Tags give the ability to mark specific points in history as being important
-
v1.4.2
6343fe81 · ·Operational metrics bucket + vivarium infrastructure reliability fixes, folded into a single 1.4.2 release. Added: - Per-task metrics bucket (timing, throughput, tokens, chunks) with reasoning_status regime flag and regime-conditioned nulls; identical key sets across OpenAI-compatible, Azure, Anthropic and vivarium endpoints - Run-level aggregates with _mean/_min/_p50/_p95/_max - Retry visibility (operational.attempts), total_chunks counter - Anthropic reasoning_tokens from thinking_tokens (SDK >= 0.69) Fixed: - Vivarium lifecycle stages (task-files upload, seed, untrusted-swap steps, teardown destroy) now retry transient infra failures via a shared retry_transient helper in palace.utils.infra_errors - Run-poll 404 treated as transient (orchestrator collapses worker hiccups into 404 when one worker is registered) - Verify-file injection retries before verify_infra_error classification - Progress lines no longer show elapsed=0.0s (read wall_seconds from the operational bucket); dispatch error entries emit operational keys Docs updated to the operational schema; new tests for retry semantics (88 passing).
-
v1.4.1
c5b9dd4c · ·Fix CI: remove deprecated CLI entrypoints - Delete palace-cli, palace-run, palace-download entrypoints and the palace/entrypoints/ package outright (banner shims from the v1.4.0 consolidation broke the CI 'Verify CLI entrypoints' check). - CI now verifies only the unified entrypoints: bare 'palace' (help), 'palace download --help', 'palace run --help'. - Docs and skill references updated from the old command names.
-
v1.4.0
97166e9d · ·Task failure isolation, agentic timeout fix, download module consolidation Fixed - Task failure isolation: untrusted-agent swap failures (VERIFY_PATHS download errors) are now per-task error outcomes instead of FatalEvaluationError — one broken task no longer aborts the whole job. Transient download errors retry with exponential backoff (2s→60s, 5 attempts); 404 retries once, then fails the task permanently. Judge config is validated in a preflight before dispatch, so misconfiguration fails in seconds instead of after a full agent run. (v1.3.1) - Agentic per-task timeout raised 2h10m → 3h30m: the dispatch timeout now sits above the agent's graceful 3h timeout, so long tasks are cancelled server-side with metrics preserved instead of killed with a bare TimeoutError and an unreconciled vivarium run. - Verify-stage infra errors (504s, connection errors) are excluded from the accuracy denominator instead of grading tasks as incorrect. Changed - Download module consolidation: engine moved to palace/download/engine.py; conversion_recipes.json is now the single source of truth for public benchmark recipes (public_datasets_info.json deleted). Legacy palace-cli and palace-run CLIs removed (entrypoints are now "removed" banner shims). - omp is now the default vivarium agent harness. - Public recipe catalogue trimmed: removed BABILong-32k/128k (superseded by the full-length BABILong), CURIE-protein, SimpleQA-Verified, VLSBench.