Empty work sessions created for every chat interaction - needs filtering and full conversation capture

[Deezinor]

Problem Description

The PAI-OpenCode plugin is creating work sessions for every single user interaction, including trivial messages like greetings ("hello"), acknowledgments ("thanks"), and ratings ("8/10"). This results in hundreds of empty work sessions with no valuable content.

Current Behavior

• 97 work sessions created in just 48 hours of normal usage
• All sessions contain only empty templates:
  • ISC.json: {"criteria": [], "anti_criteria": []}
  • THREAD.md: Just header template, no conversation
  • META.yaml: Status "COMPLETED" with no context
• Zero learning captured from assistant responses or user feedback

Expected Behavior (Per PAI Specification)

Per PAI/SKILL.md, work sessions should only be created based on conversation depth:

Depth	Content	Create Session?
FULL	Problem-solving, implementation, design	✅ Yes
ITERATION	Continuing existing work	✅ Yes
MINIMAL	Greetings, ratings, acknowledgments	❌ No

Root Cause

File: plugins/handlers/work-tracker.ts:139 - createWorkSession()

// Currently creates session unconditionally
export async function createWorkSession(prompt: string) {
// No filtering - creates session for EVERY prompt
const isc = { criteria: [], anti_criteria: [] }; // Always empty!
}

Triggers: Both chat.message and event handlers call createWorkSession() without checking message content or depth.

Additional Issues

Beyond the filtering problem, there are missing features compared to original PAI:

1. Assistant responses never captured - Only user messages (truncated to 200 chars) are written to THREAD.md
2. No items/ subdirectory - Original PAI creates structured work items at items/001/
3. ISC never populated - Empty criteria arrays, no extraction from algorithm output
4. No effort tier detection - Missing QUICK/STANDARD/DEEP classification
5. Ratings not linked to sessions - Ratings go to ratings.jsonl but not to the work session thread

Proposed Solutions

Phase 1: Trivial Message Filtering

Add isTrivialMessage() function to detect and skip:

• Greetings: "hello", "hi", "hey", "good morning"
• Acknowledgments: "ok", "thanks", "got it", "sounds good"
• Ratings: "8", "8/10", "9 - great work"
• Short messages: <20 characters

Phase 2: Full Conversation Capture

• Capture full user messages (not truncated)
• Add appendAssistantToThread() to capture assistant responses
• Link ratings to work sessions with context

Phase 3: Original PAI Structure

• Create items/001/ subdirectory with META.yaml, ISC.json, RESPONSE.md
• Detect effort tier (QUICK/STANDARD/DEEP) from prompt
• Extract ISC criteria from algorithm output patterns
• Add YAML frontmatter to all files

Acceptance Criteria

• Trivial messages (greetings, ratings, acknowledgments) do NOT create work sessions
• Meaningful work creates sessions with full directory structure
• Assistant responses captured in THREAD.md
• Full conversation preserved (not truncated)
• Ratings linked to work sessions
• ISC.json populated with criteria from algorithm output
• Effort tier detected and stored in META.yaml

References

• Original PAI Memory System: SYSTEM/MEMORYSYSTEM.md
• PAI Algorithm Depth: SKILL.md lines 47-51
• Original PAI Repository: https://github.com/danielmiessler/Personal_AI_Infrastructure

Environment

• OpenCode: v1.1.x
• PAI-OpenCode: Current main branch
• OS: Linux (tested on WSL2)

Labels

bug, memory-system, performance, enhancement

[Steffen025]

Hi @Deezinor — great issue report with an excellent phased proposal. We've implemented Phase 1 (trivial message filtering) in PR #26.

What's fixed

• isTrivialMessage() function added to work-tracker.ts — detects greetings, acknowledgments, ratings, farewells, and short messages
• Both session creation paths in pai-unified.ts now check before creating sessions
• Messages like "hello", "thanks", "8/10" will no longer create empty work directories

What's next (future PRs)

• Phase 2: Full conversation capture — assistant responses in THREAD.md, un-truncated user messages
• Phase 3: ISC population from algorithm output, effort tier in META.yaml, items/ structure

Please test with the PR branch and let us know if the filtering catches your use cases. If you find messages that should be filtered but aren't (or vice versa), feedback is welcome on the PR.

Thanks for the thorough analysis — this will clean up a lot of noise in the MEMORY/WORK/ directory.