Strip the audience for compacting by DOsinga · Pull Request #6646 · block/goose

DOsinga · 2026-01-22T20:25:09Z

Summary

We weren't stripping audience = user tags.

my conversation to be compacted came in at 229,887 before, now it is 128,516 - to be honest seems like almost too steep a change

Copilot

Pull request overview

This PR implements audience-based content filtering for conversation compaction. When compacting conversations, content tagged for specific audiences (e.g., user-only) is now filtered out before sending to the LLM, reducing token usage by approximately 44% (from 229,887 to 128,516 tokens in the test case).

Changes:

Implements content filtering by audience role for compaction
Removes verbose LLM request debug logging
Optimizes compaction formatting by using compact JSON and simplified thinking representation
Increases session import body limit to 25MB

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
crates/goose/src/conversation/message.rs	Adds `filter_for_audience()` and `agent_visible_content()` methods to filter message content by audience role
crates/goose/src/context_mgmt/mod.rs	Updates compaction to use audience filtering, changes to compact JSON formatting, simplifies thinking representation
crates/goose/src/providers/api_client.rs	Removes debug logging of LLM request payloads
crates/goose-server/src/routes/session.rs	Increases body limit for session import endpoint to 25MB

Copilot · 2026-01-22T20:30:01Z

crates/goose/src/conversation/message.rs

+    pub fn filter_for_audience(&self, audience: Role) -> Option<MessageContent> {
+        match self {
+            MessageContent::Text(text) => {
+                if text.audience().map(|roles| roles.contains(&audience)).unwrap_or(true) {
+                    Some(self.clone())
+                } else {
+                    None
+                }
+            }
+            MessageContent::Image(img) => {
+                if img.audience().map(|roles| roles.contains(&audience)).unwrap_or(true) {
+                    Some(self.clone())
+                } else {
+                    None
+                }
+            }
+            MessageContent::ToolResponse(res) => {
+                let Ok(result) = &res.tool_result else {
+                    return Some(self.clone());
+                };
+
+                let filtered_content: Vec<Content> = result
+                    .content
+                    .iter()
+                    .filter(|c| {
+                        c.audience()
+                            .map(|roles| roles.contains(&audience))
+                            .unwrap_or(true)
+                    })
+                    .cloned()
+                    .collect();
+
+                if filtered_content.is_empty() {
+                    return None;
+                }
+
+                Some(MessageContent::ToolResponse(ToolResponse {
+                    id: res.id.clone(),
+                    tool_result: Ok(CallToolResult {
+                        content: filtered_content,
+                        ..result.clone()
+                    }),
+                    metadata: res.metadata.clone(),
+                }))
+            }
+            MessageContent::Thinking(_) | MessageContent::RedactedThinking(_) => None,
+            _ => Some(self.clone()),
+        }
+    }


The new public methods filter_for_audience and agent_visible_content lack test coverage. Consider adding tests to verify the filtering logic, especially for edge cases like empty content, tool responses with mixed audience tags, and content with no audience specified.

katzdave

Nice find. Seems like lots of the provider format code is doing the audience stripping which explains the difference in the compaction vs llm call context.

Maybe we should follow up and make this more general and push this filter_for_audience more generically into the flow of provider.complete, and get rid of the provider format specific filters?

DOsinga · 2026-01-22T20:46:12Z

I think the format code is actually wrong; the LLMs are supposed to ignore this already so it should all be good to send them this; just dont want to compact it.

* main: docs: ml-based prompt injection detection (#6627) Strip the audience for compacting (#6646) chore(release): release version 1.21.0 (minor) (#6634) add collapsable chat nav (#6649) fix: capitalize Rust in CONTRIBUTING.md (#6640) chore(deps): bump lodash from 4.17.21 to 4.17.23 in /ui/desktop (#6623) Vibe mcp apps (#6569) Add session forking capability (#5882) chore(deps): bump lodash from 4.17.21 to 4.17.23 in /documentation (#6624) fix(docs): use named import for globby v13 (#6639) PR Code Review (#6043) fix(docs): use dynamic import for globby ESM module (#6636) chore: trigger CI Document tab completion (#6635) Install goose-mcp crate dependencies (#6632) feat(goose): standardize agent-session-id for session correlation (#6626)

