is not proof that the task achieved its goal. - Normal invocation remains the conversational service. Task mode does not start chat subscriptions, presence, heartbeat turns, or a persistent agent pool. Tasks are trusted executable instructions, not sandboxed input. Reusing a task ID does not prevent duplicate execution, and uncertain outcomes must not be blindly retried. See [the task contract](crates/buzz-acp/TASKS.md) for the document format, exit codes, limits, and examples. URL retrieval, a task server, and a persistent execution ledger are outside this change. ### Related issue None found in the existing issue and PR search. ### Testing - At `4e86bbc97`: 964 unit tests and 22 integration tests passed, with three existing ignored tests. The package suite ran serially with inherited harness/Git options cleared. A final test-only shutdown-result assertion was then checked with the focused memory test and all-target Clippy. - All eight real-binary task tests independently passed on the exact candidate, including blocked configuration reads, file/stdin validation, session equipment, deadlines, cancellation, and cleanup. The memory test passed 20 repeated runs. Removing the oversized-input guard fails the new error assertion; restoring it passes. - Package all-target Clippy, formatting, file-size checks, and pre-push Rust/Tauri checks passed. All executed hosted checks passed on the updated head (54 successful, 28 skipped). The separate Codex review comment still requests authorization for this exact range. - Earlier revisions passed the release build and a real Goose 1.51.0 tool-use smoke test. No new release artifact or installed Janet runtime is claimed for this feedback revision. - Repository-wide `just ci` was previously attempted; workspace Clippy passed, then the desktop dependency check stopped on a local pnpm fetch failure. Hosted checks cover the applicable PR lanes. --------- Signed-off-by: Logan Johnson <loganj@squareup.com> Co-authored-by: Larry <627498bd4bd1f281a16431e3c6cce3b5c25b6692798c78672298aefbf2f8f8b5@buzz.block.builderlab.xyz> |
||
|
Salman Mohammed
|
5621006bcf |
refactor: move agent Git bootstrap into ACP harness (#7819)
discovered in https://github.com/block/buzz/issues/7742#issuecomment-5783512985 ## Summary Goose uses its native shell, so it bypassed the Git identity and signing setup inside `buzz-dev-mcp`. Agent commits could use the host Git identity, while MCP shells had a separate config. This moves Git setup into `buzz-acp`, where both paths receive the same agent identity, Nostr signer, and relay-scoped credential helper. Before: ```text Buzz Desktop └─ buzz-acp ← Desktop supplied relay Git authentication ├─ Goose → native shell ← no agent Git author or signing setup └─ Buzz Agent └─ buzz-dev-mcp └─ shell ← MCP created the keyfile and Git signing config ``` After: ```text Buzz Desktop or Sprig └─ buzz-acp ← owns the keyfile, identity, signing, and scoped credentials ├─ Goose → native shell ← inherits the complete Git environment └─ Buzz Agent └─ buzz-dev-mcp └─ shell ← receives the same Git environment via mcpServers[].env ``` The harness keeps the signing key in a private temporary file and removes it on normal exit, startup failure, and graceful termination. It composes inherited `GIT_CONFIG_*` entries, forwards the complete block to MCP servers, and bundles the helpers through the ACP multicall entrypoint. For standard `buzz-acp` launches, Desktop and Sprig no longer write Git config. A custom Desktop ACP command bypasses the harness, so Desktop preserves its prior relay-scoped credential helper for that supported override. This changes Git settings for every ACP runtime. The startup signal path also changes so a termination during adapter initialization removes the keyfile. The temporary file can remain after SIGKILL or a machine crash. ### Related issue Related: #6177 covers stricter commit identity enforcement and touches the same Git paths. That PR and this one need to be reconciled before either lands. No issue was found for this runtime bootstrap gap. ### Testing - Ran the ignored `git_runtime_tests` with real Goose 1.50.0 and Buzz Agent ACP sessions. `real_goose_native_git_shell` passed again on this PR branch with `BUZZ_TEST_BIN_DIR` pointing to built binaries. Goose made a signed commit and tag through its native shell; Buzz Agent did the same through its MCP shell. Both verified signatures, identity, credential scoping, and keyfile cleanup. - Built the Linux Sprig image and ran `scripts/test-sprig-image.sh buzz-sprig:acp-git-bootstrap`. The container made and verified a signed commit and tag, then confirmed keyfile removal after SIGTERM. - `just ci` passed on the original branch. After rebasing onto `main` and addressing review feedback, the ACP Git bootstrap integration tests, focused desktop smoke test, custom-command regression test, formatting, and ACP/Desktop Clippy passed. The pre-push hook hit two timing-sensitive ACP deadline tests under parallel load; both passed alone and the complete ACP package passed serially (964 library, 2 Git bootstrap, 9 pool lifecycle tests). The push used `--no-verify` after those checks; GitHub CI is rerunning on the new head. - Under injected `BUZZ_ACP_ALLOWED_RESPOND_TO`, `BUZZ_ACP_LAZY_POOL`, and `BUZZ_ACP_IDLE_POOL_SLEEP`, the clean-env serial package command in `crates/buzz-acp/TESTING.md` passed: 964 active library tests, 2 Git bootstrap tests, and 9 pool lifecycle tests. Both named Databricks OAuth tests passed locally; another reviewer reproduced their environment-specific failures on `main`. - No authenticated relay clone or push was run. The local probes used disposable repositories and did not contact a relay. Generated with Codex --------- Signed-off-by: Salman Mohammed <smohammed@squareup.com> |
||
|
Salman Mohammed
|
4d08194ea9 |
Deduplicate thread context by ACP session delivery (#7620)
## Summary Buzz can repeat a thread reply in later prompts after the provider already received that event as context. The old fix treated every message signed by the agent key as retained by the current provider session. That could hide an unseen heartbeat reply because independent sessions share the same signing key. This PR now removes only exact event IDs already delivered to the live ACP session. It uses bounded overfetch so removed events do not shrink the new-context window, and it preserves unseen same-key replies from other sessions. The context header reports `truncated=true` whenever session history was omitted. The harness calls `/count` only when the relay query finds a sentinel beyond its fetch window. This deliberately does not parse event IDs from shell output. A reply that the agent just published may appear once in the next context delta; after that event reaches the provider as context, later prompts omit it. --------- Signed-off-by: Salman Mohammed <smohammed@squareup.com> |
||
|
Salman Mohammed
|
deda09c18c |
Document incoming Buzz turn structure (#7624)
## Summary Document how agents should read Buzz’s semantic turn envelope. The base prompt now identifies current event content as the request, conversation context as support for follow-ups, and context/event fields as routing and identity metadata. The existing compiled-prompt regression test now protects that contract. The lockfile also updates rustls to 0.23.45 to clear RUSTSEC-2026-0285 from the security gate. |
||
|
Salman Mohammed
|
6c35e82bd5 |
fix(acp): improve hints in pi agent setup guide (#7594)
## Summary Handle missing Pi adapter setup and update its install guidance. ### Related issue None found. ### Testing No manual testing. Generated with Amp --------- Signed-off-by: Salman Mohammed <smohammed@squareup.com> Co-authored-by: Amp <amp@ampcode.com> |
||
|
Salman Mohammed
|
e17cdd9d5c |
refactor(buzz-acp): point agents at buzz --help instead of a command table (#7586)
## Summary The ACP base prompt carried a `Group | Key commands` table for the `buzz` CLI — a second copy of a surface that already documents itself, with nothing tying it to the parser. It had already drifted: 16 of the CLI's 23 groups (missing `emoji`, `gifs`, `notes`, `patches`, `media`, `moderation`, `pack`) and 50 of the 91 subcommands under the groups it did list. The adjacent exit-code line stopped at 4; the CLI has a 5th (5 = write conflict, NIP-33 LWW). #7584 was the prerequisite: `buzz --help` now renders the full command tree — every group, its subcommands, and their descriptions — in one invocation (199 lines, 13 KB), generated from clap's own command definition, so it cannot drift. `buzz -h` still prints the group-level summary. Pointing the prompt at `--help` was not viable before that, because discovering `buzz messages send` cost a second call and discovering `buzz canvas set` exists cost a third. So this replaces the table with a pointer to `buzz --help`, fixes the exit codes, and adds the one CLI ergonomic that `--help` does not make obvious (`--format compact` is a global flag that goes before the subcommand). Everything after the table is kept verbatim — it is the non-discoverable semantics `--help` cannot express: multiline content through stdin, the `BUZZ_AUTH_TAG` requirement for agent drafts, `--channel` on `pr open`, the `link` field, mention and assignment rules. Signed-off-by: Salman Mohammed <smohammed@squareup.com> Co-authored-by: Claude Opus 5 <noreply@anthropic.com> |
||
|
Salman Mohammed
|
813bbd1412 |
fix(acp): integrate the Buzz Pi adapter fork (#7552)
## Summary PR #7335 worked around missing Pi adapter support by generating a private Pi launcher and injecting Buzz's standing prompt and skills at process launch. The Buzz Pi fork now carries the required adapter extensions, so this removes that launcher and returns prompt construction to the normal ACP session path while retaining the base-prompt composition introduced by #7335. The Pi preset now installs `salman1993/pi-acp` and launches its renamed `buzz-pi-acp` binary. Buzz adds `-- --skill <harness-cwd>/.agents/skills` only when launching that binary, sends the complete composed prompt as the `_meta.systemPrompt` replacement string on `session/new` only when `initialize.agentInfo.name` is `buzz-pi-acp`, and sends the scoped title alongside it as `_meta.sessionTitle`. The fork identity is treated as system-prompt capable regardless of its reported ACP protocol version, which prevents duplicate legacy user-message framing. Upstream `pi-acp` does not receive either fork-specific behavior. Observer transcript projection accepts the string, `{ replace }`, and `{ append }` metadata forms. The fork now stores restore metadata in one atomic file per session under `~/.pi/buzz-pi-acp/sessions/`. This prevents concurrent Buzz workers from overwriting another session's prompt or title. The fix landed in [salman1993/pi-acp#9](https://github.com/salman1993/pi-acp/pull/9). This supersedes the closed #7508. No agent-configuration rules changed; this changes the Buzz Pi adapter contract and launch arguments. ### Related issue #7329 ### Testing Installed fork commit `09cf07e436b8f18e52401558f988f31a15702313` through the documented Git URL. The installed bundle matched the committed `dist/index.js` byte for byte and contained the `~/.pi/buzz-pi-acp` metadata path. The fork's 106 non-skipped tests, typecheck, and lint pass. Ran the ignored real-Pi integration test through Buzz's production session composer. The test exercised the renamed package, `agentInfo.name`, and the new per-session metadata store. Base, persona, team, core-memory, huddle, canvas, and skill markers each appeared once after switching sessions and again after restarting the adapter, while the other session and Pi's native default prompt were absent. Added regression coverage proving `buzz-pi-acp` receives fork-specific system-prompt metadata and managed skills while upstream `pi-acp` does not. `just ci` passes. Generated with Codex --------- Signed-off-by: Salman Mohammed <smohammed@squareup.com> |
||
|
Diem Nguyen
|
051c3a270b |
fix(buzz-acp): report missing models without retrying (#7538)
## Summary When an agent reports model-not-found, Buzz retries the unavailable model and delays the failure reply until retries are exhausted. Stop retrying this error and immediately post a threaded recovery notice. The notice tells users to select a different model in agent settings, save, restart the agent to apply the configuration, and re-send their request. This adds one error-handling branch and regression coverage in `buzz-acp`. It matches `-32002` errors containing `model not found`. Other resource-not-found errors, such as stale sessions, retain the existing retry behavior. Detailed error events remain available for diagnosis. The existing restart policy is unchanged. ### Related issue None found in existing issue/PR searches for model-not-found recovery. ### Testing Playwright captured and visually checked the thread UI with seeded conversation data and the exact recovery text. The check opens the request's thread, confirms no reply before the failure, injects the notice, and verifies the full text is visible. [Before/after screenshots](https://github.com/block/buzz/pull/7538#issuecomment-5608196506) show the corrected save-and-restart instructions. These are local test captures, not a deployed provider recovery flow. Generated with Codex --------- Signed-off-by: Diem Nguyen <diem@squareup.com> |
||
|
Logan Johnson
|
c045321a7f |
fix(acp): pace targeted overflow recovery on consumer capacity (#7325)
🤖 ## Summary When a Buzz agent falls behind on incoming messages, its connection can make the backlog worse while trying to recover. The connection buffers messages from the relay server until the agent is ready to process them; if that buffer overflows, recovery previously requested history for **every subscribed channel** and paused socket reads while sending those requests. That adds traffic to an already overloaded connection. This change requests history only for affected subscriptions, once the code consuming those messages has room, with at least five seconds between attempts. The recovery path now: - Combines repeated losses into one pending recovery per affected subscription, keeping the oldest dropped timestamp so replay starts early enough. - Waits until at least half the consumer queue is free and the relay's existing rate-limit delay has expired. The queue wakes recovery when space becomes available; recovery does not periodically sample capacity or hold queue space away from live messages. - Attempts one subscription at a time, choosing the least recently attempted so a busy channel cannot crowd out other channels or membership notifications. The five-second delay starts when an attempt finishes, including a failed write; failed writes leave recovery pending. Recovery is paced by available capacity, not by how often messages are lost. This is not a larger buffer or a cutoff that abandons recovery. Subscription identifiers, message filters, replay timestamp overlap and duplicate filtering are unchanged; no downstream agent changes are required. This targets a reproducible overload **amplifier**, not every cause of overload or every catch-up limitation. The initial live overload's cause has not been established. Recovery remains best effort: a successful request write is not proof of delivery, and existing history/retention limits, bounded duplicate tracking and replay limitations still apply. There is no exactly-once or complete catch-up guarantee. A stalled write can still pause socket reads for the existing ten-second timeout; the pacing bound does not cover initial subscriptions, reconnects or other retry paths. ### Related issue Closest related: #5014 (channel re-subscription); also #6661 (membership reconciliation) and #6090 (relay backpressure gap signaling). This addresses local overflow recovery scheduling, not those separate mechanisms. ### Testing Recorded offline comparisons against the previous behavior, with the final implementation at `8000636f3073167c5a5107bb179c7d91160f1729`: | Same fixture: 18 subscriptions, three overload rounds | Before | After | | --- | --- | --- | | Recovery history requests | 108 | 3 | | Ping-response delay | About 4.6 seconds | Below the measurement's 1 ms resolution | A separate bounded-history fixture delivered all 320 events plus subsequent live traffic in **both** versions. Regression coverage exercises the real socket-handling task, including intermittent consumer capacity, fairness, failed writes and cancellation of capacity waits before live delivery. These are synthetic results, not production throughput measurements or evidence of a deployed cure. The full local `RUST_TEST_THREADS=4 just ci` run passed on September 4, 2026. Earlier unsuccessful local runs remain part of the validation history. The [recorded validation evidence and separate desktop follow-up](https://github.com/block/buzz/pull/7325#issuecomment-5540592398) preserve the original desktop mock-history scroll failure, its passing rerun and the remaining investigation. That desktop path does not run the agent connection code; neither this repair nor the passing rerun fixes the observed scroll problem. --------- Signed-off-by: Logan Johnson <loganj@squareup.com> |
||
|
Salman Mohammed
|
86c189e857 |
fix(buzz-acp): wake held ACP threads and fence forked sessions (#7340)
