Skip to content

CHANGELOG.md

Latest commit

 

History

History
742 lines (570 loc) · 37.8 KB

CHANGELOG.md

File metadata and controls

742 lines (570 loc) · 37.8 KB
Raw

Changelog

All notable changes to PAI-OpenCode are documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

3.0.1 (2026-04-13)

Bug Fixes

  • migration: add missing closing brace in migrateSkills function (#147) (1625b41)
  • update hero banner path in README (e7e1c92)

3.0.0 (2026-04-13)

PAI-OpenCode v3.0 is the OpenCode-native release — a complete re-architecture that moves from a Claude Code fork to vanilla OpenCode, removes the bootstrap loading mechanism in favour of the native skill system, and ships a zero-config Zen-free default so users are productive immediately.

⚠️ Breaking Changes

  • model_tiers removed (WP-M1). Each agent in opencode.json has exactly one model field. The model_tiers block (quick/standard/advanced) and the model_tier Task parameter are gone. Run migrate-legacy-config.ts to auto-convert. See ADR-019 and UPGRADE.md.
  • Vanilla OpenCode only. The custom Steffen025/opencode fork is archived. Install via the official opencode.ai installer — no custom build required.
  • Bootstrap loading removed (ADR-020). MINIMAL_BOOTSTRAP.md is deleted. The PAI Core Skill (Algorithm, ISC, Capabilities) now loads via OpenCode's native skill system (tier: always via GenerateSkillIndex.ts + skills/PAI symlink). Plugin loads user identity context only.
  • skills/CORE renamed to skills/PAI. All internal paths updated.
  • Provider setup is post-install. The install wizard no longer asks about providers. Connect via /connect in OpenCode, then update agent models with switch-provider.ts. See INSTALL.md.

Migration

For v2.x users with model_tiers configs:

bun run PAI-Install/engine/migrate-legacy-config.ts ~/.opencode/opencode.json

Full guide: UPGRADE.md.

Features

Core Architecture

  • Algorithm v3.7.0 ported from upstream PAI v4.0.3 — Constraint Fidelity System, Verification Rehearsal, ISC Adherence Check, Build Drift Prevention (WP1, PRs #32-35)
  • Hierarchical skill structure — 11 categories (Agents, ContentAnalysis, Investigation, Media, Research, Scraping, Security, Telos, Thinking, USMetrics, Utilities) migrated from flat layout (WP4, PRs #38-40)
  • Native skill loading — PAI/SKILL.md is tier: always via GenerateSkillIndex.ts; skills load on-demand via OpenCode's native skill tool; 233KB static context eliminated (WP2, PR #34 + PR #138)
  • Zen free out-of-box — all agents default to opencode/big-pickle (permanent Zen flagship, no API key required); volatile free Zen models removed as defaults due to rotation risk (PR #138)

Event-Driven Plugin System (WP3 + WP-A, PRs #37, #42)

  • Unified plugin pai-unified.ts — replaces 5 separate plugins and hook emulation layer
  • 5 new plugin handlers: prd-sync.ts, session-cleanup.ts, last-response-cache.ts, relationship-memory.ts, question-tracking.ts
  • Bus events: session.compacted, session.error, permission.asked, command.executed, installation.update.available, session.updated, session.created
  • shell.env hook — PAI context injected per bash call
  • loadUserSystemContext() — plugin now loads only user-specific context (PAI/AISTEERINGRULES.md + USER/* identity files); Algorithm content loads via skill system

Security Hardening (WP-B, PR #43)

  • Prompt injection detection — injection-patterns.ts with configurable sensitivity (low/medium/high)
  • Input sanitization layer — sanitizer.ts pre-processes input before LLM
  • Security event logging — audit trail for injection attempts
  • Integration runs on tool.execute.before and message.received events

Core PAI System (WP-C, PR #45)

