docs(agentic): define trajectory observability contract - #472
Conversation
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 7fffb7d2c6
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| multiple model requests. Agent code returns an ``AgentTrajectory`` containing | ||
| one or more ``AgentTrajectoryTurn`` objects. AReno merges turns that carry the | ||
| same ``AgentItem`` and exposes the completed trajectory to the reward function | ||
| as one ``RewardRecord``. This is the v0.0.3 observability boundary; it defines |
There was a problem hiding this comment.
Use the actual release version for this contract
The reviewed tree declares version 0.0.6 in pyproject.toml, so labeling this as the v0.0.3 observability boundary—and repeating that label in the artifact section—misidentifies which release the documentation describes. This can lead users of older releases to expect the documented agentic types and fields; update both version references to the intended current release or explicitly identify this as a separately versioned contract.
Useful? React with 👍 / 👎.
| - Ordered normalized ``request``, ``assistant_text``, | ||
| ``assistant_tool_call``, ``tool_result``, ``finish``, and ``error`` | ||
| events. Events carry only fields relevant to their type. |
There was a problem hiding this comment.
Do not describe the current trace as event-ordered
For a multi-turn trajectory containing tool results, the current trace is not in conversation order: _append_sample_response concatenates each call's request/assistant/finish events, while _trace_with_tool_results inserts all tool-result events before the first finish. The resulting trace can place a result before its turn finishes and move later results into the first turn, so consumers relying on the documented ordering will misassociate events; either correct trace construction and assert order in the cited test, or document that messages is the ordering source.
Useful? React with 👍 / 👎.
|
Hi maintainers, the GitHub Actions checks for this fork PR are currently blocked with action_required, and no jobs have run. I’ve addressed the review feedback and verified locally (46 passed; Sphinx -W build succeeded). Could someone please approve and run the workflows when convenient? Thanks! @adohe |
What does this PR do?
Defines the agentic trajectory observability contract in the existing observability guide.
The documentation now:
AgentTrajectory/AgentTrajectoryTurnto the merged per-prompt/sampleRewardRecord;No runtime or artifact-writing behavior changes.
Related issue
Closes #49
Type of change
How was it tested?
python -m pytest tests/test_agentic_cpu.py -q(46 passed)python -m sphinx -W --keep-going -b html docs dist/docs-issue49(succeeded without warnings)CPU and documentation validation only; no GPU validation is required for this documentation-only change.
Checklist