## Summary Adds an independent deadline wakeup so held thread work dispatches after its 10-second bound even when the relay loop is otherwise quiet. Fences session ownership by generation so a worker returning after a fork cannot make an older provider session claimable again. This follows up on the two post-merge findings from [#7337](https://github.com/block/buzz/pull/7337#pullrequestreview-5116329341). ### Related issue Follow-up to #7337. ### Testing - `cargo test -p buzz-acp` - `cargo clippy -p buzz-acp --all-targets -- -D warnings` - Pre-push file-size, differential Rust test, and desktop Tauri gates No UI changes. --- **Update Sep 4, 15:35:** Addressed both Codex review findings. - Queue-cap eviction now prunes orphaned hold deadlines. - An expired hold stays expired until a worker is successfully claimed. - Hold timers remain disabled while every worker is busy; worker return wakes dispatch directly. - Added regressions for queue eviction and pool exhaustion. Generated with Codex --------- Signed-off-by: Salman Mohammed <smohammed@squareup.com> |
||
|
Salman Mohammed
|
4beffef697 |
feat(buzz-acp): update base prompt; add buzz context and skills to Pi agents (#7335)
## Summary - launch Pi through a private Buzz wrapper so managed Pi agents receive Buzz's base prompt plus the configured thread/channel session model via `--system-prompt` - load the Buzz workspace skill directory with Pi's repeatable `--skill` flag, making `~/.buzz/.agents/skills/buzz-cli` available alongside Pi's normal global skill discovery - report Pi and `pi-acp` installation states separately so setup guidance points to the missing component `pi-acp` does not currently consume the ACP `session/new` system prompt, but its normal new-session and restore paths do honor `PI_ACP_PI_COMMAND` as the executable used to launch Pi. Buzz reserves that variable and sets it to a private generated launcher. The launcher invokes `pi` from Buzz's effective `PATH`, adds `--system-prompt <file>` and `--skill <workspace>/.agents/skills`, and forwards `pi-acp`'s RPC and session arguments unchanged. This keeps the integration entirely in `block/buzz`, without changes to Pi or `pi-acp`. ### How PI_ACP_PI_COMMAND works in `pi-acp` `PI_ACP_PI_COMMAND` selects one executable; it does not accept arguments. Therefore: ```bash # Does not work PI_ACP_PI_COMMAND="pi --skill ~/.buzz/.agents/skills" ``` Buzz does not expose this variable as user configuration. It creates a private launcher and sets `PI_ACP_PI_COMMAND` only on the `pi-acp` child. Conceptually, that launcher executes: ```sh #!/bin/sh exec pi \ --system-prompt "$PRIVATE_SYSTEM_PROMPT" \ --skill "$HOME/.buzz/.agents/skills" \ "$@" ``` ### Related issue Follow-up to #7208. No duplicate issue or open PR found. ### Testing - `just ci` - `cargo test -p buzz-acp pi_launcher` - `cargo test --manifest-path desktop/src-tauri/Cargo.toml managed_agents::env_vars::tests` - `cargo test --manifest-path desktop/src-tauri/Cargo.toml managed_agents::discovery::presets::tests` - pre-push branch checks No screenshots: the UI change is state-dependent setup guidance only. --- **Update Sep 4, 13:35:** Made `PI_ACP_PI_COMMAND` entirely Buzz-owned. - Managed agent configuration now rejects the variable as a user override. - Buzz always launches `pi` from its effective `PATH` and rejects inherited values before creating its private launcher. - Removed the custom Pi executable discovery path and its internal environment alias. --- **Update Sep 4, 14:03:** Reject inherited `PI_ACP_PI_COMMAND` values instead of replacing them. - Pi startup now fails with an actionable message telling the user to unset the variable. - Removed the Pi-specific exception from the generic ACP environment injection path. Generated with Codex --------- Signed-off-by: Salman Mohammed <smohammed@squareup.com> |
||
|
Jordan Mecom
|
e09f715c9d |
Verify ACP relay events before prompt routing (#7010)
## What changed Verify every Nostr EVENT received by buzz-acp immediately after deserialization. Events with an invalid NIP-01 ID or Schnorr signature are dropped before subscription routing, deduplication, replay-watermark updates, membership handling, or either harness queue. ## Safety Signature verification runs on the blocking pool so cryptographic work does not block the relay task. A verification failure drops only that event and keeps the connection available for subsequent valid traffic. The existing observer-control verification remains as defense in depth. Regression tests cover valid delivery; changed content, ID, signature, pubkey, tags, and timestamp; a forged owner shutdown command with a recomputed ID; forged membership notifications; and forged observer-control events. ## Testing cargo test -p buzz-acp cargo clippy -p buzz-acp --all-targets -- -D warnings just ci --------- Signed-off-by: Jordan Mecom <jm@squareup.com> |
||
|
Will Pfleger
|
b17c0776b7 |
fix(buzz-acp): bound busy-owner hold to prevent cross-channel starvation (#7337)
## Problem #6732 added a busy-owner hold to the ACP harness: when a scope's recorded session owner (`session_owners`) is checked out on **any** turn, `dispatch_pending` holds the scope's batch instead of dispatching it. The hold was added to keep one provider session per thread — but it is unconditional: it applies to `Conversation` scopes too, and it has no time bound. Under the default `session_policy=channel`, every channel collapses to a single `Conversation` scope, so once two channels' sessions land on the same worker (pass 2 of `try_claim` picks the first idle worker by index, so this happens quickly after any restart), channel A's mention starves behind channel B's in-flight turn — for up to the full `max_turn_duration` (7200s by default) — while other workers sit idle. The only signal is a DEBUG-level log, and the 👀 seen-reaction is added at queue admission *before* the hold decision, so the user sees the agent acknowledge the mention and then nothing. Observed in production on the first day of the v0.5.22 rollout: three separate incidents where a mention got 👀 but no turn started until an unrelated channel's turn ended on the shared worker (in the worst case the blocking turn sat in a single tool call for 6+ minutes). ## Fix One new seam, `AgentPool::hold_decision`, replaces the raw `should_hold_for_busy_owner` check in `dispatch_pending` (the predicate itself is unchanged and remains the inner check): - **`Conversation` scopes never hold.** Channel-policy channels and all DMs dispatch immediately; a busy owner means forking onto an idle worker, exactly the pre-#6732 behavior. This removes the cross-channel head-of-line blocking entirely for the default policy. - **`Thread` scopes hold for a bounded window.** `HOLD_BUSY_OWNER_TIMEOUT` (10s) is measured from the first time the batch is held (`held_since` stamp); once elapsed, the batch stops holding and forks a fresh session on an idle worker, rebuilding thread context from the relay. This preserves #6732's session-continuity intent for the momentary-busy case while capping the worst-case wait. No new timer is needed: held batches are requeued with preserved timestamps and re-evaluated on every dispatch trigger (turn end, relay event, 30s maintenance tick), so the effective worst-case re-check gap on a fully silent system is one maintenance tick. - **Holds are observable.** Holding logs at INFO and a hold expiry logs at WARN (previously DEBUG-only), and both emit observer-feed events (`busy_owner_hold`, `busy_owner_hold_forked`) with the scope, owner index, and held duration. `held_since` is derived state and is cleared on every removal path: dispatch/fork (inside `hold_decision`), `invalidate_channel_sessions`, `invalidate_scope_session`, and `switch_idle_agent_model`. ## Accepted trade-offs - A fork after an expired hold leaves the old owner's now-orphaned thread session in its session map until natural rotation/invalidation — benign, and identical to pre-#6732 fork semantics (`loadSession: false`; sessions are worker-pinned, so migration is not an option). - Under sustained pool exhaustion the hold stamp is cleared on the fork attempt and re-stamped next cycle, so the bound is effectively "timeout after a worker frees up," not absolute wall clock. ## Tests - New table test `hold_decision_covers_variant_session_busy_and_timeout` over the full input space (scope variant × idle-session presence × owner busyness × elapsed vs. window). The `Conversation` + busy-owner row is the cross-channel regression guard; the past-window row guards the bound. Both were mutation-checked: removing the variant gate or the timeout branch fails the suite. - `busy_session_owner_holds_batch_instead_of_forking_session` extended with the Hold → ForkAfterHold transition, the `Conversation` dispatch guard, and `held_since` pruning on channel invalidation. - Scope-invalidation and idle-model-switch tests extended to cover `held_since` cleanup alongside the existing `session_owners` assertions. Signed-off-by: Will Pfleger <pfleger.will@gmail.com> |
||
|
Salman Mohammed
|
ce9decb235 |
fix(acp): rename system tag to agent-instructions (#7332)
## Summary Rename the paired persona prompt boundary from `<system>` to `<agent-instructions>` in both modern and legacy ACP delivery paths. Update Desktop diagnostics to label the new boundary as Agent Instructions while keeping archived `<system>` captures readable. `system` is confusing (esp with system role) and agent instructions is what shows up in the UI today <img width="767" height="298" alt="Screenshot 2026-09-04 at 11 37 54 AM" src="https://github.com/user-attachments/assets/4e34ac06-db50-461e-a555-ce17885b0031" /> ### Related issue Follow-up to #6701. No duplicate open issue or PR found. ### Testing - `cargo test -p buzz-acp` - `cargo clippy -p buzz-acp --all-targets -- -D warnings` - `node --import ./desktop/test-loader.mjs --experimental-strip-types --test desktop/src/features/agents/ui/agentSessionTranscriptHelpers.test.mjs desktop/src/features/agents/ui/agentSessionTranscript.test.mjs` - `pnpm --dir desktop typecheck` - Pre-push hook: 6,237 Desktop tests, 3,164 Tauri tests, and all 13 Rust unit-test lanes passed Generated with Codex Signed-off-by: Salman Mohammed <smohammed@squareup.com> |
||
|
Matt Toohey
|
40220d5618 |
perf(desktop): publish mention sends before waking agents (#7154)
## Summary Cuts perceived agent-mention send latency by publishing the message first and waking the agent afterwards, instead of blocking the send on a synchronous agent start/deploy round-trip. A send that mentions a stopped or undeployed managed agent now shows the message immediately; the wake runs fire-and-forget after the relay accepts the publish. The already-running-agent send also gets faster via revalidation dedupe and NIP-11 caching. ## Changes ### Publish-first agent wake - Wakes for mentioned managed agents are collected during send preparation and flushed fire-and-forget only after `await send(...)` resolves. No start can fire — and no "your message was sent" toast can appear — for a message the relay never accepted; every abort path (cancel, readiness error, publish rejection, dismissed non-member prompt) simply drops the queue. Persona-create wakes ride the pending draft behind the non-member prompt for the same reason. - Each wake is bound to the tenant scope captured at send time: the new `useDetachedAgentStart` hook passes `expectedRelayUrl` + `expectedSignerPubkey` with every start, so a wake that outlives a community switch fails closed at the backend instead of spawning against the new tenant. A wake whose scope has not resolved yet (identity query still loading, blank stored relay URL) is refused with a recoverable toast rather than fired unscoped. - In-flight wakes are deduped through a module-level map keyed by `(relay URL, pubkey)` — the same tenant pair the backend keys on — so two quick sends or two composers cannot double-spawn a cold agent during the seconds-long start window. Entries are deliberately retained across community switches (the key *is* the tenant scope, so a retained entry can never affect another community, and clearing it let an A→B→A round trip deploy a provider agent twice) and self-clean when the start settles. - Wake-failure toasts are fenced to the community they fired in via a module-level scope mirror: a start that settles after a community switch logs instead of rendering community A's failure over community B's UI, and an A→B→A return re-delivers the warning where it is actionable. - Membership attach and access-policy writes stay synchronous, so the harness's first kind-39002 read still sees the channel. ### Replay floor - The send timestamp travels with the wake as `BUZZ_ACP_REPLAY_FLOOR`, threaded through both local spawns (`spawn_agent_child`) and provider deploys (`deploy_to_provider` injects it into `launch.policy_env`), so the harness's startup watermark replays back past the just-published triggering message no matter how long the spawn takes. `buzz-acp` clamps the floor to `[now − 15min, now]`. - The floor is captured at enqueue time, not flush time — the flush runs post-publish, so a flush-time stamp could exceed the message's `created_at` and skip the very message the floor exists to cover. - On local spawns the caller's floor is asserted *after* the user env layering (and the ambient parent-process value is stripped unconditionally), so a saved persona/global/agent env entry cannot shadow this send's floor — mirroring the shadow-strip the provider path applies to `launch.env`. Both halves share one `REPLAY_FLOOR_ENV_VAR` const. ### Send-path latency reductions (already-running agents) - Mention revalidation is deduped: the publish-boundary pass reuses the pre-side-effect authorization pass unless an awaited round-trip actually separated the two (background upload, link-preview settlement, DM expansion, a real access-policy/membership write, or active-huddle enrollment). This preserves the #5681 authorization boundary while making the common send single-pass. - NIP-11 `self` lookups are cached per relay URL for 5 minutes. Only verified values are cached — non-2xx and malformed responses stay retryable — and URL keying keeps community switches from serving another relay's identity. - `applyReusableAgentAccessPolicy` now reports its relay write explicitly (`{ agent, wrote }`) instead of signalling through object identity, so the revalidation trigger above is load-bearing by construction. ### File splits Four files crossed the repository file-size ratchet during this work; one cohesive unit was extracted from each rather than raising a ceiling — `runtime/setup_payload.rs`, `commands/agents_create_fields.rs`, `app_state_accessors.rs`, and `useEnsureAgentMentionsReady.ts`. The ratchet is green at the tip. ### Review follow-ups The three concrete findings from the first review round are fixed at the tip: the pre-publish wake and its false "your message was sent" toast (fixed by queueing wakes behind the publish), the stale cross-community failure toast (fixed by the scope-mirror fence), and the A→B→A duplicate provider deploy (fixed by retaining the tenant-keyed in-flight entries across switches). The fast-path admission-staleness point is answered in the review thread: deferred paths already re-validate at the publish boundary, and the remaining fast-path window is milliseconds against an irreducible network-transit race. Mid-branch send-perf instrumentation was added to attribute the residual spinner latency and reverted once that analysis concluded — it is net-zero in this diff. ### Deferred follow-ups Durable mention catch-up via `event_mentions` (option 2 step 3) and backend deploy-epoch coalescing for the wake paths that do not funnel through `useDetachedAgentStart` (Agents-panel Start, restore, inbound-persona deploys) are intentionally left for separate changes. ## Testing - `cargo test --lib` on desktop/src-tauri: 3054 passed; clippy `-D warnings` + fmt clean - Desktop unit tests: 5856 passed (the 5 failures are the pre-existing `inboxReopenNavigation` / `useRetainedProjectGitViews` baseline, present on origin/main); `tsc --noEmit` and biome clean - Full mentions (87), channels (89), and community-rail (25) Playwright smoke suites against `pnpm build:e2e` bundles, with 3× stress reruns of each new spec - The load-bearing regression specs were confirmed red on the pre-fix code: publish-failure → zero starts and no false toast, the dedupe hold (1 call vs 2), the fail-closed scope refusal, the rail-switch toast fence, and the A→B→A retention spec (1 deploy vs 2) - New unit coverage pins the queue contract (enqueue-time floors, attach-seam queueing), the scope capture and verbatim relay-URL handoff, the dedupe map's keying and settle-then-repermit behavior, the unscoped refusal, the toast-scope mirror, the `{ agent, wrote }` contract, and the replay-floor env layering on both spawn paths 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Signed-off-by: Matt Toohey <contact@matttoohey.com> Co-authored-by: Claude Fable 5 <noreply@anthropic.com> |
||
|
Will Pfleger
|
47d068e210 |
feat(cli): add buzz gifs command group and NIP-30 emoji tags on messages (#7259)