  • Skill structure fixes — flattened incorrectly nested USMetrics/USMetrics/ and Telos/Telos/ directories
  • New skills ported from PAI v4.0.3: Utilities/AudioEditor/, Utilities/Delegation/
  • PAI docs ported from v4.0.3: CLI.md, CLIFIRSTARCHITECTURE.md, DOCUMENTATIONINDEX.md, FLOWS.md, PAIAGENTSYSTEM.md, THENOTIFICATIONSYSTEM.md + 3 subdirectories (ACTIONS/, FLOWS/, PIPELINES/)
  • BuildOpenCode.ts — OpenCode-native replacement for BuildCLAUDE.ts

Installer & Migration (WP-D, PR #47)

  • Migration script migrate-legacy-config.ts — auto-converts model_tiers blocks, preserves standard tier model, creates .pre-v3.0.bak backup
  • DB Health tooling — db-utils.ts monitors DB size and session age; warns when DB > 500 MB or sessions > 90 days old
  • db-archive.ts — standalone Bun tool (--dry-run, --vacuum, --restore)
  • /db-archive custom command — session archiving in OpenCode TUI
  • DB-MAINTENANCE.md — operational guide
  • Interactive terminal install wizard — CLI-only, Electron GUI removed (PR #96)
  • PAI wrapper script — pai command separate from opencode, explicit PAI_ENABLED=1 (PR #104)

Session Intelligence (WP-N1–N3, PRs #50-53)

  • Session Registry — session_registry and session_results custom tools for post-compaction context recovery (ADR-012)
  • Compaction Intelligence — compaction-intelligence.ts hooks experimental.session.compacting; injects registry + ISC + PRD context into compaction summary to prevent silent context loss (ADR-015)
  • Algorithm Awareness — SKILL.md updated with post-compaction recovery pattern; parent_session_id links child PRDs to originating session (ADR-013)

Documentation & Tooling (WP-N4–N10, PRs #53-59)

  • LSP integration documented — OPENCODE_EXPERIMENTAL_LSP_TOOL=true opt-in, LSP vs Grep decision table (ADR-014)
  • System Self-Awareness skill — OpenCodeSystem/SKILL.md with SystemArchitecture.md, ToolReference.md, Configuration.md, Troubleshooting.md (ADR-017)
  • roborev + Biome CI — roborev-trigger.ts handler, CodeReview skill, code-quality.yml GitHub Actions workflow (ADR-018)
  • Obsidian Formatting Guidelines — FormattingGuidelines.md and AgentCapabilityMatrix.md (16 agents, models, tool access, decision rules)
  • ADR-019 — Vanilla OpenCode migration decision record
  • ADR-020 — Native OpenCode context loading (bootstrap removal) decision record

Provider Infrastructure

  • switch-provider.ts — switches all agent models with one command; profiles: zen, zen-free, zen-paid, anthropic, openai, local; --multi-research for native research agents
  • Two-step provider setup documented — /connect (credentials) → switch-provider.ts (opencode.json agent models)
  • Anthropic Claude Max ToS warning — community plugin not shipped; ToS-safe alternatives documented in INSTALL.md

CI/CD

  • GitHub Actions — ci.yml, code-quality.yml, codeql.yml, scorecards.yml, upstream sync workflows
  • CodeRabbit integration (.coderabbit.yaml)
  • StepSecurity hardening — CI workflows hardened (PR #108)
  • release-please — automated release management

Bug Fixes

  • Fix files_loaded counter — now counts all injected files including PAI/AISTEERINGRULES.md (was hardcoded 1)
  • Fix USER/ → PAI/USER/ paths in plugin, docs, and diagrams
  • Remove provider API key write from .env generation — model provider auth via /connect, not .env
  • Remove misleading ANTHROPIC_API_KEY/OPENAI_API_KEY examples from .env template
  • Rename loadMinimalBootstrap() → loadUserSystemContext() — name now matches behaviour
  • Fix stale log messages in plugin ("minimal bootstrap" → "user system context")
  • Fix ALWAYS_LOADED_SKILLS docs — correctly shows all 6 entries matching GenerateSkillIndex.ts

Removed

  • PAI-Install/electron/ — GUI installer
  • PAI-Install/engine/build-opencode.ts — fork build system (245 LOC)
  • MINIMAL_BOOTSTRAP.md — replaced by native skill system
  • model_tiers blocks from opencode.json and all provider profiles
  • model_tier parameter from agent-execution-guard.ts, session-registry.ts
  • tiers: sections from all provider YAML profiles
  • PAI-Install/web/ and PAI-Install/main.ts
  • All references to Steffen025/opencode fork and feature/model-tiers branch

