-
Notifications
You must be signed in to change notification settings - Fork 2.3k
Add filtering for agentVisible: false messages on streaming providers #4847
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Conversation
DOsinga
left a comment
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
looking at this again, I think you want to do this on a lower level.
we have a number of format_messages(..) - this feels like the right place to put this, where we convert the messages to what the client should see, not on a per provider basis. we currently already filter on two places, one time I'm pretty sure after we already have, so this is just going to be a whack a mole game I think
what do you think?
| /// // Get messages visible ONLY to agent (not visible to user) | ||
| /// let agent_only = conversation.filtered_messages(|meta| meta.agent_visible && !meta.user_visible); | ||
| /// ``` | ||
| pub fn filtered_messages<F>(&self, filter: F) -> Vec<Message> |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
ah, this is not really what I meant - meant just pass it a visibility level for the messages and then return those
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
visibility level kind of needs some kind of mask though on MessageMetadata? Don't quite follow.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
i meant something like
pub fn visible_messages(for: User|Agent) but maybe that's not the API you imagined.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can do that but I was imagining adding more tags to message metadata like summary marker, notification so wanted to keep them synced.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
ok, some other time, let's just drop this for now. or keep it but remove the comment
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
ack. leaving for now but will follow up and use it + might update it a bit.
|
pushed some changes refactoring the provider side impl. I also noticed it did some filtering of thinking messages in anthropic provider. Unsure about those, seems to have been modified to match what google is doing. Could these be related to goose repeating itself? |
| // Convert messages to Anthropic format | ||
| for message in messages { | ||
| // Filter messages to only include agent_visible ones and convert to Anthropic format | ||
| for message in messages.iter().filter(|m| m.is_agent_visible()) { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
to be honest, if we go through the loop here anyway, we might as well skip them in place rather than prefilter
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Goose claims:
The current implementation with .iter().filter() is already optimal - it combines filtering and processing in a single, lazy iteration without any unnecessary allocations. There's no performance benefit to pre-filtering in this case.
| .iter() | ||
| .filter(|m| m.is_agent_visible()) | ||
| .cloned() | ||
| .collect(); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
shouldn't we move this inside get_messages_token_counts_async?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
done, nice call.
| /// // Get messages visible ONLY to agent (not visible to user) | ||
| /// let agent_only = conversation.filtered_messages(|meta| meta.agent_visible && !meta.user_visible); | ||
| /// ``` | ||
| pub fn filtered_messages<F>(&self, filter: F) -> Vec<Message> |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
ok, some other time, let's just drop this for now. or keep it but remove the comment
| })); | ||
| MessageContent::RedactedThinking(_) => { | ||
| // Skip redacted thinking - internal use only | ||
| continue; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'm not sure about this. it sounds reasonable, but do we want to make a change we don't really really understand for this if we want to cherry-pick?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Agreed, reverted but wanted to just call this out as somewhere to investigate.
| captured_messages: std::sync::Arc<std::sync::Mutex<Vec<Message>>>, | ||
| } | ||
|
|
||
| impl MockStreamingProvider { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
this feels very LLM generated. did you read the code? delete
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
gone
| self.complete_with_model(&model_config, system, messages, tools) | ||
| .await | ||
| } else { | ||
| Err(e) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
nice. there's also snowflake & anthropic that come with their own message formats
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
done
crates/goose/src/conversation/mod.rs
Outdated
| } | ||
|
|
||
| #[test] | ||
| fn test_filtered_messages() { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
these tests don't add anything
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
done
…-unification * 'main' of github.com:block/goose: (24 commits) feat(cli): add `path` & `limit` to `session list` command (#4878) Allow better concurrent access (#4896) fix: Windows prompt cursor positioning issue with ANSI escape sequences (#4464) Fix: LiteLLM API key field not showing in UI configuration (#4105) fix: path is duplicated on tool calls causing them to fail (#4658) (#4859) add new prompt to get all available tutorials (#4802) Add filtering for agentVisible: false messages on streaming providers (#4847) alexhancock/mcp-crate-cleanup (#4885) docs: rename sub-recipe to subrecipe (#4886) docs: new multi-model section with autopilot topic (#4864) make agent manager singleton (#4880) Cli web auth token (#4456) fix(token_counter): fix panic with GitHub Copilot (#4632) Revert "Internal MCP Crate Cleanup (#4800)" (#4883) remove 2 redundant comments and one that lies (#4866) Internal MCP Crate Cleanup (#4800) Fix #4612: Return non-zero exit code when CLI session ends with error (#4621) Dead code cleanup (#4873) fix: restoring test data and correcting name (#4875) Add .goosehints file to enforce lowercase branding in documentation (#4870) ...