## What Two new agent-facing capabilities in `buzz-cli`: ### 1. `buzz gifs` command group (agent KLIPY picker path) Agents can now search and share GIFs via the relay's authenticated KLIPY proxy without holding a provider credential. ```bash buzz gifs search # trending GIFs buzz gifs search --query "celebration" # search GIFs buzz gifs share --slug <slug> # report selection to provider Recents ``` Output is a JSON array of GIF objects. Paste the `cdn_url` field directly into `buzz messages send --content` — sending a GIF is a plain message containing the CDN URL, no special send-path handling. **Implementation details:** - Gates on NIP-11 `supported_extensions` containing `buzz-gif` and `gif.provider == "klipy"` - Uses relay-relative paths from the NIP-11 `gif` descriptor — no hardcoded paths; safe-path validation mirrors `desktop/src/features/gifs/api.ts` - New `post_json_authed` helper in `BuzzClient` handles NIP-98-signed JSON POSTs and 204 No Content responses - `customer_id` derived as `SHA-256(secret_key_bytes || '\0' || relay_url_bytes)[..16]` → 32 hex chars: stable, relay-scoped, not computable from public data, no storage needed - `locale` defaults to `$LANG` (stripped of encoding suffix) or `en_US` ### 2. NIP-30 custom emoji tags on outgoing messages `buzz messages send` now automatically attaches `["emoji", shortcode, url]` tags for any `:shortcode:` patterns in the content that resolve in the workspace palette — identical to the desktop composer behavior. ```bash buzz messages send --channel <uuid> --content "hello 👋 everyone 🎉" # → event carries ["emoji", "wave", "..."] and ["emoji", "tada", "..."] tags ``` **Implementation details:** - Hand-rolled single-pass scanner (no new dependency) implementing `:([a-z0-9_-]+):` case-insensitively with canonical lowercase output — mirrors `desktop/src/shared/lib/customEmojiTags.ts` exactly - Zero extra relay round-trips when content contains no `:` character; one `query` when candidates exist but none match - Palette fetch reuses the existing `union_custom_emoji` logic from `commands/emoji.rs` - `build_message` in `buzz-sdk` gains a new `emoji_tags: &[Vec<String>]` parameter (additive — all existing callers pass `&[]`); NIP-30 tag attachment lives in the SDK alongside `imeta` tags - MCP send path (`buzz-acp`) continues to pass `&[]` and is not affected; the MCP gap is noted in a comment ## Files changed | Crate | File | Change | |-------|------|--------| | `buzz-cli` | `src/commands/gifs.rs` | New — search + share handlers, NIP-11 gating, tests | | `buzz-cli` | `src/commands/mod.rs` | `pub mod gifs` | | `buzz-cli` | `src/lib.rs` | `Gifs(GifsCmd)` variant, dispatch arm, inventory test update | | `buzz-cli` | `src/client.rs` | `post_json_authed` helper | | `buzz-cli` | `src/commands/emoji.rs` | `scan_shortcodes` + `resolve_emoji_tags_for_content` + tests | | `buzz-cli` | `src/commands/messages.rs` | Emoji scan + tag injection in `cmd_send_message` + seam tests | | `buzz-cli` | `README.md` | `buzz gifs` section + emoji-in-messages note | | `buzz-sdk` | `src/builders.rs` | `build_message` gains `emoji_tags` param + tests | | `buzz-acp` | `src/pool.rs` | Update `build_message` call site (`&[]`) | | `buzz-acp` | `src/setup_mode.rs` | Update `build_message` call site (`&[]`) | | `countdown-bot` | `src/main.rs` | Update `build_message` call site (`&[]`) | Relates to: https://buzz.block.builderlab.xyz — buzz-team channel thread on agent GIF/emoji support --------- Signed-off-by: Will Pfleger <pfleger.will@gmail.com> Co-authored-by: Duncan <dcfd242e557282d7a1e2cf2e6877522682f1e5c6156dc92ca7d90eaedd3b0f95@buzz.block.builderlab.xyz> |
||
|
Will Pfleger
|
2af9773d62 |
fix(acp): replace real user name in base prompt mention example (#7250)
Replace the real user name in the shared ACP mention guidance with the fictional `Alice Smith` example. Preserve the exact-display-name and no-inference instructions while avoiding prompt priming from a real user identity. Signed-off-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Alia <d32955ad69077062930cc46cfe2df30ca9aaf6f8e76422681265e9e9af704d78@buzz.block.builderlab.xyz> Co-authored-by: Alia <d32955ad69077062930cc46cfe2df30ca9aaf6f8e76422681265e9e9af704d78@buzz.block.builderlab.xyz> |
||
|
Salman Mohammed
|
42aeb1571c |
feat(desktop): add Pi agent preset (#7208)
## Summary Add Pi as a selectable desktop ACP runtime via `pi-acp`, including discovery guidance and its official logo. ### Related issue Follow-up to closed #6546. ### Testing - Focused Tauri preset and desktop logo tests - TypeScript typecheck Generated with Codex --------- Signed-off-by: Salman Mohammed <smohammed@squareup.com> |
||
|
Will Pfleger
|
cae158ce7f |
fix(dev-mcp): extend shell timeout cap to 20 minutes and align outer budgets (#7185)
Raise the `buzz-dev-mcp` shell timeout ceiling to 1,200,000 ms (20 minutes) while keeping the omitted-value default at 120,000 ms (2 minutes), and align the two outer timeout layers that bound every MCP tool call. ## What changed **`crates/buzz-dev-mcp`** — shell timeout cap raised to 1,200,000 ms; tool and parameter descriptions updated to match. **`crates/buzz-agent`** — `BUZZ_AGENT_TOOL_TIMEOUT_SECS` default raised from 660 s to 1,260 s (20-min cap + 60 s cleanup headroom). The old 660 s default would have killed the MCP server at 11 min, before a 20-min shell call could finish. README updated. **`crates/buzz-acp`** — `DEFAULT_IDLE_TIMEOUT_SECS` raised from 900 s to 1,500 s and its doc comment updated (1,200 s max shell + 300 s breathing room). The old value was documented as "600 s max shell + 300 s" — stale arithmetic. ## Tests - Existing boundary tests in `buzz-dev-mcp` (default, exact-cap, over-cap, `u64::MAX`) all pass. - `default_idle_timeout_is_1500_seconds` — pinned constant replaces old 900 s assertion. - `default_tool_timeout_is_1260_seconds` — new pinned constant in `buzz-agent`. - `budget_ordering_invariant_shell_cap_plus_headroom_fits_within_idle_timeout` — new const assertion in `buzz-acp` that encodes the full three-layer ordering: shell cap (1,200 s) ≤ agent tool timeout (1,260 s) < ACP idle timeout (1,500 s) < max turn duration. A future edit that inverts any tier fails to compile. --------- Signed-off-by: Alia <d32955ad69077062930cc46cfe2df30ca9aaf6f8e76422681265e9e9af704d78@buzz.block.builderlab.xyz> Signed-off-by: Will Pfleger <pfleger.will@gmail.com> Co-authored-by: Alia <d32955ad69077062930cc46cfe2df30ca9aaf6f8e76422681265e9e9af704d78@buzz.block.builderlab.xyz> Co-authored-by: Hayt <9e1c23a3fd83f61da34420e4e88ff1b16e45cafcc0cd9019eb07d4ecfa8ca9b0@buzz.block.builderlab.xyz> |
||
|
Salman Mohammed
|
674c173eb7 |
feat(buzz-acp): give each channel thread its own agent session (#6732)
## What this does In a channel, people often run several unrelated conversations at once (separate threads). Today the agent treats the whole channel as one conversation, so unrelated threads share the same running session — their context bleeds together and independent tasks can step on each other. This change gives the agent a **separate session per thread** inside a channel. Direct messages stay as one conversation (unchanged). The channel is still the boundary for who is allowed in and what is visible — only the agent's working context is now split by thread. ## How it is turned on Off by default. Operators opt in with one setting: - `BUZZ_ACP_SESSION_POLICY=channel` — default, current behavior - `BUZZ_ACP_SESSION_POLICY=thread` — new per-thread behavior Being behind a flag means we can enable it for a few agents, watch how it behaves, and roll back instantly without a code change. ## Key design decisions - **Decide the thread once, up front.** When a message arrives we work out which thread it belongs to a single time and tag it. Everything after that (which line it waits in, which session runs it, what history it sees) uses that tag instead of re-guessing later, which avoids mismatches. - **Default stays identical to today.** Under the default setting a "thread" is just "the whole channel," so existing behavior and every existing test are unchanged. The new, riskier behavior is strictly opt-in. - **Give the agent only its thread's history.** On a reply the agent sees that thread's messages (including ones that did not mention it), not the whole channel transcript — less noise and smaller prompts. - **Don't let one channel use more memory than before.** More threads means more live sessions, so the existing per-channel limit now caps all of a channel's threads together — splitting into threads can't multiply how much work is held. ## Bugs found and fixed while iterating (from review) - **Same thread, two sessions.** If the worker already holding a thread's session was busy, a new message for that thread could start a *second* session on another worker and split its history. Now it waits for the right worker instead of forking. - **Interrupting the wrong thread.** A follow-up meant for thread A could interrupt thread B in the same channel. Interrupts now target the exact thread. - **Stuck thread after a crash.** If a thread's turn crashed, its slot wasn't cleared and stayed blocked for up to ~2 hours. It now clears right away and retries. - **Lost the original request.** When a thread was interrupted and then had to wait for a busy worker, only the follow-up was kept and the original request was dropped. The full request is now preserved on retry. - **Same thread seen as two.** Two spellings of the same thread id (upper/lower case) could be treated as different threads. Normalized so they count as one. ## Not in this PR - The desktop Settings toggle and rollout wiring for managed agents — https://github.com/block/buzz/pull/6909 - One pre-existing retry edge case (present today without this flag, unrelated to this change) — tracked separately so this PR stays focused. ## Testing The full `buzz-acp` test suite passes (830+ unit and integration tests), plus new focused tests for thread routing, session reuse, interrupt targeting, crash recovery, and request preservation. Behavior with the flag off is unchanged. --------- Signed-off-by: Salman Mohammed <smohammed@squareup.com> Signed-off-by: Leo <5faf251baee50ee6bcde338aef6acdd70bb3e60115664c2cd490d94a55dfc488@buzz.block.builderlab.xyz> Co-authored-by: Leo <5faf251baee50ee6bcde338aef6acdd70bb3e60115664c2cd490d94a55dfc488@buzz.block.builderlab.xyz> |
||
|
Wes
|
2f3dd850db |
fix(relay): reject a frame on its own acknowledgement channel (#6961)
Pinky, an AI agent, updated this description on Wes's behalf after
taking over the startup investigation.
**Category:** fix
**User Impact:** An EVENT refused by WebSocket admission or handler
saturation receives a correlated `OK(event_id, false, reason)` instead
of an uncorrelated NOTICE, so the client can settle that refusal without
waiting for its publish timeout. Rate-limited refusals also arm client
backoff. This fixes a protocol failure mechanism; it does not establish
that every startup send will succeed or that the reported Desktop
startup incident is fully resolved.
**Problem:** Startup opens several live subscriptions and publishes at
once, and the relay's WebSocket admission gate is a fixed 5-second
window (`ws_admission_budget` = `human_ws_events_per_sec * 5`). If that
shared per-principal quota is exhausted, `enforce_ws_admission`
previously rejected an EVENT with a bare `["NOTICE", reason]`. Quota
pressure is a possible trigger, not proof of the original incident's
complete cause.
A NOTICE carries no event id. Both clients settle a pending publish
*only* from an `OK` keyed by event id (desktop `pendingEvents`, mobile
`_pendingEvents`), so nothing settled — and `handle_text_message`
returns early, so no `OK` ever followed either. The send **could not
fail**; it could only time out at `PUBLISH_TIMEOUT_MS` = 25s. That
explains how this rejection mechanism can produce a roughly 25-second
timeout; attributing the original report to it still requires the actual
startup/send workflow.
The handler-semaphore saturation path had the identical defect, and that
one needs no quota burst to fire.
**Solution:** NIP-01 gives each request type its own acknowledgement
channel, and a rejection is only actionable on the same one. Reject a
REQ with `CLOSED`, an EVENT with `OK(id, false, reason)`, and fall back
to `NOTICE` only where no per-request correlation exists. COUNT refusals
now also use `CLOSED(query_id, reason)` per NIP-45, covering both quota
admission and handler saturation (added in
`cd12c93804b87a24b61075dfd171dc471a0a527f`).
Reason strings are unchanged, so the `rate-limited:` prefix and `retry
in {N}s` hint that existing client gates parse keep working (desktop
`parseRateLimitHint`, mobile `RelayRateLimitGate`, buzz-acp
`set_rate_limit_gate`). Only the frame *type* changes, so
`docs/multi-tenant-relay.md` L7 stays satisfied.
Two notes on how this landed, both worth a reviewer's attention:
1. **A survived mutation became a design change.**
`send_admission_result` originally took a `RejectionTarget` parameter,
and reverting the *second* call site (the per-minute message quota)
survived the whole suite — with Redis unreachable the first quota check
short-circuits, so that line is unreachable in test. Rather than test
around it, the parameter is gone: the target is derived from the frame,
so no call site can name the wrong channel.
2. **The relay fix would have caused a client regression on its own.**
Gate arming lived only in the NOTICE branch. Once rejections arrive as
`OK:false`, `handleOk` failed the send without ever backing off — the
client would retry straight into the same quota. Desktop and Mobile now
arm on a `rate-limited:` OK rejection. ACP was subsequently fixed in
`3b06dd32493596ec650f20abf8805791c50fdc24`: it arms the gate and
re-parks only the refused observer frame, preserving other in-flight
frames. Desktop gets `activateRateLimitIfSignalled` as the single owner
of that prefix test, called from both `handleOk` and the NOTICE branch.
<details>
<summary>File changes</summary>
**crates/buzz-relay/src/rejection.rs** (new)
Owns the admission-rejection concern: `RejectionTarget`,
`rejection_target_for`, `request_rejection_message`,
`send_admission_result`, and `enforce_ws_admission`, moved out of
`connection.rs`. Six tests, two of which drive the real
`enforce_ws_admission` against a real `AppState`.
**crates/buzz-relay/src/connection.rs**
Fix the EVENT handler-semaphore rejection to correlate to the event id;
delegate admission to the new module. Add two tests that drive the real
`handle_text_message` with every handler permit held. Down from 1319 to
1116 lines.
**crates/buzz-relay/src/state.rs**
Widen the existing `test_state` helper to `pub(crate)` so the rejection
tests reuse it rather than adding a ninth copy of `AppState`
construction.
**desktop/src/shared/api/relayRateLimitGate.ts**
Add `activateRateLimitIfSignalled` — one owner for the `rate-limited:`
prefix test, since three inbound frame types now carry it.
**desktop/src/shared/api/relayClientSession.ts**
Arm the gate on a rate-limited OK rejection; route the NOTICE branch
through the same helper. Net zero lines, which keeps this
already-oversized file within the differential ratchet.
**desktop/src/shared/api/relayClientPublishRejection.test.mjs** (new)
Four tests against the real `RelayClient`: a rate-limited OK settles the
pending publish and arms the gate; an ordinary rejection does not arm
it; an accepted OK still resolves.
**mobile/lib/shared/relay/relay_session.dart**
Arm the gate in `_handleOk` for a rate-limited rejection.
**mobile/test/shared/relay/relay_session_test.dart**
Two tests driving the real `publish` + `debugHandleMessage` path.
</details>
<details>
<summary>Validation</summary>
**Mutation-tested — 5 mutations, all now killed.** Each production call
site was reverted to the defective behaviour to confirm a test fails.
This caught two false-negative tests:
| # | Mutation | Result |
|---|----------|--------|
| 1 | `rejection_target_for`: EVENT → `Connection` | 4 tests fail |
| 2 | EVENT handler-semaphore call site → bare NOTICE | **survived at
first** |
| 3 | per-minute quota call site → `Connection` | **survived**; fixed by
removing the parameter |
| 4 | desktop `handleOk` gate arming removed | 1 test fails |
| 5 | mobile `_handleOk` gate arming removed | 1 test fails |
Mutation 2 is the lesson: my first saturation test called
`request_rejection_message` directly, so reverting the real call site
inside the `match` arm left it green. It now drives
`handle_text_message` itself and dies on that mutation.
- `cargo test -p buzz-relay` — 928 passed, 1 failed:
`api::mesh_demo::tests::demo_join_forwarded_arm_round_trips_echo`,
**pre-existing**, reproduced with all changes stashed at `4dd4d73de`.
- `cd desktop && npm test` — 5721 passed, 0 failed (full suite).
- `cd mobile && flutter test` — 1876 passed, 0 failed (full suite).
- `just fmt-check`, `just clippy`, `just desktop-check`, `just
mobile-check`, `just file-size-check` — clean. Desktop's 5 biome
warnings are pre-existing (reproduced with changes stashed).
- All 9 pre-push lanes green, including `rust-tests` and
`desktop-tauri-checks`.
**Not verified:** not reproduced end-to-end against a live relay under a
forced quota burst. The causal chain is source-proven and
mutation-proven at the frame level; the ~25s attribution follows from
`PUBLISH_TIMEOUT_MS` but is not directly measured. A packaged-build
click-through would close that gap.
</details>
Related work: #6957 bounds Desktop HTTP event submission, but safe
retained-operation recovery after exhausted/ambiguous outcomes remains
unfinished. #6998 is the separately reviewable Desktop
readiness/duplicate-subscription slice. Neither is claimed to complete
native before/after startup-send validation.
Diagnosis note: `RESEARCH/DESKTOP_STARTUP_SEND_STALL_2026_08_27.md`
(Brain's workspace).
