[PR #93] [MERGED] feat: patch refiner #153

New Issue

2026-06-06T22:09:26-04:00

yindo commented

2026-06-06 22:09:26 -04:00

📋 Pull Request Information

Original PR: https://github.com/vxcontrol/pentagi/pull/93
Author: @zavgorodnii
Created: 12/17/2025
Status: ✅ Merged
Merged: 12/29/2025
Merged by: @asdek

Base: feat/graphiti-search ← Head: feat/patch-refiner

📝 Commits (1)

da9019b feat: added patch refiner

📊 Changes

12 files changed (+1129 additions, -22 deletions)

View changed files

📝 backend/pkg/providers/openai/config.yml (+2 -2)
📝 backend/pkg/providers/performers.go (+47 -12)
📝 backend/pkg/providers/provider.go (+7 -1)
➕ backend/pkg/providers/subtask_patch.go (+249 -0)
➕ backend/pkg/providers/subtask_patch_test.go (+679 -0)
📝 backend/pkg/templates/prompts/refiner.tmpl (+38 -5)
📝 backend/pkg/templates/templates.go (+1 -1)
📝 backend/pkg/templates/validator/testdata.go (+1 -0)
📝 backend/pkg/tools/args.go (+26 -0)
📝 backend/pkg/tools/registry.go (+9 -0)
📝 backend/pkg/tools/tools.go (+69 -0)
📝 docker-compose.yml (+1 -1)

📄 Description

Description of the Change

Problem

The current subtask refiner regenerates the entire subtask list on each refinement cycle, which is token inefficient - Requires the LLM to output all subtasks even when only minor changes are needed.

Solution

Implement a delta-based patch system for subtask refinement that allows the refiner to submit only the changes needed rather than regenerating the complete list.

Key changes:

Introduced SubtaskPatch with four operation types: add, remove, modify, and reorder
Each operation targets specific subtasks by ID for precise modifications
Added applySubtaskOperations() function to apply patch operations to the existing subtask list
Updated refiner prompt to instruct the LLM to use delta operations
Created dedicated GetRefinerExecutor() with the new subtask_patch tool
Added comprehensive validation and error handling for patch operations
Increased refiner max_tokens from 8192 to 16384 for complex refinement scenarios

Additional optimizations:

Changed default Graphiti model from gpt-5 to gpt-5-mini

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/vxcontrol/pentagi/pull/93 **Author:** [@zavgorodnii](https://github.com/zavgorodnii) **Created:** 12/17/2025 **Status:** ✅ Merged **Merged:** 12/29/2025 **Merged by:** [@asdek](https://github.com/asdek) **Base:** `feat/graphiti-search` ← **Head:** `feat/patch-refiner` --- ### 📝 Commits (1) - [`da9019b`](https://github.com/vxcontrol/pentagi/commit/da9019bdf3c651d83b24faa49510b170d4b251d2) feat: added patch refiner ### 📊 Changes **12 files changed** (+1129 additions, -22 deletions) <details> <summary>View changed files</summary> 📝 `backend/pkg/providers/openai/config.yml` (+2 -2) 📝 `backend/pkg/providers/performers.go` (+47 -12) 📝 `backend/pkg/providers/provider.go` (+7 -1) ➕ `backend/pkg/providers/subtask_patch.go` (+249 -0) ➕ `backend/pkg/providers/subtask_patch_test.go` (+679 -0) 📝 `backend/pkg/templates/prompts/refiner.tmpl` (+38 -5) 📝 `backend/pkg/templates/templates.go` (+1 -1) 📝 `backend/pkg/templates/validator/testdata.go` (+1 -0) 📝 `backend/pkg/tools/args.go` (+26 -0) 📝 `backend/pkg/tools/registry.go` (+9 -0) 📝 `backend/pkg/tools/tools.go` (+69 -0) 📝 `docker-compose.yml` (+1 -1) </details> ### 📄 Description ### Description of the Change #### Problem The current subtask refiner regenerates the entire subtask list on each refinement cycle, which is **token inefficient** - Requires the LLM to output all subtasks even when only minor changes are needed. #### Solution Implement a **delta-based patch system** for subtask refinement that allows the refiner to submit only the changes needed rather than regenerating the complete list. **Key changes:** - Introduced `SubtaskPatch` with four operation types: `add`, `remove`, `modify`, and `reorder` - Each operation targets specific subtasks by ID for precise modifications - Added `applySubtaskOperations()` function to apply patch operations to the existing subtask list - Updated refiner prompt to instruct the LLM to use delta operations - Created dedicated `GetRefinerExecutor()` with the new `subtask_patch` tool - Added comprehensive validation and error handling for patch operations - Increased refiner max_tokens from 8192 to 16384 for complex refinement scenarios **Additional optimizations:** - Changed default Graphiti model from `gpt-5` to `gpt-5-mini` --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

yindo added the pull-request label 2026-06-06 22:09:26 -04:00

yindo closed this issue

2026-06-06 22:09:26 -04:00

yindo referenced this issue

2026-06-06 22:09:39 -04:00

[PR #153] [MERGED] test: add unit tests for perplexity and google search tools #195

yindo referenced this issue

2026-06-06 22:09:41 -04:00

[PR #165] [CLOSED] fix: apply constructed opts to Google search service creation #206

yindo referenced this issue

2026-06-06 22:09:45 -04:00

[PR #185] [MERGED] Multi-Provider Expansion & Testing Infrastructure #221