Dependencies

  • bump @opencode-ai/plugin 1.1.39 → 1.4.3 (#118)
  • bump actions/checkout v4 → v6 (#124)
  • bump actions/github-script v7 → v9 (#129)
  • bump actions/upload-artifact v4 → v7 (#127)
  • bump github/codeql-action v3 → v4 (#114)
  • bump ossf/scorecard-action 2.4.0 → 2.4.3 (#111)
  • bump sharp 0.33.5 → 0.34.5 in Pptx/Scripts (#119)
  • bump @types/node v20 → v25 in BugBountyTool (#113)

Contributors

Special thanks to community members whose contributions are included in this release:

Contributor Contribution PR
@eddovandenboom PAI wrapper script, --fix-symlink flag #104
@ktneely Model profile updates for provider-agnostic routing #103
@step-security-bot CI security hardening, pinned Actions SHA hashes #106

[2.0.0] - 2026-02-19

Breaking Changes

  • Algorithm format changed (8 effort levels replace FULL/ITERATION/MINIMAL depth)
  • Start symbol changed from 🤖 to ♻︎
  • ISC naming convention updated to ISC-{Domain}-{N} with priority/confidence tags
  • SKILL.md completely rewritten — Algorithm v1.2.0 → v1.8.0 (upstream sync)

Added

v1.2.0 Base (2026-02-17)

  • Constraint Extraction System — Mechanical [EX-N] extraction before ISC
  • Self-Interrogation — 5 structured questions before BUILD
  • Build Drift Prevention — Re-read [CRITICAL] ISC before each artifact
  • Verification Rehearsal — Simulate violations in THINK phase
  • Mechanical Verification — No rubber-stamp PASS, require evidence
  • 8 Effort Levels — Instant, Fast, Standard, Extended, Advanced, Deep, Comprehensive, Loop
  • 7 Quality Gates — QG1-QG7 must pass before proceeding
  • 25-Capability Full Scan Audit — Replaces Two-Pass capability selection
  • PRD System — Persistent Requirements Documents for cross-session tracking
  • Anti-Criteria — ISC-A-{Domain}-{N} for what must NOT happen
  • Algorithm Reflection JSONL — Structured Q1/Q2/Q3 learning capture
  • OBSERVE Hard Gate — Thinking-only phase, no tool calls except TaskCreate
  • AUTO-COMPRESS — Drop effort tier when >150% of phase budget
  • 10 new skills — Cloudflare, ExtractWisdom, IterativeDepth, Science, Parser, Remotion, Sales, USMetrics, WorldThreatModelHarness, WriteStory
  • 5 new plugin handlers — algorithm-tracker, agent-execution-guard, skill-guard, check-version, integrity-check
  • format-reminder handler updated for 8-tier effort level system
  • PRD directory structure with templates and lifecycle management

Upstream Sync v1.3.0–v1.8.0 (2026-02-19)

Porting 14 upstream commits spanning Algorithm v1.3.0 through v1.8.0:

  • Verify Completion Gate (v1.6.0) — CRITICAL: Prevents "PASS" claims without actual TaskUpdate calls. NON-NEGOTIABLE at ALL effort levels.
  • Phase Separation Enforcement (v1.6.0) — "STOP" markers on THINK, PLAN, BUILD, EXECUTE, VERIFY phases
  • Zero-Delay Output Section (v1.6.0) — Instant output before any processing
  • Self-Interrogation Effort Scaling (v1.3.0) — Instant/Fast skip, Standard answers 1+4, Extended+ answers all 5
  • Constraint Extraction Effort Scaling (v1.3.0) — Gate added for effort levels below Standard
  • Steps 6-8 Gated to Extended+ (v1.3.0) — Constraint Fidelity System steps scale by effort
  • QG6/QG7 Gated to Extended+ (v1.3.0) — Quality gates scale by effort in OBSERVE and PLAN
  • ISC Scale Tiers Updated (v1.3.0) — Simple: 4-16, Medium: 17-32, Large: 33-99, Massive: 100-500+
  • BUILD Capability Execution Substep (v1.8.0) — Explicit capability execution within BUILD
  • Wisdom Injection OUTPUT 1.75 (v1.8.0) — Injects domain wisdom between Constraint Extraction and ISC
  • Wisdom Frame Update in LEARN (v1.8.0) — Captures new wisdom into domain frames
  • Algorithm Reflection Moved First in LEARN (v1.8.0) — Reflection before PRD LOG
  • Wisdom Frames System — MEMORY/WISDOM/ directory with 5 seed domain frames (development, deployment, security, architecture, communication)
  • WisdomFrameUpdater.ts — CLI tool for managing wisdom frames (--domain, --observation, --type, --list, --show)
  • Security Validator env var prefix fix — Upstream #620: strips export/set/declare/readonly prefixes to prevent false positives
  • Rating Capture 5/10 noise filter — Ambiguous 5/10 ratings skip learning file generation (still tracked in JSONL)
  • Symlink support in GenerateSkillIndex.ts — findSkillFiles() follows symlinks to directories, handles broken symlinks gracefully
  • SessionHarvester PAI_DIR rename — CLAUDE_DIR → PAI_DIR with process.env.PAI_DIR fallback

Changed

  • SKILL.md rewritten from Algorithm v0.2.25 to v1.8.0 (was v1.2.0)
  • Capability selection now uses 25-capability full scan (was Two-Pass)
  • ISC criteria now use domain-grouped naming convention
  • format-reminder handler enhanced with effort level detection
  • Constraint Fidelity System updated to v1.3.0
  • LEARN phase restructured: Algorithm Reflection → PRD LOG → Wisdom Frame Update → Learning → Voice
  • Voice curl commands now use {DAIDENTITY.ALGORITHMVOICEID} template variable

Not Portable (Claude Code Only)

  • Agent Teams/Swarm (requires CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1)
  • Plan Mode (EnterPlanMode/ExitPlanMode built-in tools)
  • StatusLine (Claude Code UI feature)

v2.0.1 — Legacy Installer Build Process Fix (2026-02-20)

Fixed

Critical: Legacy Installer Build-from-Source Completely Broken

  • Symptom: git clone succeeds but legacy installer reports "Failed to clone repository" and aborts
  • Root cause: execCommand() with stdio: 'inherit' returns null from execSync(). Calling .trim() on null throws TypeError, caught as failure. Every non-silent command falsely reports failure.
  • Fix: Null-safe guard: output.trim() → (output ?? '').trim()

Go Prerequisite Removed (No Longer Needed)

  • Context: OpenCode was historically a Go project (BubbleTea TUI). It has been completely rewritten to Bun/TypeScript. The build now uses Bun.build({ compile: true }) to produce native binaries — no Go toolchain needed.
  • Fix: Removed Go prerequisite check from the legacy installer and all documentation
  • Fix: Strengthened Bun version check from 1.3+ to 1.3.9+ (matches OpenCode monorepo packageManager field)

Binary Detection After Build

  • Symptom: After successful build, legacy installer couldn't find the binary
  • Root cause: Generic filename search didn't match Bun.build() output structure (dist/opencode-{os}-{arch}/bin/opencode)
  • Fix: Deterministic platform-based binary lookup using process.platform + process.arch with baseline fallback

Changed

  • Legacy installer messaging updated to reflect Bun-based build (not Go)
  • INSTALL.md — Prerequisites: removed Go, added Bun 1.3.9+ note. Manual install: complete rewrite with bun run ./packages/opencode/script/build.ts --single. WSL section: removed golang-go package.
  • README.md — Quick Start: added prerequisites line, updated installer step description
  • docs/MIGRATION.md — Replaced go install with installer command, replaced $(go env GOPATH)/bin with ~/.local/bin in PATH troubleshooting
  • EXPLORATION-SUMMARY.md (since removed) — Updated wizard flow description

v1.3.2 — Wizard Fixes + Fork Alignment (2026-02-11)