## Current review disposition (2026-08-28)
The [review on
`cd12c938`](https://github.com/block/buzz/pull/6961#pullrequestreview-5052902510)
identified ACP's missing rate-limited-OK handling. Commit
`3b06dd32493596ec650f20abf8805791c50fdc24` fixes gate arming, re-parking
the specifically refused observer frame, and the stale NOTICE comment.
Two regressions drive the real frame dispatcher. See [the implementation
and validation
response](https://github.com/block/buzz/pull/6961#issuecomment-5455032054).
The Mobile generation-check inline thread is resolved: its `async
publish` returns a failed Future when superseded; it does not throw
synchronously at invocation. No further production change was indicated
by that comment.
The validation counts above describe the original slice, not a new
rerun. At `3b06dd324`, the current GitHub check rollup has successful
completed test/build checks (non-applicable jobs skipped). The
security-review comment still requires review for the current base/head
range; do not read a green authorization job as a completed security
review. Approval and merge remain human decisions.
---------
Signed-off-by: Brain <1a02c72794dcd0f07058a353bc3a81f4028b8c77c92c87fce6d5c8b85970a20b@buzz.block.builderlab.xyz>
Signed-off-by: Wes <wesbillman@users.noreply.github.com>
Co-authored-by: Brain <1a02c72794dcd0f07058a353bc3a81f4028b8c77c92c87fce6d5c8b85970a20b@buzz.block.builderlab.xyz>
Co-authored-by: Carl <9d00794d3df50972eb8b615511783cab12a77a8fd5dd5edd58073ec73b54bd8b@buzz.block.builderlab.xyz>
|
||
|
Wes
|
93237b4a74 |
fix(acp): wake agents from workflow messages (#6953)
> Pinky, an AI agent, is opening this PR on Wes's behalf.
## Summary
Workflow-generated messages can contain a valid agent mention but still
fail the ACP inbound author gate because the relay signs the event. This
keeps the existing wake policy and gives ACP a narrowly verified
effective author:
- preserve the workflow owner's existing `p` tag and all
rendered-mention `p` tags
- add explicit `["buzz:workflow-owner", <owner hex>]` provenance to
relay-generated workflow messages
- add `["buzz:workflow-mention", <agent hex>]` authority only for
mentions resolved from the stored, unrendered workflow step template
- accept that owner only for a verified kind-9 event signed by the
relay's current NIP-11 `self` key, with unique canonical workflow
metadata and an explicit workflow mention for the receiving agent
- route the verified owner through the existing author and in-flight
mode policies in both normal and setup listeners
- refresh relay identity after reconnects, retaining the last verified
key on transient fetch errors while treating a successful response
without `self` as definitive removal
Malformed, duplicate, forged, tampered, wrong-kind, and wrong-relay
attribution all fail closed to the raw event signer. `respond-to=nobody`
remains absolute. Old/mixed-version messages without the explicit
provenance retain their current fail-closed behavior.
## Trust boundary
The workflow owner means **“scheduled by,” not “authored every rendered
word.”** Trigger-controlled substitutions may still produce ordinary `p`
mention routing for compatibility, but they cannot mint
`buzz:workflow-mention` authority. Only a target named in the durable
owner-authored step template can receive that authority.
The author gate is not bypassed: after relay signature/provenance
verification, the effective owner is evaluated under the same
`owner-only`, `allowlist`, DM, and `nobody` policies used for ordinary
messages. Owner control commands continue to use the raw event signer.
## Why this PR
This is the focused immediate fix for waking an **online** agent from a
stored workflow mention. Earlier attempts were not a finished mergeable
fix and had materially different or incomplete trust designs. Larry's
larger draft stack addresses durable delivery across restarts; that
remains valuable future work and can supersede this effective-author
path when it lands.
## Validation
At exact clean commit `fe5b55619fe44176343eefb4cb7fe180df45a7d8`:
- `buzz-relay workflow_sink`: 25/25 passed, including all four ignored
PostgreSQL cases
- `buzz-acp --lib`: 845/845 passed
- `buzz-workflow --lib`: 169/169 passed (2 unrelated PostgreSQL tests
ignored)
- warnings-denied Clippy passed for the changed Rust packages
- `cargo fmt --all -- --check` passed
- `git diff --check` passed
- repository pre-push gates passed, including branch-scoped Rust tests
- CI now selects the ACP library tests and the relay's pure + PostgreSQL
workflow-sink tests so these guards cannot silently remain unexecuted
The production event-to-author gate is shared by normal and setup
listeners and has biting regression tests for accepted explicit
attribution, legacy owner-`p` rejection, and forged-attribution
rejection.
## Exact-head local relay + ACP proof
Following the release-binary/local-relay shape in `TESTING.md`, the
exact commit above passed a fresh isolated real-process matrix using:
- a freshly recreated Postgres database with migrations
- isolated Redis
- exact-head release `buzz-relay`, `buzz`, `buzz-admin`, and `buzz-acp`
binaries
- newly provisioned owner, channel, and bot member through the CLI
- workflow creation and triggering through the running relay
- a deterministic ACP protocol subprocess capturing actual
`session/prompt` dispatches
- a NIP-11 `self` value verified against the running relay signer
Cases:
1. A stored explicit workflow mention woke an `owner-only` agent exactly
once.
2. A workflow message without an agent mention did not wake it.
3. A non-relay signer forging every workflow authority tag did not wake
it.
4. Trigger-controlled `{{trigger.text}}` containing `@Wake Agent`
retained ordinary `p` routing but received no authority-bearing
workflow-mention tag and did not wake the agent.
5. `respond-to=nobody` remained absolute for a valid relay-authenticated
workflow mention.
The deterministic ACP subprocess isolates and directly proves relay →
ACP authorization and prompt dispatch without depending on external
model behavior.
## Deployment and residual risk
Relay and ACP changes must be deployed together for the new wake
behavior; mixed versions fail closed. Production paired-deployment proof
remains distinct from the successful local integration run. Setup-mode
behavior has automated coverage but was not a separate case in the
five-case local matrix. Relay-key rotation is observed at ACP
startup/reconnect; transient NIP-11 errors retain the last verified key,
an intentional availability tradeoff documented in code.
---------
Signed-off-by: Pinky <5f5ab050ec58ae208332edd544ebf705221e24c1b86d82a6ca07038a7a8f6ac9@buzz.block.builderlab.xyz>
Signed-off-by: Brain <1a02c72794dcd0f07058a353bc3a81f4028b8c77c92c87fce6d5c8b85970a20b@buzz.block.builderlab.xyz>
Signed-off-by: Wes <wesbillman@users.noreply.github.com>
Co-authored-by: Pinky <5f5ab050ec58ae208332edd544ebf705221e24c1b86d82a6ca07038a7a8f6ac9@buzz.block.builderlab.xyz>
Co-authored-by: LioLionel <62820906+LioLionel@users.noreply.github.com>
Co-authored-by: Brain <1a02c72794dcd0f07058a353bc3a81f4028b8c77c92c87fce6d5c8b85970a20b@buzz.block.builderlab.xyz>
Co-authored-by: Carl <9d00794d3df50972eb8b615511783cab12a77a8fd5dd5edd58073ec73b54bd8b@buzz.block.builderlab.xyz>
|
||
|
Phil Azar
|
f463e726dd |
fix: retrieving cold memories; add regression task (#6950)
## Why Evaluating buzz agent memory retrieval by seeding a memory then asking the buzz agent a question it needs that memory. **Bug Found**: System prompt had no inclusion of retrieving cold memories and suggested looking in a mem/*.md directory that does not exist. Updated `system-prompt.md` to include memory CLI tools and usage. Eval Before System Prompt Change: 0/3 Eval After System Prompt Change: 3/3 ## What - Add a `memory-retrieval` benchmark that seeds agent memory with `buzz mem set` before asking a direct question. - Grade the observable threaded answer without inspecting tool calls or exposing the answer in channel history. - Teach agents to use `buzz mem set`, `buzz mem ls`, and `buzz mem get` for cold memory. - Add a wire-debug endpoint configuration for diagnosing ACP tool calls in local runs. - Add fixture, seeding, verifier, and prompt coverage. ## Risk Assessment Low. The runtime changes are limited to the benchmark harness. The production-facing change clarifies existing memory commands in the base prompt; it does not change memory storage, relay behavior, or authorization. ## References - Before the system-prompt changes, 0/3 attempts passed because agents never invoked the `buzz mem` CLI and instead searched a non existent filesystem - After the changes, 3/3 attempts passed. ACP wire logs confirmed that every agent ran `buzz mem ls` followed by `buzz mem get` and returned `net_gpv`. --------- Signed-off-by: Philip Azar <pazar@squareup.com> |
||
|
tulsi
|
69096c9a8d |
preserve channel description paragraph breaks (#6946)
**Category:** fix **User Impact:** Channel descriptions now keep their paragraph structure everywhere people and agents read them. **Problem:** Multi-paragraph channel descriptions were flattened in the desktop interface and in agent context, making operational instructions harder to scan and potentially changing their meaning. A running agent harness could also continue using a stale description after someone edited it. **Solution:** Preserve authored line and paragraph breaks across the desktop surfaces, and render descriptions as bounded, indented agent-context blocks with semantic delimiters escaped. Prompt-visible channel metadata now refreshes before each turn with cached fallback, while one shared desktop derivation keeps the header and channel intro consistent. <details> <summary>File changes</summary> **crates/buzz-acp/src/pool.rs** Refresh prompt-visible kind-39000 metadata with one bounded request per turn so edits reach running harnesses, while retaining cached metadata as a degradation fallback. Add regression coverage for description edits and update request-count fixtures for the refresh lifecycle. **crates/buzz-acp/src/prompt_framing.rs** Add semantic text escaping for untrusted values embedded inside paired prompt sections. **crates/buzz-acp/src/queue.rs** Render multi-paragraph descriptions as indented context blocks, preserve blank lines, retain the 500-character limit, and escape semantic delimiters. Add helper- and full-prompt tests for paragraph preservation, UTF-8-safe truncation, and adversarial input. **desktop/src/features/channels/lib/channelDescription.ts** Introduce one shared channel-detail derivation and preserve newlines between status context and authored detail text. **desktop/src/features/channels/lib/channelDescription.test.mjs** Cover shared field precedence, multiline descriptions, status-prefix separation, and fallback behavior. **desktop/src/features/channels/ui/ChannelPane.helpers.ts** Use the shared channel-detail derivation for the channel intro instead of maintaining a second field order. **desktop/src/features/channels/ui/ChannelPane.helpers.test.mjs** Verify description text wins consistently over legacy purpose text when both are present. **desktop/src/features/channels/ui/ChannelManagementSheetRows.tsx** Preserve paragraph whitespace in editable and read-only channel summaries and allow enough lines for compact multi-paragraph descriptions. **desktop/src/features/channels/ui/ChannelManagementSheetRows.test.mjs** Render both channel-summary variants and assert the multiline whitespace and four-line clamp contract. **desktop/src/features/messages/ui/ChannelIntroBlock.tsx** Display description line and paragraph breaks in the empty-channel intro. **desktop/src/features/messages/ui/ChannelIntroBlock.test.mjs** Add render-level coverage for the intro's multiline whitespace contract. </details> ## Reproduction steps 1. Run the desktop app and create an empty channel. 2. Give it a description with two paragraphs, including a blank line and a multi-line second paragraph, then save and reopen channel management. 3. Confirm the empty-channel intro preserves the blank line and subsequent line breaks. 4. Confirm the channel-management summary shows the paragraph structure for both an editor and a read-only member rather than a detached ellipsis. 5. Hover the channel title and confirm its tooltip keeps the description's line breaks. 6. Add an agent, send one prompt, edit the description, then send another prompt. Confirm the next generated agent context contains the updated description as an indented multiline block. ## Screenshot A real-app screenshot of the multiline channel intro and channel-management summary is attached in the PR comments below. Closes AIDA-1980. --------- Signed-off-by: tulsi <tulsi@block.xyz> |
||
|
Salman Mohammed
|
8f3566ad9e |
Deduplicate ACP thread prompt context (#6706)
## Why This is a small isolated follow-up to [#6503](https://github.com/block/buzz/pull/6503): remove two proven duplicate prompt fragments without adopting its broader framing experiment or changing Buzz-native routing and identity behavior. ## What - Remove parsed `parent=<id>` only when it exactly duplicates parsed `root=<id>`; retain `root=<id>`, nested parents where parent differs from root, raw event tags, event IDs, author identity, mention fields and rules, and the explicit `--reply-to` instruction. - Remove the full-history retrieval hint only when complete, untruncated conversation or thread context is included. Thread context must include its root and cover every event rendered in the batch; mixed-thread batches and root-missing windows retain retrieval guidance. Context that is absent, previously delivered, or truncated also retains retrieval guidance, and the context header, count, and messages remain unchanged. ## Risk Assessment Low — the change is confined to Buzz ACP prompt rendering and preserves all routing, identity, event, and retrieval information outside the two exact duplicate cases. ## Testing - `cargo test -p buzz-acp` — 804 unit tests and 9 lifecycle integration tests passed on the clean rerun - Focused mixed-thread, same-thread, and missing-root prompt regressions passed - The timing-sensitive `keepalive_resets_idle_past_deadline` test passed in isolation after one unrelated full-suite flake, then passed in the clean full rerun - `cargo clippy -p buzz-acp --all-targets -- -D warnings` - `cargo fmt --all -- --check` - `just file-size-check` - Repository pre-push gate at `1b88793574e74e22f00a1fd19bfe9aa8e86150fa` — file-size, branch, Rust unit-test, and Desktop Tauri lanes passed - `git diff --check` Generated with Codex Signed-off-by: Salman Mohammed <smohammed@squareup.com> |
||
|
thomaspblock
|
de188ebf06 |
feat(projects): add agent and CLI project-home support (#6590)
## Summary - inject bounded project-home identity and repository context into managed agent sessions - add project-aware CLI flows for creating projects, repositories, issues, and related channels - match project homes through the existing relay query surface, then filter channel metadata client-side This is Part 1 of the channel-first Projects stack. Part 2 contains project creation and model foundations. ## Testing - `cargo fmt --all -- --check` - `cargo clippy -p buzz-cli -p buzz-acp --all-targets -- -D warnings` - `cargo test -p buzz-cli -p buzz-acp` — 1,184 tests passed, 1 doc test ignored - full pre-push gate passed ## Post-Deploy Monitoring & Validation - validate project-home agent context and project-aware CLI commands against a staging relay - healthy signals: project context matches the active channel, explicit repo coordinates remain stable, and normal channels receive no project block - failure signals: cross-channel project context, duplicate project creation, or commands targeting an unrelated repository; mitigate by reverting this PR --------- Signed-off-by: Thomas Petersen <thomasp@squareup.com> Signed-off-by: Wrench <0eabe6ea5758c1e4c5b68cea4ac42b32c479072883cb28da8110b4a47c32b9a1@buzz.block.builderlab.xyz> Co-authored-by: Wrench <0eabe6ea5758c1e4c5b68cea4ac42b32c479072883cb28da8110b4a47c32b9a1@buzz.block.builderlab.xyz> |
||
|
Salman Mohammed
|
f177f4909d |
Use paired tags for standing & per-turn context (#6701)