Co-authored-by: Douwe Osinga <douwe@squareup.com> Signed-off-by: fbalicchia <fbalicchia@cuebiq.com>

* origin/main: Fix GCP Vertex AI global endpoint support for Gemini 3 models (#6187) fix: macOS keychain infinite prompt loop (#6620) chore: reduce duplicate or unused cargo deps (#6630) feat: codex subscription support (#6600) smoke test allow pass for flaky providers (#6638) feat: Add built-in skill for goose documentation reference (#6534) Native images (#6619) docs: ml-based prompt injection detection (#6627) Strip the audience for compacting (#6646) chore(release): release version 1.21.0 (minor) (#6634) add collapsable chat nav (#6649) fix: capitalize Rust in CONTRIBUTING.md (#6640) chore(deps): bump lodash from 4.17.21 to 4.17.23 in /ui/desktop (#6623) Vibe mcp apps (#6569) Add session forking capability (#5882) chore(deps): bump lodash from 4.17.21 to 4.17.23 in /documentation (#6624) fix(docs): use named import for globby v13 (#6639) PR Code Review (#6043) fix(docs): use dynamic import for globby ESM module (#6636) # Conflicts: # Cargo.lock # crates/goose-server/src/routes/session.rs

…o dkatz/canonical-context * 'dkatz/canonical-provider' of github.com:block/goose: (27 commits) docs: add Remotion video creation tutorial (#6675) docs: export recipe and copy yaml (#6680) Test against fastmcp (#6666) docs: mid-session changes (#6672) Fix MCP elicitation deadlock and improve UX (#6650) chore: upgrade to rmcp 0.14.0 (#6674) [docs] add MCP-UI to MCP Apps blog (#6664) ACP get working dir from args.cwd (#6653) Optimise load config in UI (#6662) Fix GCP Vertex AI global endpoint support for Gemini 3 models (#6187) fix: macOS keychain infinite prompt loop (#6620) chore: reduce duplicate or unused cargo deps (#6630) feat: codex subscription support (#6600) smoke test allow pass for flaky providers (#6638) feat: Add built-in skill for goose documentation reference (#6534) Native images (#6619) docs: ml-based prompt injection detection (#6627) Strip the audience for compacting (#6646) chore(release): release version 1.21.0 (minor) (#6634) add collapsable chat nav (#6649) ...

Strip the audience for compacting

12f4ab4

Copilot AI review requested due to automatic review settings January 22, 2026 20:25

Copilot started reviewing on behalf of DOsinga January 22, 2026 20:25 View session

DOsinga requested a review from katzdave January 22, 2026 20:25

Copilot AI reviewed Jan 22, 2026

View reviewed changes

that's a wrap folks

d3454aa

katzdave approved these changes Jan 22, 2026

View reviewed changes

DOsinga merged commit c57c256 into main Jan 22, 2026
19 checks passed

DOsinga deleted the strip-audience-for-compact branch January 22, 2026 21:13

fbalicchia pushed a commit to fbalicchia/goose that referenced this pull request Jan 23, 2026

Strip the audience for compacting (block#6646)

06c52e7

Co-authored-by: Douwe Osinga <douwe@squareup.com> Signed-off-by: fbalicchia <fbalicchia@cuebiq.com>

This was referenced Jan 27, 2026

chore(release): release version 1.22.0 (minor) #6725

Closed

chore(release): release version 1.22.0 (minor) salmanmkc/goose#1

Open

chore(release): release version 1.22.0 (minor) #6812

Closed

chore(release): release version 1.22.0 (minor) #6813

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Strip the audience for compacting#6646

Strip the audience for compacting#6646
DOsinga merged 2 commits intomainfrom
strip-audience-for-compact

DOsinga commented Jan 22, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Jan 22, 2026

Uh oh!

katzdave left a comment •

edited

Loading

Uh oh!

DOsinga commented Jan 22, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

DOsinga commented Jan 22, 2026

Summary

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Jan 22, 2026

Choose a reason for hiding this comment

Uh oh!

katzdave left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

DOsinga commented Jan 22, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

katzdave left a comment •

edited

Loading