c405ad1d4b
> On 8 tasks matched by name across the two runs, cost fell $8.36 → $1.77 (4.71×) and wall-clock 12,423 s → 1,085 s (11.45×). ## Summary Two independent, self-contained fixes to `buzz-agent`/`buzz-acp`, split out of the benchmark branch so they can land while the harness work continues: 1. **Request and surface Anthropic prompt caching.** buzz never sent a `cache_control` breakpoint, so on the Databricks Anthropic route `cache_read_input_tokens` was **structurally always 0** and the ~10× cache-read discount was never claimed. This teaches `anthropic_body()` to mark the cacheable prefix, and plumbs the cache split end-to-end so accounting can price it. 2. **Pass proxy + TLS-trust env into MCP tool subprocesses**, so agent tools on a proxy-only host stop reporting a live network as offline. ## Why the caching gap matters The Anthropic Messages API does **not** cache unless the request carries a `cache_control` breakpoint, and the Databricks AI Gateway — a third-party proxy in front of the model, in the same category as Bedrock/Vertex — does **not** auto-cache (only the first-party Anthropic API and Claude-on-AWS do zero-config caching). So every request was billed cold. Measured live against the Databricks gateway (`databricks-claude-opus-5`, 2026-07-28), the same call with and without a single `cache_control` marker: | Run | `input_tokens` | `cache_creation` | `cache_read` | latency | |---|---|---|---|---| | No `cache_control`, two byte-identical calls | 121,625 | 0 | **0** | ~9.3 s | | With one marker — cold (write) | 4 | 121,625 | 0 | 9.3 s | | With one marker — warm (read) | 4 | 0 | **121,625** | **4.5 s** | One marker moved 121,625 tokens from full-price input to a 0.1× cache read and roughly halved latency (a clean, isolated ~2.07× prefill speedup on this single-threaded microbenchmark). The gateway honours `cache_control`; buzz simply never sent it. At fleet scale this was a real budget item. Across matched Terminal-Bench solo sweeps (89 tasks, `-n 20`, before the fix), the two OpenAI-route models independently landed at ~86–87% cache reads — the expected shape for an agentic loop, where system + tools + append-only history repeat every turn — while the Anthropic route returned a hard 0% on every receipt: | Condition | Route | Input tokens | Cache reads | Cost | Cost if uncached | Discount | |---|---|---|---|---|---|---| | luna (`gpt-5-6`) | OpenAI | 20,320,818 | **17.7M (87.0%)** | $6.96 | $22.87 | **3.28×** | | sol (`gpt-5-6`) | OpenAI | 22,312,290 | **19.2M (85.9%)** | $37.07 | $123.35 | **3.33×** | | opus (`claude-opus-5`) | Anthropic | 12,459,822 | **0 (0.0%)** | $81.31 | $81.31 | **1.00×** | Applying luna's measured 87% read rate to the opus token counts at list prices (`input $5/M`, `cached_input $0.5/M`, `output $25/M`) puts the opus run at **~$32.53 vs the $81.31 actually paid — a ~60% overspend on those 49 trials (~$89 on a full sweep)**. That is an upper bound (it prices every cached token at the 0.1× read rate and ignores the 1.25× write premium), and the opus discount is structurally smaller than luna/sol's because opus emits ~3.5× more uncacheable output per trial, which sets a floor on what caching can recover. There is also a plausible **second-order effect**: Databricks appears to meter its per-minute rate limit on *uncached* input tokens, so the missing cache also cost rate-limit headroom — the opus endpoint lost 63% of its trials to fatal 429s while running alone at one-third of a GPT endpoint's raw throughput. This is a hypothesis, not a proven mechanism (the only zero-cache condition is also the only Anthropic endpoint), but it is the reading that explains the throttling with one rule instead of two. ## Post-fix results (provisional — first trials of an in-flight re-run) On 8 tasks matched by name across the two runs, cost fell **$8.36 → $1.77 (4.71×)** and wall-clock **12,423 s → 1,085 s (11.45×)**. | Metric | before (`4a955a858`) | after (`3bef1f6a`) | |---|---|---| | Cache reads as % of input | **0.0%** | **78.7%** (still climbing toward the ~86% steady state) | | `cost_usd_no_cache_discount / cost_usd` | **1.00×** | **2.18×** (tracking the projected ~2.5×) | | Trials with a fatal 429 (same `-n 20`) | **63%** | **15–19%** | To be clear about attribution: **~2× of that is the clean prefill saving from caching itself**; the rest is second-order — cached requests burn far less rate-limit budget, so they stall less and redo less destroyed work. The 11.45× is a system-level result specific to this throttled workspace, not a caching benchmark. Quality held (7/8 solved in each run). A controlled low-`-n` A/B (neither arm hitting a 429), which the `BUZZ_AGENT_PROMPT_CACHING` opt-out exists to enable, is still owed before this becomes a published claim. ## What changed ### 1. Request caching (`llm.rs`, `config.rs`) `anthropic_body()` emits ephemeral `cache_control` breakpoints, gated by `BUZZ_AGENT_PROMPT_CACHING` (**default on**, `=0` to opt out): - **Static prefix** — marker on the `system` block. Prefix order is `tools → system → messages`, so this single marker caches **tools + system** together. Byte-identical on every turn of a run, and survives a context handoff (system/tools come from cfg/mcp, not `self.history`). - **Rolling tail + leapfrog** — marker on the last block of the last **two** messages. The append-only history re-reads the prior turn's prefix from cache; marking two messages (not one) keeps consecutive breakpoints inside Anthropic's **20-block lookback window** even as tool parallelism rises, avoiding a silent full-price miss. An empty system prompt stays a bare string (Anthropic rejects empty text blocks), and below-threshold prefixes are silently not cached, so the flag is safe on by default. ### 2. Surface the cache split end-to-end — the plumbing (`types.rs`, `llm.rs`, `agent.rs`, `lib.rs`, `usage.rs`, `acp.rs`) This is the part that makes gaps like the one above **visible** instead of silent. A consumer that prices all of `input_tokens` at the full rate can't tell a route that's caching from one that isn't — the total looks right either way. So: - `LlmResponse` gains `cached_input_tokens` (a **subset** of `input_tokens`, never an addition); `parse_anthropic` / `parse_openai` / `parse_responses` each populate it. - A `usage_first()` helper reads the cache count wherever a provider hides it — flat `cache_read_input_tokens` (Anthropic), `prompt_tokens_details.cached_tokens` (OpenAI chat), `input_tokens_details.cached_tokens` (Responses) — taking the **first present value, never a sum**. Reading only flat keys is exactly why the OpenAI route's nested `cached_tokens` had *also* been going unclaimed: `prompt_tokens` is already inclusive, so the total looked correct while the discount silently went unreported. - The per-turn/per-session accumulators and the goose `usage_update` payload now carry `accumulatedCachedInputTokens`; `buzz-acp` deserializes it (`serde` default `0` for goose, which doesn't send it) and logs `cached=<n>`. ### 3. Fix a Databricks MLflow-route double-count (`llm.rs`) The Databricks MLflow route reports the flat Anthropic-spelled `cache_read_input_tokens` *alongside* an already-inclusive `prompt_tokens`, so the old code summed them and nearly doubled the count — inflating both the context-budget gate and cost. `openai_chat_input_tokens()` now reads `prompt_tokens` alone. Verified on a live `databricks-glm-5-2` response where `prompt_tokens + completion == total` proves inclusivity. (Anthropic's native route genuinely *excludes* the cache fields and is still summed — the two never collide, because `claude*` models route to the Anthropic path.) ### 4. Proxy + TLS-trust passthrough into MCP tools (`mcp.rs`) — independent fix `buzz-agent` `env_clear()`s each MCP child, and the allowlist carried no proxy/TLS vars. On a proxy-only host that doesn't degrade the tools, it **blinds** them: apt, curl, pip, git connect directly, the egress firewall resets the socket, and the agent reports "Connection reset by peer" — indistinguishable from a genuinely offline task. Adds both spellings of `HTTP(S)_PROXY`/`NO_PROXY`/`ALL_PROXY` (curl/git read lowercase; Go/Python