## Why PR #6503 bundled a broader context compiler with several prompt-shape and telemetry changes, making it difficult to isolate whether paired XML-like top-level boundaries improve agent behavior. This keeps the follow-up experiment focused on framing only. ## What - Replace top-level standing and per-turn bracket headers with paired tags while preserving inner section content and event/message framing - Carry only metadata already present in dynamic headers as attributes, such as thread counts, truncation, event type, and batch count - Preserve every section body verbatim so the model executes the exact bytes reviewed in Desktop - Keep routing metadata, event content, batching, and delivery behavior unchanged - Teach Desktop diagnostics to parse the new framing while retaining compatibility with archived bracket-framed sessions; ambiguous paired boundaries render as the complete captured prompt ## Risk Assessment Medium — the change is limited to agent prompt framing and diagnostic parsing, but it affects both standing context and every dynamic agent turn. Bodies and delivery behavior remain unchanged, and compatibility parsing remains for prior sessions. ## References - Follow-up to #6503, scoped only to the paired-tag framing hypothesis ## Update — 2026-08-24 - Addressed review feedback by removing entity encoding/decoding from the execution and diagnostic paths. - Added regressions for literal closing tags, placeholders, entity-looking text, ampersands, and authored boundary whitespace. - Verified `cargo test -p buzz-acp` (805 unit + 9 integration), Desktop tests (5,412 before the main sync; 5,439 in the pre-push gate), Desktop typechecking, and Tauri clippy/tests. - Live local relay startup and message ingestion passed. The live ACP reply step is blocked locally because installed Goose 1.0.23 no longer provides the documented `goose acp` command and no alternate ACP adapter is installed. --- **Update Aug 25, 12:13 EDT:** Extended paired tags to per-turn top-level sections. - Converted context, conversation/thread context, event batches, and steer/interrupt boundaries without changing their inner bodies. - Preserved existing header metadata as attributes only; added no new prompt metadata. - Verified 806 Buzz ACP unit tests plus 9 integration tests, 5,441 Desktop tests, Desktop typechecking and formatting, and all differential pre-push lanes. --- **Update Aug 25, 13:58 EDT:** Addressed framing review feedback. - Updated operational base instructions to reference the emitted `<context>` section and added a regression guard against stale `[Context]` instructions. - Restored batched steer/interrupt cardinality in Desktop diagnostic titles with focused coverage for both paths. - Verified 807 Buzz ACP unit tests plus 9 integration tests, 5,442 Desktop tests, Desktop typechecking and formatting, and every differential pre-push lane including Tauri clippy/tests. --- **Update Aug 25, 15:37 EDT:** Hardened legacy transcript compatibility. - Semantic standing-context parsing now requires paired framing at the start and complete input consumption. - Historical bracket-framed personas keep literal balanced tag examples in their original section; malformed or mixed current framing displays as the complete captured prompt. - Added regressions for the reported legacy persona case and trailing unframed text; verified 60 focused parser tests, 5,444 full Desktop tests, formatting, and typechecking. --- **Update Aug 25, 16:31 EDT:** Preserved slash-command prompt block boundaries. - Desktop now sets aside a valid connector-facing slash-command block only when the following block begins recognized paired framing, parses the semantic blocks normally, and retains the command as visible Prompt metadata. - Added an end-to-end transcript regression using the production three-block ACP shape, including the user bubble, author pubkey, event ID, and grouped prompt context. - Verified 113 combined transcript tests, 5,445 full Desktop tests, formatting, typechecking, and all applicable pre-push lanes. Generated with Codex --------- Signed-off-by: Salman Mohammed <smohammed@squareup.com> |
||
|
Salman Mohammed
|
f99532585a |
refactor(acp): clarify agent prompt sections (#6501)
## Why The ACP prompt puts a machine-specific Workspace prefix before static base guidance and labels the user-facing agent instruction layer as the generic System section. Because the cwd varies by launch and worktree, leading with it reduces reusable prompt-prefix stability. `[Workspace]` was added in [PR #1194](https://github.com/block/buzz/pull/1194) as a defensive fix after a broken `~/.sprout` → `~/.buzz` migration caused agents to scan `$HOME` and trigger macOS TCC prompts. This change retains that grounding while shrinking it to the current working directory and moving dynamic environment context after the static Base prompt. ## What - Emit the prompt in Base → Workspace → Agent Instructions order - Reduce Workspace to `Current working directory: <absolute path>` - Resolve cwd as an absolute native-platform path and preserve Windows drive/UNC paths instead of checking for a leading `/` - Emit Agent Instructions for persona and standalone agent instructions across modern and legacy ACP paths - Preserve parsing for archived observer frames that used System or the former Workspace-before-Base order, and align the persona catalog label ## Risk Assessment Medium-low — this changes prompt framing for every newly created agent session. Existing archived observer frames remain parseable, and execution still uses the same ACP working directory. Cwd resolution now fails clearly instead of substituting `/` when the process directory cannot be resolved. ## References - https://github.com/block/buzz/pull/1103 - https://github.com/block/buzz/pull/1194 Generated with Codex --------- Signed-off-by: Salman Mohammed <smohammed@squareup.com> |
||
|
Salman Mohammed
|
025425591e |
fix(benchmarks): wait for scripted event delivery (#6487)
## Why The Buzz-native runtime could stop a solo agent after one turn even when a scripted follow-up event was queued or already running. That could turn a harness timing race into a benchmark failure. Follow-up to #6448. ## What - Treat solo tasks with scripted events separately from ordinary one-turn tasks. - Wait until turn counts and authored message IDs remain idle and unchanged for several polls before stopping the agent. - Remove the message-count shortcut that could stop an active turn. - Add focused regressions for a delayed follow-up turn and an already-running turn with a `DONE:` message. ## Testing - `pytest -q tests/test_container_runtime.py -k "solo_turn_end or scripted_events"` — 3 passed - `ruff check` on the three touched Python files - `ruff format --check` on the three touched Python files - Local `buzz-native-solo-luna.yaml` Harbor run, one attempt per affected task: - `cross-thread-requests` — reward 1.0; all 6 task-specific checks passed - `interleaved-agent-reports` — reward 1.0; all 7 task-specific checks passed - 2 completed trials, 0 exceptions, 56 seconds total ## Risk Low. This only changes completion detection for solo benchmark tasks with scripted events. Ordinary solo tasks retain immediate completion after their first turn ends. --- **Update Aug 21, 13:07 EDT:** Replaced the fixed settle delay with exact delivery receipts after [review feedback](https://github.com/block/buzz/pull/6487#discussion_r3832050452). - ACP now records the event IDs delivered by each completed turn; the runtime also recognizes existing successful-steer receipts. - Scripted trials stop only after every expected event ID is acknowledged and no turn is active. Missing receipts wait for the trial budget instead of producing partial evidence. - TDD regression delayed delivery beyond the old five-poll boundary and failed before the implementation; the receipt parser and Rust receipt format are pinned independently. - Focused verification: 31 container-runtime tests passed, 6 ACP delivery tests passed, plus Ruff, rustfmt, and Clippy. - Receipt-gated Luna run: 2 completed trials and 0 exceptions. `interleaved-agent-reports` scored 1.0. `cross-thread-requests` completed both calculations and thread isolation but Luna omitted the user mention on ALPHA, so that model-output dimension scored 0. - Risk remains low: the ACP production change adds delivery receipt logging and restores a missing turn-end log on a completed-before-control race; it does not change queue dispatch behavior. Generated with Codex --------- Signed-off-by: Salman Mohammed <smohammed@squareup.com> |
||
|
Will Pfleger
|
c6e3fe7dfa |
feat(workflows): reply in-thread from send_message action (#6178)
The `send_message` workflow action could only post new top-level channel
messages. This adds a `reply_in_thread` option so a `message_posted`
workflow can reply in the triggering message's thread, plus a
`trigger_is_reply` filter variable so a workflow can fire only on
top-level messages.
Every NIP-10 thread-marker reader in the tree now routes through one
shared parser and one collapse rule in `buzz-core`, deleting four
hand-rolled copies that had drifted on id-validity and marker semantics.
## What changed
- **Schema** (`crates/buzz-workflow/src/schema.rs`): new
`reply_in_thread: bool` (serde default `false`) on
`ActionDef::SendMessage`. `validate()` rejects `reply_in_thread: true`
on `schedule`/`webhook` triggers — they carry no message to reply to.
- **Executor** (`crates/buzz-workflow/src/executor.rs`): threads
`reply_to: Option<&str>` (the trigger `message_id`) through
`ActionSink::send_message` when `reply_in_thread` is set; errors clearly
if the trigger has no `message_id`. `resolve_step_templates` carries the
new field. Adds `trigger_is_reply` (boolean) to the eval context.
- **Relay sink** (`crates/buzz-relay/src/workflow_sink.rs`): when
`reply_to` is set, resolves parent/root/depth from the known trigger
event, persists real thread metadata instead of the hardcoded top-level
`depth: 0`, and pushes the live kind:39005 thread-summary overlay after
insert so subscribed desktops update the root's reply badge without
refetching — matching the human ingest path. Emits the same NIP-10
`e`-tag shape as `buzz_sdk::builders::thread_tags`: a single `["e", id,
"", "reply"]` tag for a direct reply (parent == root), and the `root` +
`reply` pair only when nested — so every writer produces one wire shape
per reply kind.
- **Ingest resolver** (`crates/buzz-relay/src/handlers/ingest.rs`): new
`resolve_relay_reply_thread_meta` + `ReplyAncestry`. Unlike the
client-facing `resolve_nip10_thread_meta` (which validates
client-supplied tags), this computes root/depth from the known parent
and enforces the same same-channel and depth-limit invariants. When the
parent has no `thread_metadata` row, both resolvers share
`derive_ancestry_from_parent_tags`, which applies
`ThreadMarkers::resolve()`: a marked nested parent remains depth 2,
while a root-only, malformed, or unmarked parent is top-level and starts
its reply thread at itself. Both the client resolver and the
parent-ancestor fallback now read markers via the shared parser instead
of hand-scanning `e`-tags.
- **Shared NIP-10 parser** (`crates/buzz-core/src/nip10.rs`):
`parse_thread_markers` reads an event's `root`/`reply` markers, honoring
a marker only when its event id is exactly 64 ASCII-hex characters.
`ThreadMarkers::resolve()` is the single definition of the (root, reply)
→ (root_id, parent_id) collapse: `root`+`reply` as-is, `reply`-only →
`(reply, reply)`, a lone `root` or neither → top-level. A slice-based
`parse_thread_markers_from_parts` entry point serves consumers holding
raw JSON tag arrays. Relay ingest, the workflow `trigger_is_reply`
predicate, ACP anchoring, and the CLI reply resolver all call these, so
they cannot drift on marker, id-validity, or collapse semantics.
- **ACP** (`crates/buzz-acp/src/queue.rs`): `parse_thread_tags`
delegates marker parsing and collapse to `buzz_core::nip10`, keeping
only its local `p`-tag mention collection. This fixes a parity gap with
ingest: a malformed non-64-hex marker id is no longer counted as a
thread link, and a lone `root` marker is now top-level rather than being
treated as root == parent.
- **CLI** (`crates/buzz-cli/src/commands/messages.rs`):
`find_root_from_tags` routes its JSON tags through the shared slice
parser and collapse rule. CLI replies to root-only, malformed, or
unmarked parents now correctly start at that immediate parent;
reply-only and root+reply parent ancestry remain intact.
- **Trigger context** (`crates/buzz-workflow/src/lib.rs`):
`build_trigger_context` derives `is_reply` solely from a valid NIP-10
`reply` marker (no DB hit). A lone `root` marker is top-level to ingest,
so it does not count; neither does a `reply` marker whose event id is
malformed, nor a bare/unmarked `e`-tag (mentions, quotes).
- **Desktop** (`workflowFormTypes.ts`, `WorkflowStepCard.tsx`):
`replyInThread` form field with YAML round-trip, and a "Reply to
triggering message in thread" checkbox shown only for message-based
triggers. Switching to `schedule` or `webhook` clears the option on
every step before serialization, so a value hidden behind an action
change cannot resurrect when the action is switched back. The form
parser also requires `reply_in_thread` to be a boolean and refuses
`reply_in_thread: true` under an ineligible trigger, falling back to
YAML mode rather than silently normalizing a backend-invalid definition.
## Usage
```yaml
trigger:
on: message_posted
filter: trigger_is_reply == false
steps:
- id: auto_reply
action: send_message
text: "pre-written reply text"
reply_in_thread: true
```
---------
Signed-off-by: Will Pfleger <pfleger.will@gmail.com>
Co-authored-by: Duncan <dcfd242e557282d7a1e2cf2e6877522682f1e5c6156dc92ca7d90eaedd3b0f95@buzz.block.builderlab.xyz>
|
||
|
Will Pfleger
|
d274a6e949 |
fix(acp): guard against unrequested public relay skills (#6394)
Adds a shared base-prompt instruction that agents must not read or blindly follow public Buzz relay skills unless a human explicitly requests them. --------- Signed-off-by: Will Pfleger <pfleger.will@gmail.com> Co-authored-by: Duncan <dcfd242e557282d7a1e2cf2e6877522682f1e5c6156dc92ca7d90eaedd3b0f95@buzz.block.builderlab.xyz> |
||
|
Salman Mohammed
|
2a236e4137 |
refactor(prompt): simplify Buzz agent guidance (#6340)
## Why The Buzz agent prompt carries verbose agent-draft instructions that add prompt weight without improving benchmark correctness. Buzz-native channel trials also showed agents creating permanent channels while claiming they had applied a one-hour TTL. The static CLI inventory remains for efficient command discovery by smaller agents. ## What - Retain the static Buzz CLI command table and direct agents to use `--help` for full usage - Compact the owner-reviewed agent-draft guidance without changing its workflow - Clarify that `buzz channels create` is permanent unless `--ttl` is provided ## Benchmark Results - Terminal-Bench retained 22/22 correctness in the initial no-table candidate while reducing prompt size, input tokens, tool calls, and cost; total active time was effectively flat. - Across two Buzz-native runs, the non-channel cases tied at 26/30; the CLI clarification directly addresses the observed TTL omission. ## Update — 2026-08-20 - Restored the CLI command table following review feedback about multistep discovery with smaller agents. The restored-table candidate has not been re-benchmarked. Signed-off-by: Salman Mohammed <smohammed@squareup.com> |
||
|
Salman Mohammed
|
934f3325c3 |
fix(buzz-acp): loosen workspace-scan guardrail to allow named paths (#6261)
## Problem Agents refused to read a specific, user-named path outside `.buzz` (e.g. `~/.claude/skills/context-health-check/SKILL.md`), and behaved inconsistently — refusing in one channel, listing the same dir in another. This is the issue Aaron and Wes hit: buzz://message?channel=e62570dd-33ad-42c5-b92b-75f2689f9694&id=401199157f8a9734109fdddfbbce4603a1ff465fae895042f29738ad52a5d15f Root cause: two prompt fragments told agents not to search `$HOME` "or other directories". The intent (added in #1194 by Will Pfleger) was narrow — stop unscoped `$HOME` sweeps that fire macOS TCC dialogs or make a model invent its own workspace dir. But "or other directories" reads as a blanket filesystem wall, so a cautious model generalized it to "any path outside `.buzz`" and refused a targeted, named read. ## Fix Prompt-only. Reframe both fragments per context-engineering guidance — state the reason and trust judgement, rather than banning a technique: **`crates/buzz-acp/src/pool.rs` — `workspace_section()`** > …This is where you already are, so start here rather than scanning `$HOME`. Any specific path the user names is fine to read. **`crates/buzz-acp/src/base_prompt.md` line 106** > These paths are relative to your working directory — start there for your own files rather than scanning `$HOME` or `/`. When the user names a specific path, read it. Preserves the original anchor intent (start at the grounded workspace; avoid unscoped `$HOME` sweeps) while explicitly permitting a named-path read — the exact case that was broken. ## Testing `cargo test -p buzz-acp` — 816 unit + 9 integration tests pass, 0 failures. Prompt-only change; the doc comment at `pool.rs:1327-1333` already states the narrow intent and is left as-is. Signed-off-by: Salman Mohammed <smohammed@squareup.com> Co-authored-by: Leo <5faf251baee50ee6bcde338aef6acdd70bb3e60115664c2cd490d94a55dfc488@buzz.block.builderlab.xyz> |
||
|
Luke Tornquist
|
08eb46ef3c |
Revert "fix(acp): gate relay-signed workflow messages on their attributed author" (#6311)
Reverts block/buzz#6129 |
||
|
Will Pfleger
|
50a71137e6 |
feat(managed-agents): close five Claude Code agent-config gaps (#4557)
