release: v6.8.0 - #49
Conversation
- search_files: ripgrep-first with FFF fallback, one-shot bounded results (default max_results 100); cursor pagination removed from the model-facing schema and output - agent loop: leading read-only tool calls execute concurrently (max 4 workers) with results committed in model order - malformed tool-call JSON returns a corrective tool result with the raw arguments so the model can re-issue the call - strict-mode schema tightening across native tool schemas (required-with-nullable optionals, execute_command safety classification guidance) - system prompt search_files guidance updated for one-shot behavior
![matterai-app[bot]](https://avatars.githubusercontent.com/in/474549?s=60&v=4){kind=small}
There was a problem hiding this comment.
🧪 PR Review is completed: Release PR flipping search_files to a ripgrep-first one-shot model, adding parallel execution of leading read-only tool calls, and tightening tool schemas for strict mode. One real robustness gap in the new parallel batch error path; the rest of the changes look consistent.
Reviewed src/tools/executors/searchFiles.ts: no issues found.
Reviewed src/tools/executors/searchFiles/format.ts: no issues found.
Reviewed src/tools/executors/searchFiles/types.ts: no issues found.
Reviewed src/tools/schemas/search_files.ts: no issues found.
Reviewed src/tools/schemas/execute_command.ts: no issues found.
Reviewed src/prompts/system.ts: no issues found.
Reviewed test/search-files.test.ts: no issues found.
Skipped files
CHANGELOG.md: Skipped file pattern
⬇️ Low Priority Suggestions (1)
src/core/agent.ts (1 suggestion)
Location:
src/core/agent.ts(Lines 1181-1184)🟡 Concurrency
Issue:
handleToolCallis not just tool execution — it also runs PreToolUse/PostToolUse hooks,observeCommittedCode(), and emitstool-start/tool-endevents. Running up to 4 of these concurrently means user-defined hooks now fire in parallel and interleave (a hook that mutates state or rewrites inputs is not guaranteed to be concurrency-safe), and UI events for the batch arrive out of model order. This may be acceptable given the read-only restriction, but it's a behavioral change worth confirming the hook runner and TUI event consumers tolerate.Fix: No code change required if hooks and event consumers are concurrency-safe; otherwise serialize hook execution or document the concurrency guarantee for hook authors.
Impact: Avoids subtle races in user-defined hooks and out-of-order UI event rendering.
- if (batchEnd > 1) { - const batch = toolCalls.slice(0, batchEnd) - const results = new Array<string>(batch.length) - let nextIndex = 0 +
| while (nextIndex < batch.length) { | ||
| const index = nextIndex++ | ||
| results[index] = await this.handleToolCall(batch[index]) | ||
| } |
There was a problem hiding this comment.
🟠 Error Handling
Issue: In the new parallel batch, results are buffered in results[] and only pushed to this.messages after Promise.all resolves. If handleToolCall rejects for any one tool (e.g. an AbortError from user interrupt, or a throwing PreToolUse hook), Promise.all rejects immediately: the tool results that already completed are discarded and no role: "tool" messages are pushed for the batch. The assistant message with tool_calls is already in this.messages, so the next model request violates the tool_call/tool_result pairing contract and the API call will fail. The serialized path pushed each result as it completed, so this is a regression introduced by the parallel path.
Fix: Catch per-tool errors inside each worker and store the error text as that tool's result, so every tool_call_id always gets a tool response and completed results survive a sibling failure.
Impact: Keeps the OpenAI message contract intact on aborts/hook failures and prevents losing already-completed parallel tool results.
| while (nextIndex < batch.length) { | |
| const index = nextIndex++ | |
| results[index] = await this.handleToolCall(batch[index]) | |
| } | |
| while (nextIndex < batch.length) { | |
| const index = nextIndex++ | |
| try { | |
| results[index] = await this.handleToolCall(batch[index]) | |
| } catch (error) { | |
| results[index] = `Tool execution failed: ${error instanceof Error ? error.message : String(error)}` | |
| } | |
| } |
- Sync the model catalog from /v1/models at startup and on usage refresh; register new backend models and prune retired ones (empty/failed fetches never wipe the offline fallback, default model is never pruned) - Capture the catalog's iconUrl and costMultiplier on each model - Static fallback updated to the current 7-model OSS catalog (muse-spark-1.3, gpt-5.6-luna, gpt-5.6-sol, gemini-3.8-flash added) - Model picker shows provider badges ([Z.ai], [Meta], [DeepSeek], [OpenAI], [Google]) — the TUI stand-in for the webapp's provider logos - orbcode usage command + per-model usage in /status and /usage - README/CHANGELOG updated
Summary By MatterAI
🔄 What Changed
🔍 Impact of the Change
📁 Total Files ChangedClick to Expand
🧪 Test Added/RecommendedAdded
🔒Security Vulnerabilities
|
|
✅ Reviewed the changes: Reviewed src/ui/components/ModelPicker.tsx: the only change is bumping VISIBLE_ROWS from 6 to 12, which safely widens the sliding window (the Math.max/Math.min clamping already handles lists shorter than the window, and the model count remains within the 1-9 digit-selection range) — no issues found. Note: the previously flagged src/core/agent.ts parallel-tool-batch issues are not part of this diff, so their resolution could not be verified here. |
1 participant
Summary
Ports the 6.8.2 coding-harness update from the Orbital extension (MatterAIOrg/Orbital-Extension, commit cdb837e) to OrbCode, switches the built-in catalog to the OSS-first model lineup with dynamic backend sync, and bumps the version to 6.8.0.
Dynamic model catalog sync
/v1/modelsat startup and on usage refresh (fetchDynamicModels); new backend models are registered into the picker and retired ones are pruned. Empty/failed fetches never wipe the offline fallback, and the default model is never pruned.iconUrlandcostMultiplieron each model.zai/glm-5.3-flash(new default),zai/glm-5.3,deepseek/deepseek-v4-flash-0731,meta/muse-spark-1.3-contributor,gpt-5.6-luna,gpt-5.6-sol,gemini-3.8-flash— all 232K context / 64K max output, available on every plan.Provider badges in the model picker
[Z.ai],[Meta],[DeepSeek],[OpenAI],[Google]) — the TUI stand-in for the webapp's provider logos. Visible rows doubled to 12.orbcode usagecommandorbcode login)./usageand/statusshow the same per-model block.search_files is now one-shot
max_results100).Parallel read-only tool execution
read_file,search_files,list_files,list_code_definition_names,codebase_search,lsp) executes concurrently (max 4 workers) with results committed in model order to preserve tool_call/tool_result pairing.Malformed tool-call recovery
Strict-mode schema tightening
replace_all,recursive,follow_up,offset/limit,cwd/message/isDangerous, and the CLI-inactive tool schemas).execute_commandguidance asks for an explicit safety classification.System prompt
search_filesguidance rewritten for the bounded one-shot behavior.Test plan
npm run typecheckandnpm run buildpassnpm run test:search— 11/11 pass