Fixed

  • Wizard: 4 provider presets (Anthropic, Zen Paid, Zen Free, Ollama Local) with clearer auth guidance
  • Wizard: Build-from-source now clones from Steffen025/opencode fork (feature/model-tiers branch)
  • Wizard: Non-blocking Go prerequisite check (warning only, no abort)
  • Wizard: Auto-creates .opencode/opencode.json symlink → ../opencode.json (single source of truth)
  • Voice Server: macOS say TTS fallback when no API keys configured
  • Observability Dashboard: Fixed type → event_type field naming across all Vue components
  • Observability Server: Added completeSession() on session.end events

Deprecated

  • Observability Dashboard — Will be removed in a future version. The Vue-based dashboard adds significant dependency overhead with limited practical value. Server-side JSONL event logging remains unaffected.

v1.3.1 — Plugin Crash Fix + Interface Alignment (2026-02-11)

Fixed

Critical: Tool Execution Crash (All Tools Affected)

  • Symptom: ALL tools (Bash, Read, Glob, Grep, etc.) crashed with TypeError: undefined is not an object (evaluating 'Object.keys(args)')
  • Root cause: Interface mismatch between pai-unified.ts (caller) and observability-emitter.ts (receiver). Emit functions expected positional parameters but were called with object parameters after a feature sync.
  • Fix: All 14 emit functions in observability-emitter.ts converted to accept object parameters with defensive null-checks. All 17 call-sites in pai-unified.ts updated to match.

Critical: OpenCode Startup Failure

  • Symptom: OpenCode refused to start with config validation error
  • Root cause: Invalid "pai" top-level key in opencode.json — OpenCode doesn't recognize this config key
  • Fix: Removed "pai" block from opencode.json

Installation Wizard: Build-from-Source Completely Broken (#21)

  • Symptom: git clone in wizard fails with 404; even if clone succeeded, build would fail
  • Root cause 1: Clone URL referenced nicepkg/opencode.git which no longer exists
  • Root cause 2: Build commands used go build — but OpenCode is a TypeScript/Bun project, not Go
  • Root cause 3: Install paths referenced ~/go/bin which is irrelevant
  • Fix: Complete rewrite of build-from-source function:
    • Clone URL updated to anomalyco/opencode.git
    • Build process now uses bun install + bun run ./packages/opencode/script/build.ts --single
    • Added Bun 1.3+ prerequisite check with version validation
    • Binary search in packages/opencode/dist/ with platform-specific names
    • Install to ~/.local/bin or /usr/local/bin (removed Go paths)

Changed

Simplified Versioning

  • Removed @version tags from individual plugin handler files (pai-unified.ts, observability-emitter.ts, implicit-sentiment.ts)
  • Single version source: Only the repository version (in README + CHANGELOG) matters. Individual subsystems don't have their own version lifecycle since they're always released together.

v1.3.0 — Dynamic Per-Task Model Tier Routing (2026-02-10)

🚀 Major: Dynamic Tier Routing Across Provider Boundaries

The orchestrator now automatically routes each task to the right model at the right cost — and the same agent scales up or down dynamically based on task complexity.

This is a turning point for PAI-OpenCode. Up until v1.2, we were running a 1:1 port of vanilla PAI. With v1.3.0, we leverage what makes OpenCode unique: multi-provider support with dynamic model routing.

As far as we can tell, no other AI coding assistant or agent framework currently offers this pattern of dynamic per-task model tier routing across provider boundaries.

How It Works

  • Three-tier model routing — quick, standard, advanced tiers per agent in opencode.json
  • Orchestrator decides per task — Same Engineer uses GLM 4.7 for batch edits, Kimi K2.5 for features, Claude Sonnet 4.5 for complex debugging
  • You always pay exactly what the task requires — no more, no less
  • Backward-compatible: model field still works as fallback

