6.3 KiB
Buzz Agent API-provider slice — 16b4
Executable candidate: 36e2eaca0 (preceding implementation 575254ac9).
Pinned runtime/config source: 051c3a270be9c73da9ab06700bcab7d5552fceaa.
Delivered scope
Three distinct Buzz Agent API-key contracts: Anthropic, OpenAI-compatible, OpenRouter. Normal setup option 8 and local immutable add-buzz-provider share source-grounded input. Credentials, endpoint, wire selection and service paths remain local. Public model/workspace/binding selection, profile/configuration CAS, Start/Restart use existing owners. No automatic selection/restart/login/install. Existing Claude/Codex API-only restrictions and ten diagnostic presets unchanged.
Config consumers: crates/buzz-agent/src/config.rs:631–691,922–960 at pinned
source. Anthropic key/base URL; OpenAI COMPAT key/base URL plus auto/chat/responses;
OpenRouter key/base URL, Chat only. Universal BUZZ_AGENT_MODEL overrides provider
model defaults. No invented generic env map or native profile transport.
Databricks lib.rs:151–200, auth.rs:1287–1325,1356–1392 read: explicit native
service-user login with DATABRICKS_HOST, headless runtime refresh, cache identity
by discovery URL/client/scopes. Default service HOME .config/buzz-agent/oauth,
override root BUZZ_AGENT_CONFIG_DIR/buzz-agent/oauth. No auth stores read.
New Databricks token setup/richer OAuth diagnostic guidance NOT delivered.
Validation (not vendor inference)
Node 24.15.0, pnpm 11.4.0, standalone existing locked dependencies; no install/config changes. Strict PASS. Provider input/launch negatives 3/3. Focused initial normal installed journeys 3/3, 58.527s. Real wizard → host CLI → actual TUI → real installed buzz-acp/Buzz CLI with source-shaped TypeScript provider harness: four signed same-agent replies each, two channels/later turn, fresh-session Restart/exact profile/model, sibling/history/manifest unchanged. Final suite additionally passes wrong private credential, missing credential, model/native rejection and neutral Stop controls; public inventory excludes local key/path/HOME/endpoint. Unit checks cover endpoint/wire/mixed-contract/unsafe-file rejection and all three OpenAI wire mappings. No real provider network inference. Local add/diagnostic conversion is implemented but not an executed new installed journey; custom conversion also remains source-only.
One FULL DEFAULT concurrent installed-enabled suite (no concurrency override): 111/112, zero skips/cancellations, natural 223.753s. All ten installed normal journeys PASS. No rerun, timeout change, force exit or weakened assertion. Gate remains OPEN. Failure is admission control conflicting-id, aggregate 2612ms versus unchanged <2500ms. Earlier wrong-authority control passed at 1712ms; later controls in that test were not reached. This is not a provider test failure and not evidence that the original 2572ms wrong-authority failure is classified.
Durable owner evidence at this failure
Original instrumentation-complete.patch SHA256
37de1d9fd5110017a8be3caf739a5b3a4cf56ee5e512b77ea1d7619580f84129
reused/adapted narrowly. Trace inactive normally. Test activates with
BEEHIVE_LATENCY_TRACE=<fresh private evidence directory>; bounded 50k rows,
no payloads/credentials/results/URLs, exclusive 0600 file+directory fsync on
completion/failure, t.after fallback if cleanup fails. No operational timing or
authority policy change. Exact source and binary pins retained separately.
Conflicting-id trace has 472 stamps, zero dropped. Relative monotonic ms:
| Owner event | ms |
|---|---|
| Start reaches slot handler | 153.806 |
| Start receipt seal | 1344.754 |
| Conflicting Stop handler / conflict receipt seal | 1348.277 / 1348.289 |
| Conflict receipt first decoded | 1625.343 |
| Valid Stop submitted / handler | 1642.016 / 1759.741 |
| Owned Stop begin / end | 1915.733 / 2447.602 |
| Valid Stop receipt first decoded | 2593.462 |
| Aggregate timer end | 2611.454 (wall elapsed 2612) |
A synchronous writePrivate(journal.json) span of 582.982ms, 761.769–1344.750, blocks this event loop during legitimate Start completion. File fsync itself is 19.850ms and directory fsync 18.321ms; do not call the whole span fsync time. Before file-sync begin: 392.743ms; between file-sync end and dir-sync begin: 152.055ms. Those subregions still combine filesystem calls and scheduling without finer stamps. Maximum timer lateness is 582.785ms. Concurrent async relay file.write await is 615.456ms and overlaps that synchronous blocking region; it is NOT established as 615ms disk execution. Owned Stop is 531.869ms (existing 500ms containment policy), not rejection latency. Conflict handler-to-seal is 0.012ms after Start; unknown-agent bypass in the original remains distinct.
Thus this recurrence has an event-associated owner-level decomposition, including a dominant synchronous journal region and legitimate Stop contribution. Underlying OS/I/O scheduling cause remains unresolved; no speculative fix. Original admission failure remains unresolved independently. Next engineering action is review this captured failure at journal/write/scheduling owners, not another blind full suite.
Evidence and boundaries
Workspace WORK_LOGS/BEEHIVE_DESIGN_183FFAF0/PROVIDER_16B4/:
full-default.log, strict-final.log, final-pins.txt, trace-analysis.txt,
traces/83342-{wrong-authority,conflicting-id}.json, all focused iteration logs,
instrumentation-adapted.patch, post-run-processes.txt.
First journey failed because fixed fixture session ID violated fresh identity;
next two failed on later-session restoration (set_model, then prompt owner).
Fixture corrected without runtime changes or assertion relaxation. First strict
inference failure and failed patch application/strict iterations retained/documented.
Installed binaries, read-only:
- buzz-acp SHA256
10612d0025d1420bdd9e9afcc2e7129377da2414e75049a8b640e81d857441ea - buzz SHA256
147cc2ccf276ddedc5b84d13f5399a95282066303c9dca566bb2af5a37b856c5
No production/current-kind40002, OAuth login/cache access, real vendor inference, Rust/native/client changes, full repository CI, mesh/compute/live admission claim. Custom confirmation review COMPLETE SOURCE-ONLY/no new blocker; R1 source-closed plus actual wizard regressions. Previous scoped reviews remain closed; historical broker/tool report complete/unclassified. No independent review of this new delta.