Skip to content

0.7.7 — 2026-04-27

Choose a tag to compare

View all tags
@brijesh1100 brijesh1100 released this 28 Apr 05:26
· 41 commits to main since this release
v0.7.7
e05d739

0.7.7 — 2026-04-27

Added

  • Context Management v2 recall and masking tests covering policy classification, squeeze gating, marker parsing, synthesis rehydration, recall round-trips, and streaming/non-streaming masking symmetry.
  • Autonomous-mode stability tests covering Critic/Refiner routing, abstention, compute-budget escalation, and tool-budget synthesis behavior.
  • Tool deduplication metadata via opt-in BaseTool.idempotent / @tool(idempotent=True). The default remains False so live-data tools are never deduplicated unless explicitly declared safe.

Changed

  • Stabilized Context Management v2 around one Compactor, bounded observation markers, recallable offload, and synthesis-time rehydration.
  • Kept ContextConfig.squeeze_threshold=0.70 as the stable default after validating that later masking can miss hard context limits on PDF-heavy runs.
  • Bounded marker previews to a small orientation preview only for large tool results, preventing marker bloat from becoming a second evidence store.
  • Restored Task E to the normal benchmark surface (list_pdf_inventory + read_annual_report_excerpt) and quarantined evidence-pack/search helpers as experiment-only artifacts, not framework behavior.

Fixed

  • Synthesis rehydration now uses the engine's resolved model context window when ContextConfig.max_context_tokens is auto-detected, instead of silently skipping rehydration when the config field is None.
  • AgentResult now surfaces legacy mode error-string sentinels as status=error when the agent state is already ERROR, including a stable ToolCallLimitError classification for exhausted tool budgets.
  • Standard mode forces a tools-free synthesis pass when the configured tool-call cap is reached and synthesis is enabled, so autonomous runs can still reach validation/refinement instead of stopping at a raw tool-limit string.

Provider Updates

  • OpenAI provider (nucleusiq-openai 0.6.3) — keeps provider behavior in sync with the stable V2 core changes, including Responses/Chat tool conversion and response-normalization fixes covered by provider tests.
  • Gemini provider (nucleusiq-gemini 0.2.5) — keeps Gemini response normalization and base-call behavior aligned with the updated core message/tool contracts.

Validation

  • Focused stable gates: 1340 passed across src/nucleusiq/tests/unit/context, src/nucleusiq/tests/agents/unit, test_context_management_e2e.py, and test_recall_round_trip_e2e.py.
  • Full local gate excluding provider-backed memory integration: 2469 passed, 2 skipped.
  • Full local gate including provider-backed memory integration: 2469 passed, 2 skipped, 8 failed because the configured OpenAI project returned 403 model_not_found for gpt-4o-mini. These are external/provider access failures, not release-blocking framework regressions.

Known Limitations

  • This release does not claim that long evidence-heavy tasks are fully solved.
  • Task E is not production-grade on gpt-5.2; it remains an internal research benchmark.
  • Workspace state, document search, evidence dossier, and autonomous phase control are deferred to the next feature line.

Packages

Package Version Note
nucleusiq 0.7.7 Stable V2 context/autonomous framework release
nucleusiq-openai 0.6.3 Provider sync for stable V2 message/tool contracts (nucleusiq>=0.7.7)
nucleusiq-gemini 0.2.5 Provider sync for stable V2 message/tool contracts (nucleusiq>=0.7.7)
Assets 2
Loading