* remove only-pr-labels (block#4842) Signed-off-by: Angela Ning <[email protected]> * Docs: Add link to Plug & Play video for Reddit MCP (block#4852) * Fix: Token count UI doesn't re-render if it's open. (block#4822) * Update databricks flash model (block#4836) * Session manager (block#4648) Co-authored-by: Douwe Osinga <[email protected]> * Add Hacktoberfest Guides (block#4830) Co-authored-by: taniashiba <[email protected]> * docs: goose x Hacktoberfest 2025 Blog (block#4855) Co-authored-by: Tania Chakraborty <[email protected]> Co-authored-by: Angie Jones <[email protected]> * fix: delete some flaky and not-useful tests (block#4861) * can tell the system what shell it is using (block#4807) * new subrecipe blog post banner (block#4862) * docs: remove recipe generator link from next to extension search (block#4858) * lowercase g in goose (block#4832) * doc: file parameter recipe update (block#4594) * Fiie input recipe ref doc (block#4869) * Add .goosehints file to enforce lowercase branding in documentation (block#4870) Co-authored-by: Angie Jones <[email protected]> * fix: restoring test data and correcting name (block#4875) * Dead code cleanup (block#4873) Co-authored-by: Douwe Osinga <[email protected]> Co-authored-by: Michael Neale <[email protected]> * Fix block#4612: Return non-zero exit code when CLI session ends with error (block#4621) Signed-off-by: jalateras <[email protected]> * Internal MCP Crate Cleanup (block#4800) * remove 2 redundant comments and one that lies (block#4866) Co-authored-by: Douwe Osinga <[email protected]> * Revert "Internal MCP Crate Cleanup (block#4800)" (block#4883) * fix(token_counter): fix panic with GitHub Copilot (block#4632) Signed-off-by: sings-to-bees-on-wednesdays <[email protected]> main was broken. this seems important * Cli web auth token (block#4456) Signed-off-by: Evanfeenstra <[email protected]> * make agent manager singleton (block#4880) * docs: new multi-model section with autopilot topic (block#4864) * docs: rename sub-recipe to subrecipe (block#4886) * alexhancock/mcp-crate-cleanup (block#4885) * Temporary workaround for mcp server * Add filtering for agentVisible: false messages on streaming providers (block#4847) * add new prompt to get all available tutorials (block#4802) Signed-off-by: AdemolaAri <[email protected]> * Fix for auth in extension, fix for stale keychain * fix: path is duplicated on tool calls causing them to fail (block#4658) (block#4859) Signed-off-by: demetrio108 <[email protected]> * Fix: LiteLLM API key field not showing in UI configuration (block#4105) Signed-off-by: jalateras <[email protected]> Co-authored-by: Ebony Louis <[email protected]> Co-authored-by: Claude <[email protected]> Co-authored-by: Jack Amadeo <[email protected]> * fix: Windows prompt cursor positioning issue with ANSI escape sequences (block#4464) Signed-off-by: Matt Donovan <[email protected]> Co-authored-by: Matt Donovan <[email protected]> * Allow better concurrent access (block#4896) Co-authored-by: Douwe Osinga <[email protected]> * feat(cli): add `path` & `limit` to `session list` command (block#4878) * Added CMD+T keyboard shortcut that takes you to the Home tab (block#4541) Signed-off-by: Guillaume Simard <[email protected]> --------- Signed-off-by: Angela Ning <[email protected]> Signed-off-by: jalateras <[email protected]> Signed-off-by: Evanfeenstra <[email protected]> Signed-off-by: AdemolaAri <[email protected]> Signed-off-by: demetrio108 <[email protected]> Signed-off-by: Matt Donovan <[email protected]> Signed-off-by: Guillaume Simard <[email protected]> Co-authored-by: Angela Ning <[email protected]> Co-authored-by: Emma Youndtsmith <[email protected]> Co-authored-by: David Katz <[email protected]> Co-authored-by: Douwe Osinga <[email protected]> Co-authored-by: Douwe Osinga <[email protected]> Co-authored-by: Ebony Louis <[email protected]> Co-authored-by: taniashiba <[email protected]> Co-authored-by: taniandjerry <[email protected]> Co-authored-by: Tania Chakraborty <[email protected]> Co-authored-by: Angie Jones <[email protected]> Co-authored-by: Jack Amadeo <[email protected]> Co-authored-by: Michael Neale <[email protected]> Co-authored-by: w. ian douglas <[email protected]> Co-authored-by: Alex Hancock <[email protected]> Co-authored-by: Jarrod Sibbison <[email protected]> Co-authored-by: Rizel Scarlett <[email protected]> Co-authored-by: Jim Alateras <[email protected]> Co-authored-by: sings-to-bees-on-wednesdays <[email protected]> Co-authored-by: Evan Feenstra <[email protected]> Co-authored-by: Yingjie He <[email protected]> Co-authored-by: dianed-square <[email protected]> Co-authored-by: Ademola Arigbabuwo <[email protected]> Co-authored-by: Demetrio ⚡️ <[email protected]> Co-authored-by: Claude <[email protected]> Co-authored-by: Jack Amadeo <[email protected]> Co-authored-by: Matt Donovan <[email protected]> Co-authored-by: Matt Donovan <[email protected]> Co-authored-by: Robert Mcgregor <[email protected]> Co-authored-by: Guillaume Simard <[email protected]>
* main: (206 commits) Tiny: fix github casing (block#4903) remove anyOf from create_task tool (block#4897) chore(deps): bump tracing-subscriber from 0.3.19 to 0.3.20 (block#4442) fix optional recipe schema zod validation (block#4900) Added CMD+T keyboard shortcut that takes you to the Home tab (block#4541) feat(cli): add `path` & `limit` to `session list` command (block#4878) Allow better concurrent access (block#4896) fix: Windows prompt cursor positioning issue with ANSI escape sequences (block#4464) Fix: LiteLLM API key field not showing in UI configuration (block#4105) fix: path is duplicated on tool calls causing them to fail (block#4658) (block#4859) add new prompt to get all available tutorials (block#4802) Add filtering for agentVisible: false messages on streaming providers (block#4847) alexhancock/mcp-crate-cleanup (block#4885) docs: rename sub-recipe to subrecipe (block#4886) docs: new multi-model section with autopilot topic (block#4864) make agent manager singleton (block#4880) Cli web auth token (block#4456) fix(token_counter): fix panic with GitHub Copilot (block#4632) Revert "Internal MCP Crate Cleanup (block#4800)" (block#4883) remove 2 redundant comments and one that lies (block#4866) ...
…block#4847) Signed-off-by: HikaruEgashira <[email protected]>
This is the root cause of the input too long issues we're seeing.
I also saw some quirks with auto-compact triggering right after compact; this is from the provider metadata but I updated the token estimation approach to also filter.
perform_compaction doesn't need filtering since it falls thru to complete_fast().