mirror of
https://github.com/block/buzz.git
synced 2026-08-18 06:50:31 +02:00
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>