Dynamic Tier Routing Table

Agent Default Scales Down To Scales Up To
Architect Kimi K2.5 GLM 4.7 (quick review) Claude Opus 4.6 (complex architecture)
Engineer Kimi K2.5 GLM 4.7 (batch edits) Claude Sonnet 4.5 (complex debugging)
DeepResearcher GLM 4.7 MiniMax (quick lookup) Kimi K2.5 (deep analysis)
GeminiResearcher Gemini 3 Flash — Gemini 3 Pro (deep research)
PerplexityResearcher Sonar — Sonar Deep Research
GrokResearcher Grok 4.1 Fast — Grok 4.1 (full analysis)
CodexResearcher GPT-5.1 Codex Mini — GPT-5.2 Codex
Writer Gemini 3 Flash MiniMax (quick drafts) Claude Sonnet 4.5 (premium copy)
Pentester Kimi K2.5 GLM 4.7 (quick scan) Claude Sonnet 4.5 (deep audit)
Intern MiniMax M2.1 — —
explore MiniMax M2.1 — —
QATester GLM 4.7 — —

Agent System Changes

  • Model routing centralized — Agent .md files no longer contain model: in frontmatter. ALL model routing now lives exclusively in opencode.json
  • 15 specialized agents with dynamic tier routing
  • DeepResearcher replaces ClaudeResearcher (provider-agnostic naming)
  • Removed PerplexityProResearcher (redundant), researcher (lowercase duplicate)

Migration Notes:

  • ⚠️ If you have custom workflows referencing ClaudeResearcher, update them to DeepResearcher
  • ⚠️ If you have custom skills referencing PerplexityProResearcher, migrate to PerplexityResearcher with model_tier: standard (Sonar Pro). Use model_tier: advanced only for Sonar Deep Research.

3 Configuration Presets

  • zen-paid (Recommended) — 75+ providers via Zen AI Gateway. Combine providers freely.
  • openrouter — OpenRouter routing with familiar model names.
  • local-ollama — Fully local with Ollama. Zero cloud, complete privacy.

Provider Profiles (v3.0)

  • New YAML format with default_model + agents structure including tiers
  • New profile: zen-paid.yaml for privacy-preserving pay-as-you-go models
  • Renamed: zen.yaml is now ZEN FREE (community/free models)
  • Removed: google.yaml (use manual config via ADVANCED-SETUP.md)
  • switch-provider.ts v3.0 — Updated for new profile format with model_tiers generation

Documentation

  • NEW: ADVANCED-SETUP.md — Guide for multi-provider research, custom models, and manual configuration
  • Updated: PAIAGENTSYSTEM.md fully rewritten with model tier guide and dynamic routing
  • Updated: README.md with dynamic tier routing table and new presets
  • 47 documentation gaps fixed across 11 files

Image Optimization

  • All 17 images in docs/images/ resized and compressed
  • Total reduction: 12.4 MB → 2.6 MB (79%)

Breaking Changes

  • Profile YAML format changed (models: → default_model: + agents: with tiers)
  • ClaudeResearcher renamed to DeepResearcher (update any custom workflows)
  • PerplexityProResearcher removed (use PerplexityResearcher with standard tier for Sonar Pro)
  • Agent .md files no longer accept model: field — use opencode.json exclusively
  • Google profile removed — configure manually if needed

Profile Format Change (Before → After)

Old format (v1.2.x):

models:
  - model: anthropic/claude-opus-4-6
    agents: [Algorithm]
  - model: anthropic/claude-sonnet-4-5
    agents: [Architect, Engineer, Writer]

New format (v1.3.0):

default_model: anthropic/claude-sonnet-4-5
agents:
  Algorithm:
    model: anthropic/claude-opus-4-6
    tiers:
      quick: anthropic/claude-haiku-4-5
      standard: anthropic/claude-sonnet-4-5
      advanced: anthropic/claude-opus-4-6

Stats

  • 113 files changed, 2,824 insertions, 1,792 deletions
  • 15 agents with dynamic tier routing
  • 3 presets ready to use

Migration Guide

  1. Re-run the installer update flow: bash PAI-Install/install.sh --update
  2. Or switch profile manually: bun run .opencode/tools/switch-provider.ts zen-paid
  3. Custom agent models → Edit opencode.json agent section directly

[1.2.1] - 2026-02-06

Major Feature: Provider Profile System + Multi-Provider Research

One-command provider switching for all 18 agent models, with optional multi-provider research routing for diverse AI perspectives.

