Skip to content

Bound eval waits and compact study responses - #65

Merged
satyalyadav merged 4 commits into
mainfrom
fix/bounded-evals-compact-results
Nov 6, 2026
Merged

satyalyadav merged 4 commits into
mainfrom
fix/bounded-evals-compact-results

Conversation

@satyalyadav

@satyalyadav satyalyadav commented Nov 6, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Closes #64. Routine eval runs now use one model and bounded waits, and the conical/pointed study tools return a small setup reference with a getter for the full accepted evidence.

eval: deepseek-v4-1-flash -> 180 s request / 600 s conversation
  -> result row: scores + small failure progress
  -> requests row: bytes, duration, usage, request ID, error category

study: prepare / finish / fetch -> setup_revision + configuration_digest + snapshot
  -> get_study_setup(draft_id | result_id) -> full immutable accepted setup

Defaults are one model, one repeat, and two workers. Explicit --models IDs are sent as given; only the routine default is checked against the LibreChat catalog. Default result rows no longer carry prompts, replies, or tool text; --save-traces restores them. One retry is allowed for connection, 429, 502, or 503; read timeouts, context overflow, and client errors are not retried. The batch deadline is the documented external timeout -k wrapper, because the harness has no separate watchdog.

The pointed fixture now accepts all three named approximations explicitly. It previously accepted two and left the along-track antenna beam width for the model to question.

Evidence

  • Before: 240 s request with 3 retries, so a stalled request could hold a case for 12.5 minutes; five models by default; conical prepare 40,895 B, finish 47,144 B, events 39,554 B; pointed prepare 39,438 B, finish 44,355 B, events 38,905 B.
    After: 180 s request, 600 s conversation, one retry; DeepSeek V4.1 Flash by default; conical prepare 27,143 B, finish 20,012 B, events 11,706 B; pointed prepare 24,904 B, finish 16,433 B, events 10,423 B. The full setup is available on demand through get_study_setup.
  • Baseline 20261106T011438Z-bounded-baseline: 3/3 pass (point, conical, pointed). First final 20261106T015951Z-bounded-final-all: 2/3, with the pointed model pausing after preparation to ask about the one approximation the prompt did not accept. After the fixture clarification, 20261106T021304Z-pointed-final-r3 passed 3/3 with saved traces. Combined post-fix result: 6/6 conversations (point 1/1, conical 1/1, pointed 3/3). No checker was weakened.
  • Python 3.12 full suite: 501 passed. Focused HTTP-seam and public MCP regressions pass. The full-suite run used the Cartopy coastline preload.
  • Code-review Standards and Spec axes completed; fixes are in 7a0854e and 81ac0d0. Kiro Opus 5.5 reviewed and approved with caveats; the fixture clarification and dead-branch removal address the code caveats.

Merge Danger

Door: two-way

The branch is revertible, but the public response shape changes. Fetched pages no longer repeat approximations, source conflicts, or geometry checks; those stay in the accepted prepare/result snapshot and the full getter. Clients that read those fields from a page must call get_study_setup.

Blast Radius: server

Merging rebuilds and deploys the MCP image, then restarts LibreChat. The 180 s default is close to this model's observed turn latency; two conical repeats hit it during measurement. The final comparison runs used --request-timeout 240 to keep provider latency from masquerading as a score; the default remains 180 s, and callers can use the same override when a full comparison needs more headroom.

Bloat review: the diff is about 1,200 lines, roughly half tests. Removed the duplicate full configuration from the getter, a dead retry constant, and an unreachable context-overflow branch. Remaining small duplication is the acceptance identity at top level and in the snapshot reference, and the getter rebuilding the setup shape that the result builder also creates.

Redact raw error text from default result rows, keep explicit model selection independent of the catalog default check, bound stream probes and success records, remove duplicate full configuration from the setup getter, and add a redaction regression.
Accept the along-track antenna beam approximation explicitly in the pointed fixture so the case tests the full workflow, and remove the unreachable context-overflow branch after the retry loop.
@satyalyadav
satyalyadav merged commit b0c6fe0 into main Nov 6, 2026
1 check passed
@satyalyadav
satyalyadav deleted the fix/bounded-evals-compact-results branch November 6, 2026 19:14
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Bound eval waits and compact large study responses

1 participant