fix: multi-pass response extraction, schema sanitization, workspace safety, and SandboxJS try-catch#583
Open
buger wants to merge 12 commits into
Open
fix: multi-pass response extraction, schema sanitization, workspace safety, and SandboxJS try-catch#583buger wants to merge 12 commits into
buger wants to merge 12 commits into
Conversation
…to SSE server Two bugs fixed: 1. Workflow output value_js received raw ReviewSummary wrappers instead of unwrapped step outputs. This caused every workflow tool (slack-search, slack-read-thread, discourse-read-thread, discourse-reply) to return "Unknown error" because outputs['step'].success was undefined (actual data was nested in .output). Script steps already unwrapped correctly via buildProviderTemplateContext, but workflow-executor's value_js, if conditions, and Liquid contexts did not. 2. executeHttpClientTool in mcp-custom-sse-server only handled bearer auth, not oauth2_client_credentials. When the AI called http_client tools with oauth2 auth (e.g. MongoDB Atlas), no token exchange happened and requests went out without Authorization headers, causing 401 errors. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Fix on_message trigger dispatch to match normal message path: seed setFirstMessage so human_input checks auto-resolve and the full intent-router → build-config → generate-response chain runs with proper tool loading (Jira MCP, Slack, etc.) - Inject trigger.inputs.text as the AI message with original Slack message appended, so triggers can give specific instructions - Fix live update race condition: serialize publish() calls via a promise queue in SlackTaskLiveUpdateSink to prevent duplicate Slack messages when tick() and complete() run concurrently - Track inflightTick promise so complete()/fail() await in-flight ticks before publishing the final update - Fix self-bot message detection for bot_message subtypes by also checking ev.bot_id against the bot's own bot_id from auth.test - Add resolveChannelName() to SlackClient for #channel-name support in scheduler output targets via conversations.list with caching - Allow cron jobs without workflow (inputs.text as user message) - Make StaticCronJob.workflow optional in types - Fix workflow output warning to only fire for undefined (not null) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add debounce-manager for throttling check executions and integrate it into level-dispatch. Supports configurable throttle settings per check via config types. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add global uncaughtException handler that suppresses transient I/O errors (EIO, EPIPE, ECONNRESET, ERR_STREAM_DESTROYED) from dying child processes instead of crashing the entire visor process. Three layers of defense: - Global handler in child-process-error-handler.ts (imported early) - Worktree manager skips process.exit(1) for transient I/O errors - Stream-level error handlers on MCP transport stderr pipes Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Update @probelabs/probe to v0.6.0-rc313 with enriched task telemetry (agent scope fields, full task state on events, task.items_json) - Parse task.items_json from batch events for proper titles on batch created/updated/completed/deleted operations - Collapse sub-agent scopes (engineer, code-explorer) that lack meaningful task titles into deduplicated single-line entries instead of showing repetitive generic "Engineer Task" items - Preserve sub-agent task titles when they exist (from task tool snapshots) - Group repeated sub-agent iterations under a single scope label Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Teaches the assistant to reuse inner ProbeAgent sessions via continue_session when making follow-up calls to the same tool, avoiding expensive cold-start re-execution. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
…ce safety - track-execution: replace naive single-pass response extraction with multi-pass approach (unwrapJsonText, isToolOutput, extractBestResponseText) that filters tool output, unwraps JSON-wrapped text, and falls back through generate-response entries before defaulting to "Execution completed" - mcp-custom-sse-server: add fixRequiredFields to strip invalid entries from required arrays (Gemini rejects schemas referencing non-existent properties); apply normalizeInputSchema to external MCP tools (regularTools) - workflow-tool-executor: fix argsOverrides filter from truthiness check to proper key existence (!(r in argsOverrides)) so falsy override values work - workspace-manager: add safety guards to cleanupStale() — skip directories that are real git repos, validate worktree gitdir paths, refuse to touch parent repos outside the workspace basePath - sandboxjs: bump to probelabs/SandboxJS@d0d8c8a which fixes try-catch with async/await (catch variable extraction, ExecReturn leaking, try-finally without catch) Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
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
1 participant
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
track-execution: Replace naive single-pass response extraction with multi-pass approach that filters ProbeAgent tool output (
├─┬,[task:), unwraps JSON-wrapped text ({"text":"..."}with control char handling), and falls back throughgenerate-responseentries before defaulting to "Execution completed". Fixes tasks showing raw JSON or tool tree output instead of actual AI responses.mcp-custom-sse-server: Add
fixRequiredFieldsrecursive sanitizer that stripsrequiredentries referencing non-existent properties (Gemini rejects these schemas). Also applynormalizeInputSchemato external MCP tools (regularTools) — previously only workflow/HTTP/UTCP tools were normalized.workflow-tool-executor: Fix
argsOverridesfilter from truthiness check (!argsOverrides[r]) to proper key existence (!(r in argsOverrides)) so falsy override values like0,"", orfalseare correctly handled.workspace-manager: Add safety guards to
cleanupStale()— skip directories that are real git repos (have.gitdirectory, not worktree file), validate worktree gitdir paths actually point to.git/worktrees/, and refuse to remove worktrees from parent repos outside the workspace basePath. Prevents accidental deletion of user repositories.SandboxJS: Bump
@nyariv/sandboxjstoprobelabs/SandboxJS@d0d8c8awhich fixes three bugs in try-catch handling for async/await: catch variable extraction (regex group 2→3), ExecReturn leaking that prevented code after try-catch from executing, and try-finally without catch silently swallowing errors.assistant.yaml: Add session continuation prompt guidance for workflow tools.
Test plan
🤖 Generated with Claude Code