Repository navigation
Add emit_message_history to record an agent transcript without seeding it - #74811
pankajastro wants to merge 2 commits into
Conversation
…g it Single-turn agents that want their transcript on XCom had to pass message_history=[], a parameter for seeding a conversation, not for observing a run. A dedicated flag states that intent directly, and do_xcom_push=False now also suppresses the transcript push, as it does the other XCom pushes. Co-Authored-By: Claude <noreply@anthropic.com>
|
|
||
| def _emit_message_history(self, context: Context, result: Any) -> None: | ||
| """Push the full post-run transcript to XCom for the next turn to resume.""" | ||
| if not self.do_xcom_push: |
There was a problem hiding this comment.
This gate also stops the push for message_history users, not only the new flag. In 1.0.0rc1 _emit_message_history pushed unconditionally, and stability.rst lists message_history as stable, where narrowing what a feature does needs a major release. A session Dag that sets do_xcom_push=False to keep a big answer out of XCom now pulls nothing on the next turn: with default='[]' every run silently starts fresh, and the store pattern in example_agent.py (no default) saves "None", which fails to parse on the turn after.
Could the gate apply only to the new flag, e.g. if self.message_history is None and not self.do_xcom_push: return? A test that message_history=[] still pushes with do_xcom_push=False would pin it; nothing covers that case today.
There was a problem hiding this comment.
Good catch, and agreed: the gate narrowed a stable feature. I removed it in 39a4c59.
_emit_message_history now pushes unconditionally again, so a seeded run behaves as in
1.0.0rc1, including with do_xcom_push=False.
Drafted-by: Claude Code (Sonnet 5.5); reviewed by @pankajastro before posting
| durable: bool = False, | ||
| cache_prompt: bool = True, | ||
| message_history: list[ModelMessage] | str | bytes | None = None, | ||
| emit_message_history: bool = False, |
There was a problem hiding this comment.
emit_message_history=False doesn't stop the push when message_history is set, so False here doesn't really mean "don't emit". Would bool | None = None, resolving to message_history is not None, read better? True opts a single-turn run in, False lets a seeded run opt out, and the transcript gets its own switch instead of riding on do_xcom_push.
There was a problem hiding this comment.
Agreed, done in 39a4c59: emit_message_history is now bool | None = None, resolved in
__init__ to message_history is not None when unset. True opts a single-turn run in,
False opts a seeded run out, and the transcript no longer depends on do_xcom_push.
test_transcript_emitted_to_xcom covers the seeded, opt-in and opt-out cases, and only
message_history seeds the run.
Drafted-by: Claude Code (Sonnet 5.5); reviewed by @pankajastro before posting
| .. note:: | ||
|
|
||
| ``message_history`` cannot be combined with ``enable_hitl_review`` -- the | ||
| ``message_history`` and ``emit_message_history`` cannot be combined with |
There was a problem hiding this comment.
stability.rst and troubleshooting.rst need a line each for this. The stable row names only message_history, so is emit_message_history stable or experimental? And the troubleshooting entry for message_history and enable_hitl_review=True cannot be used together doesn't cover the new emit_message_history error.
There was a problem hiding this comment.
Added both, in 39a4c59. stability.rst has a row for emit_message_history, marked stable
like message_history (it is additive and the default keeps today's behaviour). The
troubleshooting.rst entry now lists the emit_message_history and enable_hitl_review=True
error alongside the message_history one.
Drafted-by: Claude Code (Sonnet 5.5); reviewed by @pankajastro before posting
Gating the transcript push on do_xcom_push also stopped it for message_history users, narrowing a stable feature. A tri-state emit_message_history defaults to whether the run is seeded, so existing sessions keep pushing, a single-turn run can opt in, and a seeded run can opt out. Co-Authored-By: Claude <noreply@anthropic.com>
This branch has not been deployed
2 participants
AgentOperatorpushes the transcript to XCom only whenmessage_history is not None.A single-turn agent that only wants its transcript recorded has to pass
message_history=[],which is the parameter for seeding a conversation.
This adds
emit_message_history: bool | None = None:None(default): push the transcript whenmessage_historyis set, as before.True: also push it for a run withoutmessage_history; the run is not seeded.False: keep a seeded run from pushing it.emit_message_historyresolving toTruetogether withenable_hitl_reviewraises, for thesame reason
message_historydoes: the post-review transcript is not recoverable.Was generative AI tooling used to co-author this PR?
Generated-by: Claude Code (Sonnet 5.5) following the guidelines
🤖 Generated with Claude Code