RCP-CWK-014 Cowork Error Interceptor
Detects errors in tool output, classifies severity, translates them to plain-language explanations, and suggests executable fix commands.
Detects errors in tool output, classifies severity, translates them to plain-language explanations, and suggests executable fix commands.
Extends the CRAFT Operations Manager (Core COM) with Cowork-specific monitoring: file system awareness, git state tracking, enhanced token threshold actions, monorepo push safety, branding detection, file size watchlists, and a post-response priority tracker footer. Runs as a persistent overlay throughout the entire session.
Tracks multi-step CRAFT workflows visually using Cowork’s native TodoWrite widget for real-time progress display, while maintaining CRAFT’s structured workflow state for handoff serialization. Extends the Core Progress Tracker by bridging CRAFT progress semantics to Cowork’s built-in UI.
Logs estimates to a session-scoped file and integrates with the COM Extension (CWK-012) for automatic threshold-based triggers. Extends the Core Token Monitor with persistent logging and Cowork-specific capacity planning.
Creates the standard 7-folder hierarchy for a CRAFT project in Cowork: framework/, project/ (with child/), templates/, sessions/, research/, branding/ (with assets/), and transfers/ (with inbound/outbound/). Also generates root files — .gitignore, a CLAUDE.md placeholder, a PROJECT-MANIFEST.txt scaffold, and a roadmap .docx template.
Supports four workflow patterns: Builder-Validator (one agent creates, another reviews), parallel research (multiple agents investigate different angles simultaneously), sequential handoff (output of one agent feeds the next), and single delegate (one task, one agent).
CLAUDE.md is the bridge between CRAFT’s infrastructure and Cowork’s native memory system — it gets loaded automatically at session start so the AI has your project identity, personas, cookbooks, priorities, and key files without you pasting anything.
Classifies every action as ROUTINE, NON-ROUTINE, or NON-ROUTINE-HIGH, then emits a structured 7-field recommendation block (ACTION, CLASS, CONFIDENCE, IMPACT, REVERSIBILITY, FILES, RECOMMENDATION). Non-routine actions are gated — the AI stops and waits for human approval before proceeding. Three invocation modes: GATE (blocking), LOG (non-blocking file-write summary), and ASSESS (standalone evaluation).
Derives target repos dynamically from PROJECT_ID and MONOREPO_ROLE. Never hardcodes repo names. For Global pushes, delegates to CWK-045. For Local pushes, manages repo discovery, remote validation, archive detection, auto-fix of wrong remotes (with human confirmation), and PAT-based push. Supersedes CWK-038 for two-repo model pushes.
Creates compressed long-term archives of project state as a single timestamped .tar.gz file suitable for cold storage or off-site transfer.Â
Walks through a 5-phase workflow: assess current state, identify recovery sources, execute recovery, verify restored state, and document the incident. Every destructive step requires explicit human confirmation.
Performs an 8-point validation covering repository state, branch verification, remote synchronization, scope validation, force-push detection, structure integrity, changelog awareness, and pre-push backup. Returns an explicit GO, CONDITIONAL GO, or NO-GO verdict. Push proceeds only on GO.
End of content
End of content