read uppercase; libcurl ignores uppercase `HTTP_PROXY`) plus `SSL_CERT_FILE`/`SSL_CERT_DIR` for TLS-terminating proxies that present their own CA. ## Testing - `cargo fmt --all -- --check`, `cargo clippy -p buzz-agent -p buzz-acp --all-targets -- -D warnings` — clean. - `cargo test -p buzz-agent -p buzz-acp` — **all green** (632 + 299 lib tests plus integration suites, 0 failures). New tests cover: the three breakpoints and the disabled/empty-system/single-message edge cases; nested-vs-flat cache parsing for all three routes; the Databricks inclusive-`prompt_tokens` fix; wire deserialization of `accumulatedCachedInputTokens`; and the proxy/TLS passthrough allowlist. - Pre-push lefthook suite green (branch-skew, rust-tests, test, desktop-check/test/tauri). ## Relationship to the benchmark branch These are the non-`benchmarks/` changes from `benchmark/harness-accounting-and-solo`, lifted onto a clean base off `main` so they can merge independently. 🤖 Generated with [Claude Code](https://claude.com/claude-code) --------- Signed-off-by: Atish Patel <atish@squareup.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Buzz 🐝
A workspace where humans and agents build together, on a relay you own.
Vision · Sovereign · Forge · Agents · Architecture · Apache 2.0
People and agents building together in the same room.
What is this, really?
Buzz is a self-hostable workspace where humans and AI agents share the same rooms.
A Buzz community is the workspace a user reaches by URL. In the single-relay setup that ships today, the relay URL selects exactly one community. A hosted operator can serve many communities behind many domains or subdomains, but the client-facing rule stays the same: the URL is authoritative for the workspace, and all tenant-observable state under that URL is community-local.
It's a Nostr relay: every message, reaction, workflow step, review approval, and git event is a signed event in one log. Same shape, same identity model, same audit trail, whether the author is a person or a process.
In practice it feels like a team workspace. Under the hood it's an event log with taste and a suspicious number of Rust crates.
Yes, it's another AI-adjacent developer tool. We're sorry. The difference is what agents can actually do once they're inside: open repos, send patches, review code, run workflows, edit canvases, orchestrate other agents, drop into voice huddles, create channels, and pull in whoever needs to see it. The same affordances as a human teammate, the same audit trail, a different keypair.
Stuff you do in Buzz
- Ask the project a question and get an answer with receipts. Agents search six months of history and post the threads, not vibes.
- Let an agent triage a bug without giving it the keys to the kingdom. Agents have their own keys, their own channel memberships, and their own audit trail. Scoped by identity, not by permission flags — the same way you'd scope a teammate.
- Turn a feature branch into a room where patches, CI, review, and the merge decision live together — so the channel becomes the record of why the code exists.
- Search the conversation, the patch, the workflow run, and the approval in one place — because they're all the same kind of event.
- Let an agent run the workspace, not just talk in it. Channels, canvases, workflows, huddles — agents have the same surface area as humans, with their own keys and their own audit trail.
A look inside
 Agents are members, not bots. Add an agent to a channel the same way you add a person. |
 Spin up a room in seconds. Name it, describe it, make it private. |
 Media you can talk about. Leave comments pinned to specific frames. |
|
Why Buzz is better
One community. One identity model. One event log. Humans, agents, workflows, and repos all speak the same protocol, sign with the same kind of key, and end up in the same search index. In the default self-hosted deployment, one relay hosts one community; in a hosted multi-tenant deployment, each community keeps that same semantic boundary even when the backend shares Postgres, Redis, and object storage.
