Route session auto-naming through the provider adapter - #1137
Conversation
Session auto-naming was the last auxiliary request still issuing a bare `/chat/completions` POST with URL-sniffed provider selection, duplicating credential, wire-dialect, and failure-classification logic that the provider adapters already own. Add `call_naming_provider()`, mirroring the compaction summarizer's adapter path: it streams one non-tool turn through `provider.chat`, inheriting the adapter's wire dialect, credential handling, failure classification, and usage accounting (via the same `account_provider_stream` wrapper). The orchestrator (`generate_session_title`) now routes through it. `call_naming_llm()` is retained as a legacy direct-POST helper for callers that still construct a target from a raw URL + API key; its wire contract is unchanged and covered by the existing mocked-httpx tests. Verification: 62 naming tests pass (9 new for `call_naming_provider`, including a real-adapter integration test); full `tests/test_session/` plus `tests/test_engine/test_usage_http.py` = 624 passed, 1 skipped; ruff clean.
|
Thanks for addressing this. I validated the change against the I recommend keeping this PR open and updating it before merge. The local integration needed one import-conflict resolution, and the adapter path currently introduces these regressions:
Validation: the two naming test files pass 102 tests on the baseline. The local integration has 105 passing and 6 failing tests: two new tests conflict with main's updated prompt/title filtering, and four existing HTTP tests need SSE fixtures after the transport migration. The 23 focused usage-accounting tests and Ruff pass. The synthetic end-to-end probes above independently demonstrate the production regressions; they are not conclusions drawn only from those outdated test fixtures. This was targeted offline validation on macOS, not full CI or live-provider acceptance. The adapter-based approach is useful; a rebase preserving the current naming behavior, these regression tests, and fresh required CI should make this ready for another review. After validation, main advanced to |
|
Merge queue CI: failure — run 35286935546, attempt 1.
The run summary lists reused suites and executed checks. A green PR check alone does not mean queue validation has passed. |
|
Merge queue CI: success — run 35292862877, attempt 1. The run summary lists reused suites and executed checks. A green PR check alone does not mean queue validation has passed. |
|
Merge queue CI: success — run 35293033811, attempt 1. The run summary lists reused suites and executed checks. A green PR check alone does not mean queue validation has passed. |
2 participants
Scope
Session auto-naming previously sent OpenAI chat-completions requests even when the selected provider used Anthropic's native protocol. Route production naming through the selected provider adapter so its protocol, credentials, request constraints, and usage accounting match normal provider calls.
Scope boundary: preserve explicit naming model/tier selection and session overrides using an isolated selector; keep the physical request model and usage ledger consistent. Preserve resolved output limits (normally 512 tokens, TokenRhythm 1024, subject to lower configured limits), reasoning controls, correlation, title sanitization, manual rename precedence, and bounded stream cleanup. Structured refusal/content-filter responses retain the fallback title while completed usage remains accounted.
Non-goals: change routing policy, ordinary chat behavior, the public gateway event schema, or remove the legacy direct-HTTP naming helper.
Branch
Base branch: main
Target exception: N/A
Issue
Linked issue: None
If None, reason: this original PR tracks the native-provider naming defect and its regression coverage.
Release Note
Release note: Fix session auto-naming for native provider protocols while preserving configured naming models, usage accounting, and fallback titles.
Tests
Ruff:
ruff check src testspassed.Mypy:
mypy src/opensquilla --show-error-codespassed (1608 source files).Pytest: 658 focused cases passed: 48 terminal-cleanup cases plus 133 naming/real-adapter/refusal/target-selection cases and 477 provider, accounting, request-golden, and stream-golden cases. HTTP integration fixtures use real selectors/adapters with offline MockTransport and check the resulting usage ledger. Coverage includes native Anthropic messages, explicit model/tier and session overrides, registry aliases, resolved token caps, timeout/cancellation/cleanup, refusal evidence, sanitization, and concurrent manual rename.
Build: the previous updated PR passed its full CI. The merge queue exposed a 2-second timeout in an unchanged Gateway test while formatting an intentionally raised exception. A controlled baseline/candidate comparison reproduced the same failure by delaying only the renderer. The test now captures and asserts its expected warning while preserving the timeout and all durable-ack assertions. Current main is integrated; required CI runs again on this final update, and final merge-queue checks remain authoritative.
Regression tests: added
Notes: the 15 stream-golden changes each add only the default internal Provider DoneEvent
refusal: falsefield; a value-level comparison verified no other snapshot changes, and the normal non-regenerating suite passed. Local focused validation ran on macOS/Python 3.12 with explicit source imports. These provider changes are platform-neutral; local results do not claim native Linux/Windows or full-suite acceptance.The default test path remains offline, deterministic, credential-free, and safe for forks.
Maintainer Live Check
Maintainer live check: no
Surface: provider
Maintainer-only note: protocol and accounting regression checks use synthetic offline responses; contributors do not need credentials.
Safety
All fixtures use synthetic requests, dummy credentials, and isolated local storage. No real prompts, transcripts, secrets, machine-specific paths, runtime data, or assistant scratch artifacts are committed. Refusal metadata is an internal boolean without retaining refusal content; the separate public gateway Done event is unchanged.
Third-Party Origin
Third-party origin: none
Original contribution by @QinLuza remains in this PR's history. Maintainer follow-up commits integrate current main and repair regression gaps without replacing the contributor's PR.
Documentation Changes