Conversation
- the managed model went through Anthropic's OpenAI-compatible endpoint, which does not support prompt caching; Pi's anthropic-messages adapter now targets a backend /chat/v1/messages route that keeps auth, quota and usage accounting - the route adds top-level automatic cache_control, prices 5m/1h cache writes and cache reads, rejects server tools, and classifies Anthropic errors - upgrade Pi to 0.82.1, the first release with claude-opus-5; bind provider auth explicitly since 0.82 rejects an unconfigured provider - raise the Opus 5 default context window to 300K (defaultModelsVersion 8)
- the Pi stream never emitted usage metadata, so the thread contextSize stayed null and the context-usage ring was hidden for every Pi-routed model - map Pi usage to the AI SDK shape at turn_end, including cached input tokens
- wire-level assertions for cache_control on the managed Messages route, cache-aware pricing, server-tool rejection and Anthropic error telemetry - Pi adapter and harness tests for the managed alias, context window and the turn usage emitted for the context indicator
Semgrep Security ScanNo security issues found. |
|
Preview environment deployed 🚀
Stack: Auto-destroys on PR close/merge. Login via the bundled Keycloak realm — |
- cli/ pins its own copies of the four @earendil-works/pi-* packages and CI requires them to match the root pi-ai version - AgentHarness no longer takes an execution environment in 0.82; the coding tools are bound to cwd directly, so the unused NodeExecutionEnv goes away
Sancus
changed the title
Sancus
changed the title
|
There is no small context price discount on Anthropic models anymore at <200K, so I raised the default context to 300K intentionally as part of this patch. 300K aligns well with the small context of other models(OAI) so if you're not doing 1M I think it's the sweet spot. |
PR Metrics
Updated Fri, 18 Sep 2026 20:41:02 GMT · run #3054 |
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
There was a problem hiding this comment.
🔭 thunder-deep-review (advisory)
Complements the other bots — surfaces only what they did not flag. Never approves, never requests changes, never gates merge.
head: eb893423245b · mode: deep · deferred 0 item(s) already reported by other bots (best-effort dedup)
There was a problem hiding this comment.
🔭 thunder-deep-review (advisory)
Complements the other bots — surfaces only what they did not flag. Never approves, never requests changes, never gates merge.
head: 16ce8409d3d6 · mode: deep · deferred 0 item(s) already reported by other bots (best-effort dedup)
There was a problem hiding this comment.
🔭 thunder-deep-review (advisory)
Complements the other bots — surfaces only what they did not flag. Never approves, never requests changes, never gates merge.
head: d2213da3f90c · mode: deep · deferred 0 item(s) already reported by other bots (best-effort dedup)
…ilures The native route's stream onError now spreads getApiErrorMetadata like /completions does, so requestId and errorType reach telemetry. Also share the latency recorder and admission gate between the two inference routes so timing headers and quota semantics cannot drift.
1 participant
Linear
THU-899
Summary
opus-5model through Pi's nativeanthropic-messagesadapter and a new backendPOST /v1/chat/v1/messagesroute instead of Anthropic's OpenAI-compatible endpoint, which does not support prompt caching. Pi marks the system prompt, last tool and last user block withcache_control; the backend adds top-level automaticcache_controlon every request./chat/completionsauth, quota admission, usage recording and latency telemetry. It validates the body (400 on malformed input), rejects Anthropic server tools, and classifies Anthropic SDK and mid-stream errors into the existing telemetry kinds withrequestId/errorType.@earendil-works/pi-ai,pi-agent-core,pi-coding-agent(andpi-tuiincli/) 0.80.7 -> 0.82.1, the first release whose catalog includesclaude-opus-5. Provider auth is bound explicitly (boundApiKeyAuth) because 0.82 rejects an unconfigured provider;AgentHarnessno longer takes an execution environment, so the app and CLI harness builders drop it.contextWindow200K -> 300K;defaultModelsVersion7 -> 8 and the lineage migration carries the new value to unmodified rows.usagemetadata atturn_end, sochat_threads.contextSizeis persisted and the context-usage indicator renders for Pi-routed models (it previously only worked on the legacy AI SDK path).@anthropic-ai/sdkto the backend.Test Plan
bun run tsc --noEmit,bun run lint,bun run test,bun run test:agent-corebun run type-check,bun run lint,bun testcli/:bun run typecheck,bun run test,bun run build, artifact--version/--helpsmokemessage_startusage reportscache_read_input_tokens > 0; the context ring appears in the composer after the first reply.Changes
backend/src/inference/routes.ts-POST /v1/messageshandler; sharedcreateUsageCallbacks/streamingResponseHeadershelpers used by both routesbackend/src/inference/client.ts-getAnthropicMessagesClientbackend/src/utils/anthropic-streaming.ts- re-encodes SDK events as SSE, collects final usage, reports stream errorsbackend/src/inference/usage-ledger.ts,error-kind.ts- cache pricing; Anthropic error classificationshared/agent-core/anthropic-model.ts,build-app-harness.ts,bound-api-key-auth.ts,openai-compat-model.ts- managed alias / base URL / context window on the native descriptor; Pi 0.82 auth and harness changesshared/agent-core/pi-to-aisdk-stream.ts-usagemessage-metadata atturn_endsrc/acp/built-in-adapter.ts,src/ai/fetch.ts- routethunderbolt+vendor: anthropicmodels to the native path (Pi and legacy AI SDK);resolveManagedAnthropicConnectionderives from the shared thunderbolt auth policyshared/defaults/models.ts,src/lib/data-migrations/upgrade-model-defaults.ts- 300K defaultcli/package.json,cli/bun.lock,cli/src/agent/harness.ts- Pi 0.82.1 pins and harness construction