perf(agent): prewarm session integrations before first turn#2454
Merged
M3gA-Mind merged 8 commits intoMay 22, 2026
Conversation
Add a task-local depth guard around run_subagent so runtime delegation cannot recurse beyond the documented harness limit. Co-authored-by: Cursor <cursoragent@cursor.com>
Prevents usize wrap-around that could bypass MAX_SPAWN_DEPTH guard in release builds. Addresses CodeRabbit review on PR tinyhumansai#2234.
The new with_spawn_depth task-local scope adds a second stacked task_local wrapper around run_typed_mode. Under cargo-llvm-cov instrumentation, the combined future state machine overflows tokio's default 2 MiB per-thread test stack in turn_dispatches_spawn_subagent_through_full_path. Box::pin the inner future so the large state machine lives on the heap.
Contributor
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (10)
✅ Files skipped from review due to trivial changes (2)
📝 WalkthroughWalkthroughAdds a Tokio task-local spawn-depth budget (MAX_SPAWN_DEPTH = 3) and enforcement that returns SubagentRunError::SpawnDepthExceeded when exceeded; threads a best-effort Composio integration prewarm through AgentBuilder, adds session flags/config to ensure integrations are initialized once per session, and updates related types and tests. ChangesSpawn-depth limiting and integration prewarming
Estimated code review effort🎯 3 (Moderate) | ⏱️ ~25 minutes Possibly related PRs
Suggested reviewers
Poem
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches⚔️ Resolve merge conflicts
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
coderabbitai
Bot
added
agent
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 1
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
src/openhuman/agent/harness/session/builder.rs (1)
1305-1315:⚠️ Potential issue | 🟠 Major | ⚡ Quick winTrack only actually inserted synthesized tool names.
synthesized_tool_namesis currently derived before collision filtering. If a synthesized name collides with an existing direct tool, it won’t be inserted now, but it will still be marked synthesized and can be removed on the next refresh.Suggested fix
- let synthesized_tool_names: std::collections::HashSet<String> = delegation_tools - .iter() - .map(|t| t.name().to_string()) - .collect(); let existing_names: std::collections::HashSet<String> = tools.iter().map(|t| t.name().to_string()).collect(); - tools.extend( - delegation_tools - .into_iter() - .filter(|t| !existing_names.contains(t.name())), - ); + let filtered_delegation_tools: Vec<Box<dyn Tool>> = delegation_tools + .into_iter() + .filter(|t| !existing_names.contains(t.name())) + .collect(); + let synthesized_tool_names: std::collections::HashSet<String> = filtered_delegation_tools + .iter() + .map(|t| t.name().to_string()) + .collect(); + tools.extend(filtered_delegation_tools);🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@src/openhuman/agent/harness/session/builder.rs` around lines 1305 - 1315, synthesized_tool_names is computed from delegation_tools before filtering collisions, so names that were not actually inserted (because they collided with existing tools) are still marked synthesized; update the logic to compute synthesized_tool_names only from the delegation_tools that are actually inserted into tools — for example, apply the same filter used in tools.extend (the closure that checks !existing_names.contains(t.name())) and collect the names from that filtered set (or collect inserted tools into a temporary vec and derive synthesized_tool_names from it) so synthesized_tool_names reflects only truly added tools.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@src/openhuman/agent/harness/session/builder.rs`:
- Around line 1523-1524: The first issue is a use-after-move on
prewarmed_integrations: check prewarmed_integrations.is_some() first (assign
connected_integrations_initialized = prewarmed_integrations.is_some()), then
assign agent.connected_integrations = prewarmed_integrations.unwrap_or_default()
(or clone prior to unwrap) so you don't read after move; update references to
prewarmed_integrations, connected_integrations, and
connected_integrations_initialized accordingly. The second issue is that
self.synthesized_tool_names is populated from delegation_tools before removing
name collisions, causing refresh_delegation_tools to drop direct tools; fix by
first filtering delegation_tools to remove any whose names collide with existing
direct tools, extend tools with that filtered list, and then set
self.synthesized_tool_names to the names of the actually inserted delegation
tools (use the filtered collection) so refresh_delegation_tools uses the correct
drop mask (affecting synthesized_tool_names, delegation_tools, tools, and
refresh_delegation_tools).
---
Outside diff comments:
In `@src/openhuman/agent/harness/session/builder.rs`:
- Around line 1305-1315: synthesized_tool_names is computed from
delegation_tools before filtering collisions, so names that were not actually
inserted (because they collided with existing tools) are still marked
synthesized; update the logic to compute synthesized_tool_names only from the
delegation_tools that are actually inserted into tools — for example, apply the
same filter used in tools.extend (the closure that checks
!existing_names.contains(t.name())) and collect the names from that filtered set
(or collect inserted tools into a temporary vec and derive
synthesized_tool_names from it) so synthesized_tool_names reflects only truly
added tools.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro
Run ID: dedc5c1f-db40-4392-b7d7-c1d181373745
📒 Files selected for processing (12)
gitbooks/developing/architecture/agent-harness.mdsrc/openhuman/agent/harness/harness_gap_tests.rssrc/openhuman/agent/harness/mod.rssrc/openhuman/agent/harness/session/builder.rssrc/openhuman/agent/harness/session/runtime.rssrc/openhuman/agent/harness/session/tests.rssrc/openhuman/agent/harness/session/turn.rssrc/openhuman/agent/harness/session/types.rssrc/openhuman/agent/harness/spawn_depth_context.rssrc/openhuman/agent/harness/subagent_runner/ops.rssrc/openhuman/agent/harness/subagent_runner/ops_tests.rssrc/openhuman/agent/harness/subagent_runner/types.rs
graycyrus
reviewed
May 21, 2026
Contributor
graycyrus
left a comment
graycyrus
left a comment
There was a problem hiding this comment.