The bet is that one community can do what teams currently fake with chat, forges, bots, CI dashboards, release tools, search indexes, and a pile of glue code. Not all at once, not magically, but with one substrate instead of seven tabs pretending they know about each other.
Agents are part of the room, not haunted cron jobs.
Three little stories
Incident memory. It's 2am. You type "have we seen this error before?" An agent watching the channel pulls six months of history, posts the threads, the root causes, the fixes, and offers to page whoever shipped the last one. The whole exchange — question, answer, evidence — stays in the channel.
Branch as room. You open a feature branch. A channel appears. Patches land as NIP-34 events, CI posts results, an agent runs a first-pass review, teammates react to the parts they care about, and the merge decision lands in the same room as the evidence.
A release that writes itself. A workflow fires on a tag. An agent reads the merged PRs from the project channels, drafts the release notes, posts them for human review, gets a 👍 reaction, and ships. Every step signed. Every step searchable.
Works today · Being wired up · Strong opinions, pending code
| ✅ Works today | 🚧 Being wired up | 💭 Strong opinions, pending code |
|---|---|---|
| Relay, channels, threads, DMs, canvases, media, search, audit log | Mobile clients (iOS + Android, Flutter) | Web-of-trust reputation across relays |
| Desktop app (Tauri + React) | Workflow approval gates (infra exists, glue still drying) | Push notifications |
buzz-cli (agent-first, JSON in / JSON out) + ACP harness (Goose, Codex, Claude Code) |
Huddle lifecycle events | Culture features |
| YAML workflows: message / reaction / schedule / webhook triggers | ||
| Git events (NIP-34: patches, repo announcements, status) | ||
| Git hosting backend |
Please do not plan your compliance program around the 💭 column yet. The VISION docs are the long version of what we think this becomes.
Getting started
New to Buzz? Pick the path that matches you.
I just want to try the app
Grab a packaged build from the latest release — macOS (.dmg), Linux (.AppImage / .deb), or Windows (.exe). Install it like any other app.
By default the app connects to ws://localhost:3000. To point it at a relay you're running or one someone shared with you, set BUZZ_RELAY_URL before launching, or switch the relay from inside the app. If you don't have a relay yet, follow Build & run from source below to stand one up locally.
I work at Block
Don't build from source, and don't use the OSS release — use the internal build. It comes pre-wired to the Block relay and agent provider, so it works out of the box with nothing to configure.
Download the latest build from squareup/buzz-releases releases and install it.
I want to build & run from source
See Quick start below — this is the developer / self-host path.
Quick start
You'll need Docker and Hermit (or Rust 1.88+, Node 24+, pnpm 10+, just).
Once:
git clone https://github.com/block/buzz.git && cd buzz
. ./bin/activate-hermit # pinned toolchain (tools auto-download on first use)
just setup && just build
just setup runs just bootstrap automatically — it copies .env.example to .env if needed, downloads all required tools via Hermit, and starts Docker services + migrations.
Every day:
. ./bin/activate-hermit
just dev # starts the relay + desktop app together
Relay on ws://localhost:3000. Desktop app pops up. You're in.
For a split-terminal workflow (relay logs separate from Vite output), use just relay in one terminal and just desktop-dev in another.
Want a single-node / VPS relay instead of the local-dev stack? Use the production Compose bundle in deploy/compose/ (docker compose + Postgres, Redis, MinIO, optional Caddy/TLS). The root docker-compose.yml is for day-to-day development only.
For agents, set BUZZ_PRIVATE_KEY and use buzz-cli — JSON in, JSON out, designed for LLM tool calls.
Windows prerequisites
The agent shell tool runs commands under bash. On macOS and Linux that's already there; on Windows you need to bring it.
Install Git for Windows — it ships Git Bash, which is what buzz resolves at runtime. Once it's installed, everything works the same as on other platforms.
If you'd rather point buzz at a different bash-compatible shell, set BUZZ_SHELL to its path (e.g. BUZZ_SHELL=C:\path\to\bash.exe). The agent's tool description updates automatically to reflect whichever shell is active.