Closes five Claude Code agent-config gaps in Buzz Desktop, split into three commits that share the spawn-time and live-switch surfaces. ## Config isolation, model authority, and Auto mode (#2692, #2884, #3493) - **`CLAUDE_CONFIG_DIR` isolation (#3493).** `config_bridge` resolves both `settings.json` and `.claude.json` panel paths from the agent's effective env (`resolve_effective_agent_env` — baked floor → definition → global → persona → record), so the panel reads the same directory the agent runs against. `mcp_config_file_path_for_runtime` honors a custom dir; empty/blank is treated as unset, matching Claude's `CLAUDE_CONFIG_DIR || homedir()` semantics. `AgentConfigPanel` shows a Keychain caveat when a custom dir is active (a custom dir maps to a fresh Keychain namespace unless `CLAUDE_SECURESTORAGE_CONFIG_DIR` is also set). - **Single startup model authority (#2692).** `ANTHROPIC_MODEL` is the sole startup model authority for Claude. Local spawns write `ANTHROPIC_MODEL` and strip `BUZZ_ACP_MODEL` so the harness never sees two authorities; remote deploys send `ANTHROPIC_MODEL` in `policy_env` instead of `BUZZ_ACP_MODEL`. Non-Claude runtimes are unchanged. - **`PermissionMode::Auto` (#2884).** Wire string `"auto"`, model-gated, degrades to the agent default when the active model doesn't advertise it. ## Thinking effort end-to-end for local Claude agents Effort flows from the running session's discovered `thought_level` config option through the config surface to a local-only write control and a read-only two-facts display. - **Reader.** Discovers the `thought_level` config option from the session cache (never hardcoded) and populates `effort_config_id` / `effort_options` on `RuntimeConfigSurface`. The canonical effort tier orders record env > `record.effort_level` (`BuzzExplicit`) > ACP > persona > global > definition > file, so the panel shows the effort the next spawn will launch with while `resolve_with_override` surfaces the running ACP value as the struck-through override — neither masks the other silently. - **Write control.** `persist_agent_effort_level` is a direct-write Tauri setter (writes `record.effort_level` + `updated_at`, `save_managed_agents`) that rejects non-local backends — remote effort is set at deploy time via `policy_env`. `EffortPickerField` mounts in `AgentInstanceEditDialog` beside the Model block, gated on a local backend **AND** a discovered `effortConfigId`. It persists directly and invalidates the config surface, mirroring the `setManagedAgentAutoRestart` standalone-setter precedent, so the frozen `UpdateManagedAgentInput` shape stays frozen. - **Display.** The read-only configured-vs-current two-facts display is the `thinkingEffort` normalized field in `AgentConfigPanel`, fed by the reader's canonical tier ordering. - `buzz-acp` applies the startup effort env at session start. ## Distinguish a rejected model switch from silent success A live model switch collapsed adapter rejection into success: `apply_model_switch` returned `Ok` on both an accepted switch and an application-level refusal, so the caller cached pre-switch capabilities as if they described the target model and Desktop reported the pick as landed. - `ModelSwitchOutcome::{Applied(Value),Rejected}`. Transport-class errors still propagate as `Err` (respawn the poisoned stdio); an application-level refusal is now `Rejected`. - The caller drives everything off `post_switch_snapshot`: `Applied` refreshes `model_capabilities` from the target model's echoed `configOptions` (or drops to `None` when none are echoed, so a pre-switch snapshot is never mistaken for the target's); `Rejected` preserves pre-switch caps and emits a `failure` `control_result`. Effort resolution and the session-config capture read the post-switch snapshot so they converge on the model the session actually runs; `modelOverridden` is gated on `switch_succeeded`. - `liveSwitchOutcome.ts` gains a distinct `"failed"` outcome for the adapter `failure` frame and treats the busy-path `"sent"` ack as provisional — it never counts toward success. Success is confirmed only by a real positive terminal frame (the busy-path deferred apply emits a correlated `switched` `control_result` when it lands), and the fallback timeout resolves an honest `"pending"` (accepted, apply deferred), never a false `"ok"`. `ModelPicker` surfaces a distinct toast per outcome — failed, unsupported, and pending. ## Scope explicitly excluded Per-agent config dir provisioning, `CLAUDE_SECURESTORAGE_CONFIG_DIR` sentinel injection, `settings.json` projection, protected-key stripping, MCP inheritance, spawn serialization, and the `last_spawn_warnings` surface are absent from this diff. Silent-fallback machinery for non-Claude runtimes (#2265/#4004) is a tracked follow-up. ## Sanctioned follow-ups - **Live mid-conversation effort switching.** Effort is spawn-scoped only in this PR: the worker reads `BUZZ_ACP_EFFORT_LEVEL` once and applies it at session creation. The live effort-switch machinery (mid-conversation effort RPC + ack frame) was deliberately removed and is archived on `archive/claude-config-gaps-live-effort` for a future plan-gated revival. - **Idle-path late model-switch rejection is unobservable.** An idle switch acks `switched` immediately after catalog validation, but the real `set_config_option` runs at the next session creation — potentially much later — so a rejection there is not surfaced back to the picker (holding a subscription that long is not sensible). Pre-existing, catalog-gated behavior; a durable fix is a tracked follow-up. Closes #2692, #2884, #3493 --------- Signed-off-by: Will Pfleger <pfleger.will@gmail.com> Co-authored-by: Duncan <dcfd242e557282d7a1e2cf2e6877522682f1e5c6156dc92ca7d90eaedd3b0f95@buzz.block.builderlab.xyz> |
||
|
Salman Mohammed
|
d2cfd377e2 |
fix(prompt): simplify pickup follow-through (#6186)
## Why The base prompt prescribed a todo lifecycle for every task needing follow-up tools, which added control calls even for work completed in one turn. This keeps the important behavioral contract—continue after publishing pickup—without prescribing the mechanism. ## What - Replace the 40-word todo lifecycle with a concise pickup follow-through rule - Preserve the requirement to publish the outcome or blocker before stopping ## Risk Assessment Low to medium — this changes managed-agent instructions, not runtime code. The terminal benchmark covers single-turn task completion but does not cover restart, compaction, or long-lived multi-turn recovery. ## References - Builds on the prompt simplification in https://github.com/block/buzz/pull/6161 - Benchmark setup: GPT 5.6 Terra at high effort; the same 11-task Terminal-Bench 2.1 slate; four concurrent trials; 4 CPU and 8 GiB per trial; 3× timeout | Prompt | Pass | Active-h | Median active | Tool calls | |---|---:|---:|---:|---:| | PR 6161 baseline | 22/22 | 0.3438 | 0.91 min | 216 | | Benchmarked 14-word rule | 22/22 | 0.2926 | 0.75 min | 198 | The benchmarked rule used 14.9% less active time, 16.9% lower median active time, and 8.3% fewer tool calls. Across the screen and confirmation runs it passed 33/33 trials with every completion report present; results are directional because model and tool behavior is stochastic. --- **Update Aug 18, 10:47 EDT:** Expanded the completion outcomes following review feedback. - The follow-through rule now covers a verified result, blocker, or key decision or information that needs to be surfaced. - The benchmark was not rerun; the table reflects the prior 14-word formulation. This is a completion-taxonomy clarification, not a return to a prescribed todo mechanism. Generated with Codex Signed-off-by: Salman Mohammed <smohammed@squareup.com> |
||
|
Salman Mohammed
|
f64899e5d1 |
Remove Startup Recovery section in base prompt (#6161)
## Why Managed channel sessions already receive authoritative per-turn context. The old startup recovery checklist told every new session to scan the global feed ## What - Remove `Startup Recovery` with concise channel and heartbeat turn contracts. ## Risk Assessment Low. This changes prompt guidance and its test only; routing and runtime behavior are unchanged. Generated with Codex --------- Signed-off-by: Salman Mohammed <smohammed@squareup.com> |
||
|
Luke Tornquist
|
54f11219ef |
fix(acp): gate relay-signed workflow messages on their attributed author (#6129)
## Problem Scheduled workflow `send_message` actions fire and land in the channel with correct `p` tags for the mentioned agents — but the agents never wake. The wake-up is silently dropped. **Root cause:** workflow messages are signed by the **relay keypair** (`workflow_sink.rs` signs with `state.relay_keypair`), so `event.pubkey` is the relay's pubkey, not the workflow owner. In `buzz-acp`, the inbound author gate (`author_allowed`) runs **before** the `p`-tag mention check. Under the default `respond_to = owner-only`, the relay pubkey is neither the owner nor a sibling, so every workflow wake-up dies at the gate with a debug-level `"inbound author gate — dropping event"`. The relay-side comment even says the mention `p` tags exist *"so mentioned agents are woken (wake is p-tag gated)"* — but wake is also author-gated, and that path was missed. ## Fix Gate relay-signed workflow messages on their **attributed author** — the pubkey that created the workflow — instead of the relay pubkey: - **Relay:** `workflow_sink.rs` now emits an explicit `buzz:workflow-owner` tag carrying `workflow.owner_pubkey` (the workflow creator, which the executor already passes as `author_pubkey` and whose channel access the relay verifies before emitting). Ownership is never inferred from `p`-tag order; mention `p` tags play no role in attribution. - **Harness:** at startup, `buzz-acp` fetches the relay's NIP-11 `self` pubkey (new `RestClient::fetch_relay_self`, public `/info` endpoint). Best-effort: fetch failure just logs a warning and preserves pre-fix behavior. - **Gate:** an event that is (a) authored by the relay `self` key, (b) tagged `buzz:workflow`, and (c) carries a well-formed `buzz:workflow-owner` pubkey is gated on that owner, through the exact same owner/sibling/allowlist policy as a direct author. ## Security notes (all fail closed) - No NIP-11 `self` pubkey → no exemption. - `buzz:workflow` / `buzz:workflow-owner` tags on a non-relay-signed event → ignored (a member cannot forge the exemption; the relay verifies signatures on submission and only the relay holds its key). - Relay-signed event without the tags, or with a malformed owner value (not 64-hex) → plain author gate. - Who is @mentioned in the message has no bearing on whose authority is evaluated. - A workflow owned by a random channel member still cannot wake an owner-only agent — the owner's pubkey must pass the same policy. ## Testing - 7 unit tests (`workflow_attributed_author_tests`) covering attribution, fail-closed paths, p-tag independence, malformed owner values, and the forgery case. - Extended the PG-gated `workflow_send_message_p_tags_mentioned_member` integration test to assert the `buzz:workflow-owner` tag. - `cargo test -p buzz-acp`: 785 passed, 0 failed. `cargo test -p buzz-relay --lib workflow_sink`: 17 passed. Clippy + fmt clean. (9 pre-existing `buzz-relay` failures in unrelated `api::media`/`api::admin` tests fail identically on the base commit without this change.) Found while debugging scheduled automations in a Buzz review-pipeline channel: two cron workflows fired daily @mentions at agents that never responded, while direct human @mentions woke them instantly. --------- Signed-off-by: Luke Tornquist <tornquist@squareup.com> Co-authored-by: Fizz <3a9f8a30fbb462abec1e2977b2280a7ae50c7ff794433790be15bd48bfd52d0b@buzz.block.builderlab.xyz> |
||
|
Atish Patel
|
5b3f0375a2 |
fix(acp): replace Goose native system prompt (#5964)
## Why Buzz currently appends its managed prompt to Goose's native prompt, so managed agents receive both instruction sets instead of the intended Buzz-only system prompt. ## What - Send Goose's custom session system-prompt request with `mode: "set"` - Lock the replacement contract in the ACP request test ## Risk Assessment Low — the change is limited to Goose session setup; adapters that do not implement Goose's custom method keep the existing method-not-found fallback behavior. ## References Goose v1.46.0 routes `set` to `override_system_prompt`, and its prompt builder selects that override instead of rendering the native `system.md`: [ACP handler]( |
||
|
thomaspblock
|
122a8b8988 |
Projects v3: unify sharing, discussions, and issue ownership (#5792)
## Summary Projects v3 makes repository work shareable, discussion-aware, and easier to scan in one coherent workspace. People can copy canonical links, reopen the exact workspace tab, understand issue and pull-request context at a glance, find related channel conversations, and assign or unassign issues across Desktop and CLI. - **Unified workspace** — top-level sections sit above repository controls in one rounded workspace, with navigation positioned close to the page heading. README and Files retain branch selection; every section has a labeled icon header, and Issues and Pull Requests expose creation from a consistent right-aligned action. - **Repository management** — the repository selector is always available, including single-repository projects. Its integrated add flow lets project owners create a repository manually or select an existing repository without a separate toolbar button. - **Readable work-item lists** — issue and pull-request rows use plain-language context instead of opaque metadata. Files, commits, issues, pull requests, channels, and contributors share consistent row density and right-aligned timestamps, while deterministic fallback-avatar colors keep participants distinct on light backgrounds. Inbox pull-request metadata wraps between complete phrases and truncates long channel names instead of compressing copy into narrow columns. - **Reliable entity links** — projects, repositories, issues, pull requests, and commits have canonical `buzz://` links, preview cards, OS deep-link routing, and tab-aware navigation. Reopening the same link re-applies its destination instead of leaving the user on a locally selected tab. - **Related conversations** — repository and work-item views surface channels discussing the current entity, including participants, channel navigation, message context, and an explicit notice when discovery reaches its 500-result cap. - **Reversible issue ownership** — trusted assignment and unassignment events work across Desktop, Tauri, `buzz-sdk`, and `buzz issues`. Assignees appear in project views and the assigned inbox, while authorized users can remove assignments directly from the assignee row. Assignment state is derived chronologically from labeled Nostr notes. Issue authors and repository owners may change any assignee; other users may only assign or unassign themselves. Shared golden fixtures keep entity-link grammar and validation aligned across TypeScript and Rust. The branch also updates `webbrowser` to the patched release for RUSTSEC-2026-0257. ### Related issue N/A. ### Testing - [x] `just ci` — formatting, lint, typechecking, unit tests, and builds passed - [x] Full pre-push suite — organization, branch-skew, Desktop checks, typechecking, and tests passed on the latest push - [x] `cargo test -p buzz-cli` and focused `buzz-sdk` assignment tests passed - [x] Focused Tauri recipient-note and 500-result search-limit tests passed - [x] Desktop entity-link and issue-assignment unit tests passed - [x] Playwright smoke coverage passed for assignment, repeated entity-link navigation, repository create/select flows, section headers and actions, timestamp alignment, timeline icons, sentence-style issue/PR metadata, header spacing, avatar contrast, and Inbox metadata at stacked and side-rail breakpoints - [ ] Manual staging pass: link round-trips, Channels tab, assignment flows, and inbox routing ### Screenshots Pull requests explain who opened the request, where it lives, and which branch it comes from; fallback avatars remain visually distinct.  Issues use the same sentence-style hierarchy while keeping status and recency easy to scan.  The wide Inbox detail keeps author, timestamp, and origin context readable beside its metadata rail.  [View the complete six-state Projects v3 screenshot set](https://github.com/block/buzz/pull/5624#issuecomment-5268039672) and [the compact/wide Inbox comparison](https://github.com/block/buzz/pull/5624#issuecomment-5268614585). --- > Supersedes #5624, whose head commit accumulated permanently-queued required check suites (block-dco-check et al.) that GitHub never dispatched. History flattened into a single signed-off commit on latest main; tree verified byte-identical (`git merge-tree`) to merging the original branch into main. --------- Signed-off-by: Thomas Petersen <thomasp@squareup.com> Co-authored-by: Wintermute <3f1797424fd9ad6653a83665c660517777cd7f8c228c0d5907f49e01537f3ca5@buzz.block.builderlab.xyz> |
||
|
Wes
|
1d51081b8a |
Teach agents to inherit Buzz product intent (#5875)
## Summary - make `VISION.md`, relevant `VISION_*.md`, and applicable testing guides explicit planning and review inputs for non-trivial Buzz changes - teach managed agents to load repository-root and path-local `AGENTS.md` files after selecting a checkout - distinguish CI evidence from exercising the live workflow for user-visible and integration behavior - turn repeatable mistakes into same-session durable lessons, keeping only load-bearing rules in core memory and promoting shared lessons to team guidance - pin the new managed-agent prompt invariants in tests - preserve the exact display name shown in Buzz when mentioning or addressing someone; never infer or look up a surname merely to sound more complete ### Related issue None found after searching `block/buzz` issues and PRs for agent instruction, vision, and product-intent routing. ### Testing At commit `07ef705b42f58d3be6981165c6959d541ada0ba7`: - `cargo fmt --all -- --check` - `cargo test -p buzz-acp agent_draft_prompt_tests` (4 passed) - mandatory pre-push hooks passed on the exact pushed head: `branch-skew`, `desktop-check`, `desktop-typecheck`, `mobile-test`, `desktop-test`, `rust-tests`, and `desktop-tauri-checks` - `git diff --check origin/main...HEAD` --------- Signed-off-by: Wes <wesbillman@users.noreply.github.com> Co-authored-by: Carl <c7ebe626f000404285d3686e1dc74cc07cc60a9754a150041ba132e14bd3e2ec@buzz.block.builderlab.xyz> |
||
|
Tyler
|
068a83b097 |
feat(huddle): cut voice-turn time-to-first-audio from ~1.0 s to ~0.35 s (env-gated latency levers) (#5671)
## Overview
**Category:** feat (env-gated experiment + one exact always-on
optimization)
**Problem:** Speech-end -> first TTS audio through the desktop huddle
pipeline measures **924–1087 ms** on an M4 Max with a 0 ms LLM leg.
