Skip to content

features.md

Latest commit

 

History

History
92 lines (73 loc) · 5.69 KB

features.md

File metadata and controls

92 lines (73 loc) · 5.69 KB

Features

Feature Status Description
Hook Translation 🔄 Adapted from PAI Maps 20 PAI hooks → 9 semantic events → 8 OpenCode hooks
Config Translation 🔄 Adapted from PAI settings.json → opencode.json with merge semantics
Session State 🔄 Adapted from PAI Per-session Map<sessionId, T> with auto-cleanup
Security Validator 🔄 Adapted from PAI Input sanitization + 7-category injection detection (tool.execute.before)
Secret Scrubber ✅ Native to OC Redacts env secrets and API key patterns from LLM input messages (experimental.chat.messages.transform)
Plan Mode 🔄 Adapted from PAI Read-only mode via /plan command, blocks destructive tools
Model Routing ✅ Native to OC User-configurable model-per-role mapping with fallback chains
Voice Notifications 🔄 Adapted from PAI ElevenLabs TTS, ntfy.sh, Discord webhooks
StatusLine 🔄 Adapted from PAI tmux status-right integration with phase, tokens, learning signals
Compaction (Proactive) 🔄 Adapted from PAI Injects survival context during experimental.session.compacting
Compaction (Reactive) 🔄 Adapted from PAI Rescues learnings after session.compacted event
Learning Tracker 🔄 Adapted from PAI Captures ratings, sentiment, tool outcomes to JSONL
Context Loader 🔄 Adapted from PAI Loads TELOS + context files on session start
Message Deduplication 🔄 Adapted from PAI 5s TTL dedup cache prevents double-fire
Session Lifecycle 🔄 Adapted from PAI JSONL session tracking with memory summary
Terminal UI (Kitty) ⚠️ Limited Support Kitty tab integration (requires Kitty terminal)
CLI Shim 🔄 Adapted from PAI claude command → opencode wrapper script
Self-Updater ✅ Native to OC Monitors PAI + OC for updates, creates draft PRs
File Logging ✅ Native to OC /tmp/pai-opencode-debug.log (never console.log)
Event Bus ✅ Native to OC Internal pub/sub for adapter events
Audit Logger ✅ Native to OC Security audit JSONL for compliance
Env Loader ✅ Native to OC Auto-loads API keys from ~/.config/PAI/.env for skills
Skill Loader ✅ Native to OC Native OpenCode skill tool support for PAI skills
Agent Model Sync ✅ Native to OC Syncs model: field in agent .md files from pai-adapter.json on startup
Subagent Error Detection ✅ Native to OC Checks top-level + stringified Task output for provider errors
Subagent Model Fallback ✅ Native to OC Injects alternative subagent_type suggestions on provider failures
Subagent Stall Detection ✅ Native to OC 3-minute heartbeat monitor per subagent, warns primary agent
Subagent Loop Detection ✅ Native to OC Hashes reasoning text in rolling window, detects repetitive thinking
PAI Protection Rule ✅ Native to OC Prevents accidental modification of upstream PAI files

Status Legend

  • ✅ Native to OC — OpenCode native feature, adapter uses it directly
  • 🔄 Adapted from PAI — PAI feature translated to OpenCode events
  • ⚠️ Limited Support — Feature available with constraints or dependencies

What This Adapter Does

  1. Event translation — Maps 20 PAI hook files → 9 semantic events → 8 OpenCode plugin hooks
  2. Config translation — Converts settings.json to opencode.json format
  3. State management — Per-session state with automatic cleanup
  4. Security validation — Tool gating and input sanitization
  5. Compaction handling — Dual proactive and reactive session compaction
  6. Voice notifications — ElevenLabs TTS for task completion alerts
  7. Subagent reliability — 4-layer suite: error detection, model fallback, stall detection, reasoning loop detection

What This Adapter Does NOT Do

  • Modify PAI source files (read-only wrapper)
  • Add npm dependencies beyond TypeScript
  • Auto-merge updates (human review always required)

Subagent Reliability Suite (v0.9.1)

The adapter monitors all subagent sessions and provides automatic guidance when things go wrong. All mechanisms are fail-open — they inject warnings into system prompts but never block execution.

Layer 1: Enhanced Error Detection

Detects provider errors in Task tool output by checking:

  • Top-level error fields in the Task result
  • Full stringified body for error patterns (rate limits, model not found, connection errors, auth failures)

Layer 2: Actionable Model Fallback

When a provider error is detected:

  • Identifies the failed subagent_type and its model
  • Looks up the fallback chain for that role in pai-adapter.json
  • Injects a <system-reminder> with specific alternative subagent_type suggestions and their models
  • Progresses through the fallback chain on repeated failures

Layer 3: Stall Detection

Monitors subagent activity via message.part.updated events:

  • Tracks last activity timestamp per subagent session
  • If 3 minutes pass with no activity, warns the primary agent
  • Warning suggests the subagent may be hung and recommends retrying with a different model

Layer 4: Reasoning Loop Detection

Detects when a subagent gets stuck repeating the same reasoning:

  • Hashes each reasoning text chunk (via message.part.updated for reasoning parts)
  • Maintains a rolling window of 8 hashes per subagent
  • If 3+ identical hashes appear in the window, warns the primary agent
  • Session cleanup removes all tracking state on session end

← Back to README · Agents · Compatibility Registry