Skip to content

0.9.7

Choose a tag to compare

View all tags
@RobotSe7en RobotSe7en released this 04 Sep 07:19
· 22 commits to main since this release
v0.9.7
662a6d5

dagent 0.9.7 unifies model-context execution for private vLLM deployments across Chat Completions and stateless Responses.

Added

  • Added one provider-neutral model request/response contract used by SDK runtime model calls, with request-aware Chat Completions and Responses selection.
  • Added reasoning_replay="none|active_run|all_runs" for explicit same-run and cross-run reasoning replay policy.
  • Added private-vLLM capability discovery, exact /tokenize accounting, automatic max_model_len discovery, explicit output limits, and model-call audit metadata.
  • Added independent compaction_reasoning_effort and token-pressure-based context compaction.

Changed

  • Responses calls are stateless: store=False, without previous_response_id, encrypted reasoning content, or persisted provider item IDs.
  • Context compaction now triggers at 80% of usable input capacity, targets 16% retained recent history, and allows summaries up to 8,192 tokens by default.
  • context_window_tokens=None uses the vLLM-discovered model window; an explicit value overrides discovery but cannot exceed the server limit.
  • max_output_tokens now controls both input budgeting and the protocol-specific wire output limit.

Fixed

  • Protocol selection now validates the capabilities required by the exact request and final reasoning projection.
  • Failed, cancelled, incomplete, and output-truncated generations are no longer accepted as successful runs.
  • Hard context overflow compacts additional oldest cross-run history before reporting an avoidable window error.
  • Review checkpoint resumes preserve the frozen context window and explicitly unset output limit.

Compatibility and migration

  • Remove ReasoningConfig.enabled and budget_tokens; use typed effort only.
  • Replace ContextPolicy.keep_recent_turns with reasoning_replay.
  • Replace output_reserve_tokens with optional max_output_tokens; leaving it unset sends no output limit.
  • New checkpoints use schema V8, RunState V5, and ConversationState V4. Older checkpoints are rejected, so finish pending reviews before upgrading.

Known limitations

  • Built-in discovery and unified protocol behavior target private vLLM deployments. Custom Chat providers continue through the explicit compatibility adapter.
  • Automatic discovery requires vLLM to expose /openapi.json, /version, and /tokenize; otherwise the documented warnings and 32,768-token fallback apply.

Verification

  • Python suite: 1116 passed, 3 skipped.
  • Web suite: 125 passed.
  • Web production build completed.
  • Wheel and source distribution built successfully and passed Twine metadata checks.

See the English and Simplified Chinese migration guides for complete wire examples and upgrade details.

Assets 2
Loading