Voice turns feel sluggish no matter how fast the agent replies. Baseline
breakdown: ~300 ms hardcoded VAD silence flush + ~150–250 ms Parakeet
decode + ~380–550 ms TTS synthesis before the first player append.
**Outcome:** With all levers enabled, e2e time-to-first-audio measures
**347–384 ms** (307–357 ms on a longer utterance) on the same hardware,
harness, and production pipelines. Defaults preserve production behavior
everywhere except one deterministic, bit-exact cache win.
## What's in here
### Levers (all default-off, env-gated)
| Lever | Env | Effect (measured medians, short utterance) |
|---|---|---|
| Speculative Parakeet decode | `BUZZ_STT_SPECULATIVE=1` | STT leg ->
~max(flush, decode) |
| Streaming TTS synthesis | `BUZZ_TTS_STREAMING=1`,
`BUZZ_TTS_EMIT_FRAMES` | first audio 380–550 -> 211–320 ms (emit=12,
bit-exact) |
| ONNX intra-op threads | `BUZZ_STT_THREADS`, `BUZZ_TTS_THREADS` | TTS
first audio 211–320 -> 129–180 ms (4 threads) |
- **Speculative decode** starts the Parakeet decode at the *first*
silent VAD frame, overlapping it with the flush window. Resumed speech
invalidates the result (voiced-frame-count check); held silence emits it
instantly at the flush boundary.
- **Streaming TTS**: new `synth_chunk_streaming` (buzz-voice)
interleaves the Flow LM frame loop with incremental *stateful* Mimi
decoding, emitting PCM deltas to the player via the existing
`PlaybackChunkAudio` decoration. At `emit_frames=12` (the decoder's
native chunk) streamed audio is **bit-identical** to the batch path —
verified by the ignored test
`incremental_stateful_decode_matches_batch_decode` (max|diff|=0).
Smaller deltas are faster but diverge (~23 dB SNR; decoder intra-chunk
lookahead), hence the default of 12.
> **Removed after live testing:** the `BUZZ_STT_FLUSH_MS` flush-window
override. Lowering the silence window below natural mid-sentence pauses
(the fast-path recipe said 150 ms) split single spoken sentences into
multiple messages and confused the listening agents. The window is a
turn-taking quality knob, not a latency lever — it is now fixed at the
production 300 ms value.
### Push-to-talk grouping fix (always-on)
A held push-to-talk shortcut is an explicit "I am not done talking"
signal, so silence never ends the utterance while it is held — even when
the microphone is also manually open. The utterance flushes on shortcut
release (existing transmit-edge flush); a manually open mic with the
shortcut up keeps normal VAD pause flushing. Gate is the pure
`vad_flush_allowed` function with a unit-test truth table.
### Always-on (exact): voice-conditioning cache
Phase profiling (`BUZZ_TTS_PHASE_LOG=1`) showed a fixed ~160 ms
`condition_voice` Flow-LM pass on *every* chunk, re-deriving the same
post-conditioning state for the same reference voice. The state is now
snapshotted after first computation and restored per chunk (dtype-tagged
tensor copies, keyed identically to the existing `cached_voice`).
Deterministic — same tensors in, same tensors out. The default path's
TTS leg drops from 380–550 ms to 225–355 ms with no configuration.
### Bench harness
`huddle::latency_bench` (`#[cfg(test)]` + `#[ignore]`) drives the real
`SttPipeline` and `TtsPipeline`, feeding a 48 kHz WAV in real-time 100
ms batches (AudioWorklet cadence) with a configurable fake LLM in place
of the relay leg, timing speech-end -> transcript -> speak() -> first
accepted player append.
```
BUZZ_STT_SPECULATIVE=1 BUZZ_TTS_STREAMING=1 \
BUZZ_TTS_THREADS=4 BUZZ_STT_THREADS=2 \
BUZZ_BENCH_WAV=<48k f32 mono wav> \
cargo test --release -p buzz-desktop --lib huddle::latency_bench -- --ignored --nocapture
```
## Tradeoffs to weigh before promoting any lever to a default
- **Speculative decode**: the speculative buffer has ~1 silent tail
frame vs ~19; observed one CTC wobble ("fail" vs "failed") in 24 turns.
Mitigation if productionized: zero-pad the speculative buffer to match
the flush-path shape.
- **Threads**: defaults stay 1 pending the min-spec (4-core Intel) A/B
flagged in the existing `STT_NUM_THREADS` comment.
- **Streaming at emit<12** is NOT the same waveform — don't ship below
12 without an ear pass.
## Validation
- Full desktop lib suite: **2408 passed / 0 failed** at this head
(`18fab2e1c`).
- buzz-voice suite green; bit-exactness test passes against the
production batch decode.
- Defaults-only bench rerun stays in the baseline family everywhere
except the exact conditioning-cache win (stt 525–532, tts 225–355).
- `cargo clippy --workspace --all-targets -- -D warnings` + fmt clean
(pre-push hook battery green).
Measurement notes with per-lever logs: Eva's workspace,
`RESEARCH/HUDDLE_E2E_LATENCY_OPTIMIZATION_2026_08_12.md` +
`RESEARCH/HUDDLE_E2E_STT_FAKELLM_TTS_BASELINE_2026_08_12.md`.
## Suggested promotion order
1. Conditioning cache (in this PR, always-on, exact).
2. Streaming TTS at emit=12: bit-exact audio, biggest UX win — needs the
env-gate removed + barge-in soak + an ear pass on a real huddle.
3. Speculative decode with silence padding: near-free ~100–150 ms.
4. Threads: after min-spec A/B.
---------
Signed-off-by: Eva <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz>
Signed-off-by: Wren <5217c5c2f7bfb4333e46d17c98a9255a52dadee18dcd43a43536b95e6776dfa0@buzz.block.builderlab.xyz>
Signed-off-by: Max <d8473ee32b973aa31a21a65adddcc4b69cc2a8a4dee8121ecd51926e0cddbc02@buzz.block.builderlab.xyz>
Signed-off-by: Mari <95cae996907d7cab9f5dbf43c0f53edeac6ab0b032a6feae4abfd784e467b3f5@buzz.block.builderlab.xyz>
Signed-off-by: tlongwell-block <tlongwell@block.xyz>
Co-authored-by: Eva <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz>
Co-authored-by: Wren <5217c5c2f7bfb4333e46d17c98a9255a52dadee18dcd43a43536b95e6776dfa0@buzz.block.builderlab.xyz>
Co-authored-by: Max <d8473ee32b973aa31a21a65adddcc4b69cc2a8a4dee8121ecd51926e0cddbc02@buzz.block.builderlab.xyz>
Co-authored-by: Mari <95cae996907d7cab9f5dbf43c0f53edeac6ab0b032a6feae4abfd784e467b3f5@buzz.block.builderlab.xyz>
Co-authored-by: tlongwell-block <tlongwell@block.xyz>
|
||
|
Atish Patel
|
4749bc7be3 |
feat(acp): report standard adapter usage (#4950)
## Why Claude Code and Codex expose standard ACP prompt-response usage, but Buzz only consumed Goose’s private cumulative usage notification. Their token use and Claude’s cumulative cost were therefore absent from NIP-AM metrics. ## What - Read per-turn `session/prompt` response usage for known Claude and Codex adapters - Publish Claude’s raw cumulative cost separately from per-turn tokens without changing the NIP-AM schema - Keep Goose usage exclusive and cover Claude/Codex wire serialization ## Risk Assessment Low-to-medium: changes best-effort observability only and does not affect prompt execution. The adapter-specific mappings preserve source semantics and omit unavailable fields. ## References - Validated with `cargo fmt --check`, `cargo test -p buzz-acp --no-run`, and full `cargo test -p buzz-acp` (678 passed at `652e373a` before merge-trailer amendment). Generated with Codex --------- Signed-off-by: Atish Patel <atish@squareup.com> Signed-off-by: Brainy Bumble <0ed7657b57c0e8a9f5288390dd6c8d5d0a3a06abe9b01b9006814f52077d6cdf@buzz.block.builderlab.xyz> Co-authored-by: WorkerBeeGPT <c5c455215c2506cb8ba776518cec804af62d3a0526e32d496a22072e395042b9@buzz.block.builderlab.xyz> Co-authored-by: Brainy Bumble <0ed7657b57c0e8a9f5288390dd6c8d5d0a3a06abe9b01b9006814f52077d6cdf@buzz.block.builderlab.xyz> |
||
|
Wes
|
dc2dbfe0f5 |
feat(buzz-acp): idle re-sleep for woken lazy pools (#5682)
## What Adds an opt-in **idle re-sleep** for woken lazy ACP pools. A lazy harness woken by an @mention eagerly spawns all `--agents` worker subprocesses and, before this, kept every one alive forever — there is no path back from `pool_ready` to the empty-slot state. Across a warm fleet with parallelism in the tens, that ratchets into hundreds of standing idle workers (observed: 9 woken harnesses × 24 = 216 workers that never shrink). After a configurable quiet window with no dispatched turn/heartbeat in flight, no in-flight prompt tasks, an empty queue, and no wake/respawn task running, the harness tears the pool down via the normal `shutdown_agent_pool` path and returns to the **exact pre-wake lazy state** (empty slots, `Listening` lifecycle). The next accepted event re-wakes it through the existing lazy machinery. **No second pool lifecycle.** ## Why it's safe - **Race-safe with enqueue/wake by construction.** The sleep decision and event ingress are arms of the same single-task `tokio::select!`. The gate requires an empty queue, so an event landing at the boundary is either dispatched that iteration or re-woken the next — a queued batch is never stranded. - **Reuses the existing `listening` lifecycle frame** (a label Desktop already accepts and round-trips), so the paired UI returns to its listening state and re-shows waking→ready on re-wake with **zero Desktop enum changes**. - **Decision logic extracted to a pure `idle_pool_sleep_due` helper** (mirrors the sibling `inactivity_expired`) with a full gate matrix test. ## Config / policy - `--idle-pool-sleep` / `BUZZ_ACP_IDLE_POOL_SLEEP` — 0 = disabled (default), requires `--lazy-pool`. - Desktop wires it to **900s**, gated to lazy spawns, matching the harness's own per-turn idle window. Reserved key (desktop-owned lifetime policy) so user env can't disable it. ## Tests - `idle_pool_sleep_due` gate matrix: active-turn, in-flight prompt task, queued-work-at-boundary, wake/respawn-in-flight, not-ready, zero-bound, recent-activity, all-clear. - Config parse (`--idle-pool-sleep`), reserved-key membership. - `cargo test -p buzz-acp` → **761 passed, 0 failed** at base `63f961c7e`. Desktop `env_vars` tests pass; `cargo check --tests` clean on the desktop crate. > Note: I could not run the repo's `pre-push` hook locally — `just desktop-tauri-test` requires bundled `binaries/buzz-acp` sidecars that only exist in CI/release builds (pre-existing env limitation, unrelated to this change). Pushed with `--no-verify`; CI runs the authoritative gate. ## Scope Idle re-sleep only. Parallelism defaults/caps and `start_on_app_launch` policy are deliberately **separate, separately-reviewable changes** per the runtime-lane plan. --------- Signed-off-by: Wes <wesbillman@users.noreply.github.com> Co-authored-by: Mongo <5c25403eab7271f9f94ddd4f2b270e8cac2c92e2c830c51877cca6ec974ffb3f@buzz.block.builderlab.xyz> |
||
|
Will Pfleger
|
6e0631f6b5 |
feat(acp): deliver channel description in prompt [Context] (#4552)
Channels carry a kind-39000 `about` description that the harness never surfaced to agents. This delivers it in the per-turn `[Context]` block so an agent knows what a channel is for without having to ask. ## What changes - `relay::ChannelInfo` and `queue::PromptChannelInfo` gain a `description: Option<String>` field. - The `about` tag is parsed in both metadata paths: the startup discovery map (`merge_discovered_channels`) and the lazy `fetch_channel_info` lookup. Blank or whitespace-only values become `None`. - `format_context_hints` renders a `Description:` line under `Channel:` for channel- and thread-scope turns. DM turns never render it. ## Safety - The description is newline-collapsed to a single line before rendering, so a multi-line `about` value can never spoof another `[Context]` field. - It is capped at 500 characters on a UTF-8 char boundary, with a `…` truncation marker. - Unresolved channel metadata renders no `Description:` line. Session creation is untouched — the description rides the existing per-turn `[Context]` block that already carries `Channel:`. Signed-off-by: Will Pfleger <pfleger.will@gmail.com> Co-authored-by: Duncan <dcfd242e557282d7a1e2cf2e6877522682f1e5c6156dc92ca7d90eaedd3b0f95@buzz.block.builderlab.xyz> |
||
|
Wes
|
563e4346da |
Reduce repeated ACP session context (#5423)
## Summary - deliver legacy ACP standing context once per live session, committing delivery state only after a successful turn - send only new thread/DM event deltas on later turns, with fail-open behavior for missing IDs and failed/cancelled prompts - fence native steer delivery acknowledgements by ACP session identity so stale acks cannot poison replacement sessions - keep context hints truthful when a fetch contains only the triggering event versus history delivered earlier ## Validation The pre-push hook passed on exact pushed head `6a768f1bc80fe63c686acf8d730f177fff8add3c`: - `branch-skew` - `desktop-check` - `desktop-typecheck` - `desktop-test` - `rust-tests` - `desktop-tauri-checks` Focused regression tests were also run while iterating: - `channel_prompt_commits_delivery_state_only_after_acp_success` - `in_flight_stale_native_steer_ack_cannot_update_replacement_session` - thread/DM trigger-only versus previously-delivered context hint tests ## Known limitations and follow-ups A local Goose smoke timed out at `session/new`. This diff does not change code that executes at or before `session/new`; its earliest affected runtime behavior is delivery-state insertion after session creation succeeds. The smoke failure is therefore bounded as environmental or pre-existing, but no successful live-provider turn was obtained. Scripted ACP wire/lifecycle tests carry the regression coverage. - #5421 — distinguish post-delta, already-delivered, and fetch-truncated context counts - #5422 — define a standing-context re-delivery policy if a legacy provider compacts it away Durable process-restart/session resume remains out of scope for this slice of #5342. #5386 also remains separate pending upstream adapter support. --------- Signed-off-by: Wes <wesbillman@users.noreply.github.com> Co-authored-by: Carl <c7ebe626f000404285d3686e1dc74cc07cc60a9754a150041ba132e14bd3e2ec@buzz.block.builderlab.xyz> |
||
|
Will Pfleger
|
5e4c05f90b |
feat(desktop): NIP-AM agent-usage backend — P2 emission/transport/archive + P4a aggregation/D6 (#4000)