LGTM — clean, well-structured performance optimisation.
Walkthrough: this PR prewarms session connected_integrations from the shared Composio cache during from_config_* agent construction, synthesises delegation tools against the prewarmed snapshot, and carries the runtime Config on the session agent to avoid Config::load_or_init() on the hot path. Cold-cache sessions preserve fallback behaviour.
| Area | Files | Key changes |
|---|---|---|
| Rust core / session | builder.rs |
Prewarm integrations from cache, build delegation tools against cached slice, track synthesized_tool_names post-collision-filter |
| Rust core / session | turn.rs |
Gate turn-1 fetch behind connected_integrations_initialized, use stored config for mid-session cache probes |
| Rust core / session | types.rs |
New fields: connected_integrations_initialized, integration_runtime_config |
| Rust core / session | runtime.rs |
set_connected_integrations now sets initialized flag + updates hash |
| Tests | tests.rs |
Regression test for initialized/hash bookkeeping |
CodeRabbit findings (already addressed): use-after-move on prewarmed_integrations ✅, synthesized tool mask alignment ✅ — both confirmed fixed in current diff.
Additional observations:
- Fallback paths are well-preserved: cold-cache → empty slice → turn-1 fetch → delegation rebuild. No correctness regression.
config.clone()for the snapshot is a one-time construction cost, acceptable for the hot-path savings.- Good test coverage for the bookkeeping; the prewarm path itself is integration-level (requires warm cache) which is reasonable.
No new findings beyond what CodeRabbit already covered. Moving to approval queue.
M3gA-Mind
approved these changes
May 22, 2026
Contributor
|
nice work @srikaanthh, prewarming the session integrations so turn-1 skips the fetch and delegation rebuild is a really clean win 🔥 always love seeing perf improvements that cut latency on the hot path like this. glad to have it merged, keep it coming! |
This was referenced May 23, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
connected_integrationsfrom the shared Composio cache duringfrom_config_*agent constructiondelegate_<toolkit>surfaceConfigsnapshot on the session agent so mid-session integration-cache probes stop reloading config on the hot pathProblem
Agent::turn()before the first provider call.Configinside the hot path, and the session builder always synthesized delegation tools against an empty integration set, guaranteeing a repair pass on turn 1.Solution
composio::cached_active_integrations(config)during session construction to prewarmconnected_integrationswhen the shared cache is already warm.&[], then persist the synthesized-tool name set onto the builtAgent.connected_integrations_initialized; turn 1 now only fetches integrations and refreshes delegation tools when the builder could not prewarm the cache.Configsnapshot on the session agent so mid-session cache reads and fallback integration fetches do not callConfig::load_or_init()on the hot path.Arcreconciliation failures so correctness stays unchanged when prewarming is unavailable.Submission Checklist
pnpmunavailable on PATH, focused Rust tests blocked by missingcmake).N/A: behaviour-only change## Relateddocs/RELEASE-MANUAL-SMOKE.md)Closes #NNNin the## RelatedsectionImpact
Config::load_or_init()on subsequent cache probes.Related
AI Authored PR Metadata (required for Codex/Linear PRs)
Linear Issue
Commit & Branch
Validation Run
pnpmon PATH, so this command could not be run here.pnpmon PATH, so this command could not be run here.whisper-rs-sysrequirescmake, which is not installed in this environment.cargo fmt --manifest-path Cargo.tomlpassed;git diff --check origin/main...HEADclean.cargo check --manifest-path app/src-tauri/Cargo.tomlattempt was also blocked because the vendoredtauri-cefdependency tree is missing in this environment.Validation Blocked
command:GGML_NATIVE=OFF cargo test --manifest-path Cargo.toml set_connected_integrations_marks_session_initialized_and_updates_hash -- --nocaptureandGGML_NATIVE=OFF cargo test --manifest-path Cargo.toml turn_without_tools_returns_text -- --nocaptureerror:whisper-rs-sysbuild script failed becausecmakeis not installed in the local environmentimpact:focused Rust tests did not complete locally; correctness is based on source review plus the added regression coverageBehavior Changes
Parity Contract
Arcreconciliation fallback, mid-session cache-driven refresh, and config-load fallback behavior remain intactDuplicate / Superseded PR Handling
Summary by CodeRabbit
New Features
Tests
Documentation