Release v6.8.2: parallel read-only tools, malformed tool-call recovery, and one-shot search - #127
Merged
Merged
Conversation
…y, and one-shot search - Execute independent read-only native tools (read_file, search_files, list_files, list_code_definition_names, codebase_search, lsp) concurrently (max 4) with results committed in model order; mutating and interactive tools stay serialized. parallel_tool_calls enabled for native JSON providers. - Recover from malformed tool-call JSON: dropped calls are tracked by the parser and retried with a corrective message instead of silently stopping the agent. - Bound tool-repetition auto-retry at 2 attempts to prevent infinite model loops. - search_files: ripgrep-first with FFF fallback, one-shot 100-match results, cursor pagination removed from model-facing schemas. - Condense system prompt rules/guidelines/objective; codebase_search is recommended, not mandatory-first; legacy text guidance only for XML tool style. - Tighten native tool schemas for strict mode (required nullable params, execute_command safety metadata) and respect empty allowedTools. - Reduce DEFAULT_WRITE_DELAY_MS 1000ms -> 250ms; stub KiloCode notification polling. - Bump extension version 6.8.1 -> 6.8.2 and update CHANGELOG. Co-authored-by: matterai-app[bot] <bot@matterai.so>
Contributor
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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
Bumps the extension 6.8.1 -> 6.8.2 and ships the accumulated agent-loop work:
Added
read_file,search_files,list_files,list_code_definition_names,codebase_search,lsp) returned in one native-JSON assistant message now run concurrently (max 4,MAX_PARALLEL_READ_ONLY_TOOLSin the newsrc/core/tools/toolExecutionPolicy.ts), with results committed in model order to preserve tool_use/tool_result pairing. Mutating/interactive tools stay serialized; read approvals inside a batch use non-interactive progress rows.parallel_tool_callsis enabled for native tool call providers.Changed
search_filesone-shot and ripgrep-first. Ripgrep is the default engine (FFF fallback), default/max results 50 -> 100, model-facingcursorremoved from XML and native JSON schemas; capped pages direct the model to refine the query instead of paginating.codebase_searchrecommended (not mandatory-first), smallest sufficient tool preferred, independent reads batched in JSON mode while edits/commands stay sequential. Legacy text-tool guidance only included for XML tool style.execute_commandrequirescwd/message/isDangerous;file_edit/multi_file_editrequirereplace_all;generate_filerequirespath;check_past_chat_memoriesrequiresworkspace(nullable where optional);lspdescription condensed.allowedToolsmetadata respected when empty inaddNativeToolCallsToParams(intentional empty list for tool-less modes no longer falls back to the full tool set).DEFAULT_WRITE_DELAY_MS1000ms -> 250ms.Fixed
AssistantMessageParserand retried with a corrective message (raw-args preview) instead of silently stopping the agent; bounded by the consecutive-mistake limit.MAX_TOOL_REPETITION_AUTO_RETRIES) so an unchanged model cannot loop forever.Changelog
## [v6.8.2] - 2026-08-28section and relabeled the previous## [Unreleased]block to## [v6.8.1] - 2026-08-21(that content shipped in v6.8.1), so the publish workflow's[v<version>]extraction finds real release notes.Verification
tsc --noEmit(extension) and turbocheck-types/lintacross the monorepo: clean.presentAssistantMessage,nativeToolCallHelpers(new),tool-contracts(new),objective,tool-use-guidelines,search-files.main. On the specs overlapping this diff the release strictly reduces failures (63 -> 48;system-prompt.spec.ts13 -> 8) and introduces none. The remainingsystem-prompt.spec.tsfailures are stale expectations from the v6.8.1 prompt changes and are untouched here.