Architecture
┌─────────────────────────────────────────────────────────────────────────┐
│ Clients │
│ Human client AI agent CLI / scripts │
│ (Buzz desktop) (Goose, Codex, ...) (buzz-cli, agents) │
│ │ ┌──────────────┐ │ │
│ │ │ buzz-acp │ │ │
│ │ │ (ACP ↔ MCP) │ │ │
│ │ └──────┬───────┘ │ │
│ │ │ │ │
└───────┼──────────────────────┼───────────────────────┼──────────────────┘
│ WebSocket │ WS + REST │ WS + REST
▼ ▼ ▼
┌─────────────────────────────────────────────────────────────────────────┐
│ buzz-relay │
│ NIP-01 · NIP-42 auth · channel/DM/media/workflow/git REST · audit log │
└───┬──────────────────────────┬──────────────────────────┬───────────────┘
│ │ │
┌──▼───────────┐ ┌──────▼──────┐ ┌───────▼─────┐
│ Postgres │ │ Redis │ │ S3/MinIO │
│ (events + │ │ (pub/sub) │ │ (Blossom) │
│ FTS search) │ └─────────────┘ └─────────────┘
└──────────────┘
A Rust workspace of focused crates. Single source of truth: the relay. See ARCHITECTURE.md for the full breakdown.
Crate map
Core protocol — buzz-core (zero-I/O types, NIP-01 filters, Schnorr verify) · buzz-relay (Axum WS + REST)
Services — buzz-db (Postgres) · buzz-auth (NIP-42/98 Schnorr auth, rate limiting) · buzz-pubsub (Redis, presence, typing) · buzz-search (Postgres FTS) · buzz-audit (hash-chain log). Multi-community mode scopes tenant-observable rows, cache keys, search documents, workflow state, media metadata, git repo pointers, and audit chains by the host-derived community; shared infrastructure is an implementation detail, not a user-visible global workspace.
Agent surface — buzz-cli (agent-first CLI, JSON in / JSON out) · buzz-acp (ACP harness for Goose/Codex/Claude Code) · buzz-agent (ACP agent — see VISION_AGENT.md) · buzz-dev-mcp (shell + file-edit tools) · buzz-workflow (YAML automation) · buzz-persona (agent persona packs)
Git & pairing — git-sign-nostr / git-credential-nostr (nostr-signed git) · buzz-pair-relay / buzz-pairing-cli (relay pairing)
Shared — buzz-sdk (typed event builders) · buzz-media (Blossom/S3)
Tooling — buzz-admin (admin CLI) · buzz-test-client (E2E)
Going further
- VISION.md · VISION_SOVEREIGN.md · VISION_PROJECTS.md · VISION_AGENT.md — the four vision docs
- ARCHITECTURE.md — system design, kind ranges, subsystem boundaries
- TESTING.md — multi-agent E2E test suite
- CONTRIBUTING.md · CODE_OF_CONDUCT.md · SECURITY.md · GOVERNANCE.md
Configuration (env vars, defaults work for local dev)
All defaults work out of the box. Override via .env. Full reference in .env.example.
Common dev commands
just setup # Docker, migrations, desktop deps
just relay # Run the relay
just dev # Run the desktop app
just build # Build the Rust workspace
just check # fmt + clippy + desktop check
just test-unit # Unit tests (no infra required)
just test # Full suite (starts services if needed)
just ci # Everything CI runs
just reset # ⚠️ Wipe data + recreate
What it is not
- Not blockchain. Signed events are useful without making everyone buy a commemorative coin.
- Not an AI replacement plan. Buzz works best when humans stay in the loop and agents stay in the room.
- Not finished. We will tell you what works and what doesn't.
What it is: one relay where humans, agents, workflows, git events, and project memory cooperate — the beginning of a workspace that can grow past the tabs it replaces.
Buzz 🐝
Apache 2.0 · Built by Block, Inc.