## What Implements Phases 2 and 4a of the Usage v2 plan (plan events `d0268cd0`/`0e95b035`), extending the archive backend to emit, transport, archive, and aggregate both cache categories and billing identity fail-closed. ### P2 — emission, transport, archive **Tri-state accumulators** (`Unseen`/`Exact`/`Unknown`) for cache-read and cache-write in `buzz-agent` turn and session state. Absent field = Unknown (never zero) through the full pipeline. No `unwrap_or(0)` on the cache path. Both cache folds are gated on usage-bearing responses (same gate as the total-state and identity folds) — a response with no usage at all must not poison either accumulator. **Overflow-aware input token parsing and accumulation** — closed end-to-end from parse through wire to ACP: - `sum_usage()` returns `SumUsageResult` (`Exact(u64)` | `Overflow`) — checked arithmetic, never clamps. `anthropic_input_tokens()` returns `Option<SumUsageResult>` since it sums three fields (`input_tokens + cache_read_input_tokens + cache_creation_input_tokens`) that can collectively overflow. Single-field callers (`prompt_tokens`, `completion_tokens`, etc.) convert via `.into_exact()` — their single-field sums cannot overflow. - `LlmResponse.input_tokens_overflowed: bool` propagates the parse-layer signal into the run loop. When set, `input_tokens` is `None` (clamped value discarded), the context-gate baseline (`last_request_input_tokens`) is frozen at its prior reading, and `turn_input_tokens` is poisoned to `TurnIOState::Poisoned` before any emission — including mid-turn `emit_usage_update` calls. A dedicated enum on `LlmResponse.input_tokens` would ripple into ~20 existing test assertions on `r.input_tokens == Some(...)`; the bool flag confines the change to the two call sites that check it. - `TurnIOState` (`Unseen`/`Exact`/`Poisoned`) for input and output: per-round fold uses `checked_add`; overflow poisons permanently at turn and session level, no healing. Absence does not poison (pass-2-cleared contract unchanged). Wire emission omits `accumulatedInputTokens`/`accumulatedOutputTokens` when poisoned — never null, never `u64::MAX`. ACP treats absent = publisher-poisoned: `delta_reliable: false`, null turn fields, null cumulative for that category; session cumulative stays unknown for all subsequent turns once poisoned. **Conditional wire emission** for `accumulatedCachedInputTokens` and new `accumulatedCacheWriteTokens`: fields are omitted when the cumulative is Unseen or Unknown. ACP `_goose/unstable/session/update` contract documented next to the payload with tests for all absence/zero variants. **`PricingIdentity` stamping (publisher-side)**: - `pricing_authority()`: canonical parsed-URL endpoint comparison against the official allowlist — HTTPS only, exact allowlisted host (lookalike-safe), default port (omitted or explicit :443), required API base path, rejects userinfo/query/fragment/path-prefix lookalikes. - Model: the actually-requested `request_model` after mesh/auto resolution (not `effective_model_str`). - Turn discipline: identity retained only while ALL usage in the current turn carries one identical proven identity; any mismatch, unproven-usage-bearing response, or unpaired cumulative snapshot poisons to absent; a later matching notification does not heal a mixed turn. **ACP `UsageTracker` identity fold**: per-in-flight-turn tri-state identity accumulator replacing last-update-wins. Any absent identity on a token-advancing notification or exact mismatch poisons to absent; poison survives later updates; reset in `begin_turn()`/`take()`; reset also when a request fails (baseline cleared so preflight gate cannot stay frozen sub-threshold on retries). **M3 migration**: adds `turn_cache_write_tokens`, `cumulative_cache_write_tokens`, `pricing_authority`, `pricing_model`, `pricing_cache_class` to `agent_metric_index`. Additive, idempotent, guarded per-column by marker. M2 migration also guarded per-column (turn and cumulative cache-read columns checked and added independently; marker commits only after both are present). Fresh-DB schema includes all columns. **First-turn baselines**: `seed_zero_baseline` seeds `last_input: Some(0)`, `last_output: Some(0)`, `last_cached_input: Some(0)`, `last_cache_write: Some(0)`, and `last_total: Some(0)` — all have the known-zero-at-spawn argument. Absent fields from incoming snapshots still produce unknown (tri-state unchanged). Sessions buzz-acp did not spawn (no seed) remain fail-closed on turn one. **`ReportedUsage` TS mirror**: `cacheReadTokens`, `cacheWriteTokens`, `freshInputTokens` added to `tauriArchive.ts` as `UsageField` members, field-for-field with the Rust struct. ### P4a — aggregation layer **Extended S-1 ladder** to cache-read and cache-write via the same `ladder_token` path as the existing token fields. **`freshInputTokens` derivation**: checked arithmetic, fail-closed — absent cache fields produce Unknown (not zero), overflow and `cacheRead+cacheWrite > input` both produce `incomplete: true`. Aggregated as a `UsageField`. **D6 comparator**: `sort_value()` = provider total when known, else `input+output` when both known, else `None` (unknown-last). Replaces the prior total-only comparator for both agent-level and model-level sort. Ships a pinned test vector that the TS render layer (P5) must match. ## Test coverage - `buzz-agent`: 440 lib + 15 integration (golden_transcripts) — includes 13 new `cache_total_state_tests`; 14 new `turn_io_state_tests`; 3 new `sum_usage_*` tests (exact single-field, exact two-field, overflow signals correctly); 3 new `parse_anthropic_*` tests (overflow flag set + value cleared, normal sum no flag, absent usage no flag); end-to-end golden transcript drives real subprocess with Anthropic-shaped `input_tokens: u64::MAX, cache_read: 1` response and asserts `accumulatedInputTokens` absent from the emitted `usage_update` — no logic duplication; 3 wire pin tests; 4 `fold_pricing_identity_*` tests; `pricing_authority()` explicit-:443 acceptance - `buzz-acp`: 700 tests (691 lib + 9 integration) — 4 new usage tests (absent input → unreliable+null; absent output → unreliable+null; goose-shaped both present unchanged; poison mid-session); 3 ACP behavior tests; 7 pool lifecycle tests - Desktop (Rust): 2259+ tests — 14 new P4a pinned tests; 2 M3 round-trip tests; 1 serde key-shape test; 2 M2 partial-schema migration tests; first-turn cache round-trip test ## Related PRs - P1 NIP-AM spec: [#4632](https://github.com/block/buzz/pull/4632) - P3 pricing table: [#4629](https://github.com/block/buzz/pull/4629) - UI (P5): [#4001](https://github.com/block/buzz/pull/4001) --------- Signed-off-by: Will Pfleger <pfleger.will@gmail.com> Signed-off-by: Duncan <dcfd242e557282d7a1e2cf2e6877522682f1e5c6156dc92ca7d90eaedd3b0f95@buzz.block.builderlab.xyz> Co-authored-by: npub1mn7jgtj4w2pd0g0zeuhxsa6jy6p0rewxz4kujt98my82ahfmp72sxjexk7 <dcfd242e557282d7a1e2cf2e6877522682f1e5c6156dc92ca7d90eaedd3b0f95@buzz.block.builderlab.xyz> |
||
|
Wes
|
6a17d035f7 |
Revert "fix(acp): reject unattended permission requests" (#5323)
Reverts block/buzz#4609 |
||
|
Tyler
|
4da7264d90 |
fix(acp): pace observer telemetry at 1/s with per-channel batch envelopes (#4917)
## Problem Observer telemetry is the noisiest client of the relay: the old pacer (167ms spacing + 90/min rolling cap) let a busy session bill up to 6 events/second against the owner's message quota, and the rolling cap silently *dropped* frames once exceeded. Ruling from the rate-limiting investigation thread (channel `826fc99b-1472-40e7-a529-6b9db8943b8c`): pace at 1/s, always emit, minimal PR. **Review round 1 (Max, Sami)** found the first cut wrong in three ways — tick burst (all pending frames per tick), startup burst (`interval` fires at t=0), and per-channel quota arithmetic. All fixed and mutation-verified in round 1. **Review round 2 (Sami, Max)** found two more against the round-1 head: 1. **Drain-rate collapse (Sami, blocker):** front-run-only packing meant a frame held ONE event whenever channels interleaved — measured 275 B/s vs 63.5 KB/s, so an ordinary 2-channel session fell minutes behind with zero drops and no warning. Silent unbounded latency. 2. **Coalescer byte-cap bypass (Max):** chunks pending in `ObserverChunkCoalescer` were unbounded and outside the 4 MiB cap — 500 distinct-`messageId` 50KB chunks retained ~48MB with `pending_bytes == 0` and zero drops. **Review round 3 (Max)** found the drop accounting undercounted merged chunks: a coalescer entry that merged N same-`messageId` chunks counted as **1** in `dropped_events` when evicted (50 merged 1KB chunks evicted → counter read 1, 49 generated events unaccounted). Fixed: accounting is now denominated in **source (generated) observer events** end to end — each pending entry tracks how many chunks it absorbed, eviction charges that count, and the count survives flush into the publish FIFO. **Review round 4 (Sami, Max)** found three more against the round-3 head: 1. **Coalescer byte undercount (both, independently):** a pending merged entry retains its first chunk's text **twice** until flush — once inside the serialized event skeleton and once in the extracted text accumulator — but was charged only `serialized_len`, so true retention overshot the 4 MiB cap ~2× (measured 8.3 MB). Fixed: `push_pending` charges `serialized_len(&event) + text.len()`. 2. **Cap regressions asserted the accumulator against itself,** which is how the undercount hid. All three cap tests now assert on independently **walked** retained bytes (`serialized_len` per FIFO entry + `serialized_len + text.len()` per coalescer entry), with a secondary `accumulator >= walked` sanity check. Reverting the fix makes them fail at exactly 8,328,386 / 8,328,272 bytes. 3. **FIFO-arm source accounting was implemented but untested (Sami M13; Max reproduced at `cc9333b7c` with 102/151):** the round-3 regression only evicted a merged entry while still in the coalescer. New test forces a merged entry (50×1KB, `source_events=50`) through flush into the publish FIFO, then evicts it from there — mutating the FIFO eviction to `dropped += 1` fails with the reviewers' exact numbers (102 vs 151). **Review round 5 (Sami 9/9/9, Max 9/9/9)** — production judged merge-safe by both; remaining items are tests only, all landed at `63d821620`: 1. **The walker instrument was itself unverified (Sami M17–M20; Max independently confirmed the `return 0` mutant survives):** every cap test asks `walked_retained_bytes()` only for `<= CAP`, so a blinded walker passes everything — and paired with a reverted `push_pending` fix the two mutations cancel, hiding exactly the 8.3 MB overshoot it exists to detect. New two-sided pin: the walker must SEE the first chunk's text twice, and must agree with the accumulator EXACTLY while both stores are non-empty. Kills M17, M18, M19, M20. 2. **Two pre-existing snapshot-clone siblings (Sami D5b/D5c; byte-identical at merge-base `7334ad1e1` — not this PR's regression, but the PR made the class visible):** aliasing the inner turns map leaks a post-save turn into the snapshot; aliasing the inner tombstones map leaks a post-save terminal that blocks a legitimate post-restore resurrection. Two isolation tests with in-test controls — all three inner-map clones in `saveActiveAgentTurnsForCommunity` are now pinned. ## Change **Harness (`crates/buzz-acp`)** - **Global pacer: AT MOST ONE relay frame per second**, regardless of channel count or backlog size. `interval_at(now + 1s)` restores the no-startup-burst property; `MissedTickBehavior::Skip` is now pinned by a paused-time test (a stalled tick arm fires one catch-up frame, not one per missed deadline). At 1 frame/s telemetry spends ≤60/min of the shared 120/min quota; `OBSERVER_PUBLISH_TICK` documents the tradeoff as the knob. - **`ObserverPublishQueue` with gather-packing:** events wait as byte-accounted events (FIFO). `next_frame()` packs the front event's channel **gathered queue-wide in FIFO order** — frames never mix channels, and each channel's events keep their FIFO order, but cross-channel frame order MAY differ from arrival order. That is what keeps the drain rate in **bytes per slot** (one ~64KB frame/s) instead of front-run-length events per slot. **Null-channel events (`agent_panic`-class) are packing barriers** nothing gathers across, so causally-global events keep exact order against every channel. - **One byte cap over BOTH stores:** the event FIFO and the coalescer's pending chunk buffer count against the 4 MiB budget together; eviction is oldest-first across both (queue front, then coalescer front — structural age order) with accounting (warn + counter). A high-cardinality chunk flood is bounded exactly like a plain event flood. Coalescer entries are charged their **true** retention (`serialized_len + text.len()` — the first chunk's text lives in both the serialized skeleton and the extracted accumulator until flush). - **Shutdown is not a burst bypass:** paced one-frame-per-tick drain until empty. **Desktop** - `unwrapObserverBatch` expands envelopes on the live relay path and archive-ingest seam (round 1, unchanged). - **`activeAgentTurnsStore` watermark re-keyed per (agent, channel)** with a dedicated null-channel bucket: the per-agent `(timestamp, seq)` gate would silently skip a delayed channel's frames as stale under gather-packing's intentional cross-channel reorder. Safe because every turn-mutating path is channel-scoped by the event's own `channelId` (endTurn's null-turnId fallback matches `turn.channelId`; resurrectTurn keys on `event.channelId`), so per-channel serialization preserves each guard the per-agent gate provided. The tombstone-cap justification is rewritten for the new keying (worst case for an evicted tombstone is a ghost badge the prune reaps — bounded cosmetic staleness, not corruption). Community-switch save/restore deep-clones the nested map. Other per-agent maps stay agent-keyed: the clock offset is a running minimum (order-insensitive); turns/tombstones mutate only through channel-scoped paths. ## Version skew — old desktop + new harness Gather-packing *intentionally* emits cross-channel-reordered frames. An **old desktop** (per-agent watermark) against a **new harness** will silently skip a delayed channel's turn-state events as stale — working badges on that channel can go stale/missing until its next fresh event. Transcript and archive are unaffected (the transcript store sorts + rebuilds on out-of-order arrival; the archive is per-channel by construction). Ship desktop and harness together; skew degrades badges only, not data at rest. ## Throughput ceiling — "lossless" is qualified Sustained lossless rate is what fits in one ~64KB frame per second, now genuinely in bytes under interleaving: | event payload | events per frame | sustained ceiling | |---|---|---| | 100 B | 250 | 250 ev/s | | 500 B | 99 | 99 ev/s | | 2 KB | 30 | 30 ev/s | | 10 KB | 6 | 6 ev/s | With C channels producing concurrently, publish slots round-robin between them: per-channel drain is ~64KB/C per second and the 4 MiB burst budget (~64s single-channel) shortens accordingly. Beyond budget, oldest-first drops **with accounting** — visible, designed loss. **Accounting semantics:** `dropped_events` counts SOURCE (generated) observer events, not retained entries — evicting a coalesced entry that merged N chunks charges N. On the published side, a merged entry ships all N sources' text in ONE event, so the reconciliation invariant is `ingested == dropped_events + Σ source_events over published events` (for unmerged events, source_events = 1). ## Verification At `63d821620d3513505e8766ac691a8002f9d4a96f` (this head; `git rev-parse HEAD` matched in the same shell as every run), rustc 1.95.0: - `cargo test -p buzz-acp`: **689 lib + 9 integration, 0 failed** — regressions: interleaved 2-channel drain packs into ≤4 frames not 200 slots; null-channel barrier; queue-wide gather with within-channel FIFO; distinct-key 50KB chunk flood bounded by the cap with event-level accounting (published + dropped == ingested, survivors newest); paused-time `MissedTickBehavior::Skip` pin (verified to fail under `Burst`: 3 frames vs 1); merged-key eviction accounts every absorbed source chunk in BOTH arms — coalescer-side (Max's round-3 probe) and post-flush FIFO-side (Sami M13 / Max's round-4 probe: fails 102 vs 151 under `+= 1`). All three cap tests assert on independently walked retained bytes, not the accumulator (verified to fail without the `+text.len()` fix: 8,328,386 / 8,328,272 vs 4 MiB); the walker itself is pinned two-sided against the accumulator (all four blinding mutants M17–M20 verified to fail it, including the walker+fix cancellation pair). - `cargo clippy -p buzz-acp --all-targets -- -D warnings` clean, `cargo fmt --check` clean - Desktop: `tsc --noEmit` clean; node tests **4366 passed, 0 failed** — snapshot-clone family fully pinned: watermark aliasing (round 4), turns aliasing and tombstone aliasing (round 5, pre-existing gaps; each mutant verified to fail exactly its target test with an in-test control). Prior rounds: cross-channel reorder processed, cross-channel-delayed null-turnId `turn_error` evicts only its own channel's turn, null-bucket replay idempotency, same-channel stale/duplicate still skipped, watermark survives community-switch save/restore - All pre-push hooks green at the pushed commit (branch-skew, desktop-check, desktop-test, rust-tests, desktop-tauri-checks) Part of the rate-limiting fix stack; independent of `eva/rate-limit-fixes` by design (separate minimal PR per Tyler's ruling). --------- Signed-off-by: Eva <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz> Signed-off-by: Sami <f4a42a97e594b77bdbd8ee35191c8b28a94a4cb871d96f32921558275421fb68@buzz.block.builderlab.xyz> Co-authored-by: Eva <011987e296fd5006292d2f930b574be47c7801048d1983c46c425d3c95f0cffd@buzz.block.builderlab.xyz> Co-authored-by: Sami <f4a42a97e594b77bdbd8ee35191c8b28a94a4cb871d96f32921558275421fb68@buzz.block.builderlab.xyz> |