Added

Provider Profile System

  • 5 Provider Profiles (profiles/*.yaml): Anthropic, OpenAI, Google, ZEN (free), Local (Ollama)
  • 3-Tier Model Strategy: Each profile maps agents to Most Capable → Standard → Budget tiers
  • switch-provider.ts v2.0: CLI tool to switch all agent models with one command
    • bun run switch-provider.ts anthropic — switch to Anthropic
    • bun run switch-provider.ts --list — show available profiles
    • bun run switch-provider.ts --current — show active configuration
    • bun run switch-provider.ts --researchers — show researcher routing status

Multi-Provider Research

  • --multi-research flag: Routes research agents to their native providers for diverse perspectives
    • GeminiResearcher → google/gemini-2.5-flash
    • GrokResearcher → xai/grok-4-1-fast
    • PerplexityResearcher → perplexity/sonar
    • PerplexityProResearcher → perplexity/sonar-pro
    • CodexResearcher → openrouter/openai/gpt-4.1
  • researchers.yaml: Native researcher-to-provider mapping configuration
  • Graceful fallback: Missing API keys → researcher uses primary provider instead
  • User-driven opt-in: No automatic detection — user decides via --multi-research flag

Installation Wizard Updates

  • New Step 1b: "Research Agent Configuration" — asks user to choose single or multi-provider research
  • Profile-based generation: Wizard now uses applyProfile() from switch-provider.ts (single source of truth)
  • Research mode display: Success screen shows research configuration status

Changed

Provider Profile Models (Verified from anomalyco/opencode source)

Profile Most Capable Standard Budget
Anthropic anthropic/claude-opus-4-6 anthropic/claude-sonnet-4-5 anthropic/claude-haiku-4-5
OpenAI openai/gpt-5.1 openai/gpt-4.1 openai/gpt-4.1-mini
Google google/gemini-2.5-pro google/gemini-2.5-flash google/gemini-2.0-flash-lite
ZEN opencode/big-pickle opencode/kimi-k2.5-free opencode/gpt-5-nano
Local ollama/qwen2.5-coder:32b ollama/qwen2.5-coder:7b ollama/qwen2.5-coder:1.5b

Documentation Overhaul

  • README.md: Updated Quick Start, research agent models, provider switching docs
  • INSTALL.md: Added "Existing OpenCode Users" section addressing symlink workflow (fixes #14)
  • INSTALL.md: Replaced outdated "Option A/B/C" API Configuration with profile-based switching
  • INSTALL.md: Updated API Keys table with current model names

Fixed

  • ZEN profile: Replaced non-free models (opencode/claude-sonnet-4-5) with actual free models
  • OpenAI profile: Updated from deprecated gpt-4o/gpt-4o-mini to gpt-5.1/gpt-4.1/gpt-4.1-mini
  • Google profile: Added proper 3-tier (was all gemini-2.5-flash), uses gemini-2.0-flash-lite for budget
  • Local profile: Added guidance comments for Ollama users on which models to pull
  • switch-provider.ts: Module export guard prevents CLI execution when imported by wizard

Technical Details

  • Profile Format: YAML files in .opencode/profiles/ with provider/model format
  • Researcher Overlay: researchers.yaml defines native model + required API key per researcher
  • API Key Detection: Reads ~/.opencode/.env to check for available provider keys
  • Settings Tracking: settings.json records multiResearch state and active profile

[1.2.0] - 2026-02-05

Major Feature: Real-Time Observability Dashboard

This release introduces a complete monitoring infrastructure for PAI-OpenCode with real-time event streaming, SQLite persistence, and a Vue 3 dashboard.

Added

Observability Server

  • Bun HTTP Server on port 8889 with REST API and SSE streaming
  • SQLite Database for event persistence with 30-day retention
  • 14 Event Types captured across all plugin hooks
  • Real-time SSE Stream at /api/events/stream

API Endpoints

Endpoint Method Purpose
/health GET Health check with server stats
/events POST Event ingestion from plugins
/api/events GET Query events with filters
/api/events/stream GET SSE real-time stream
/api/sessions GET Query sessions
/api/stats GET Aggregated statistics

Vue 3 Dashboard

  • Dashboard Page: Real-time stats cards + live event stream
  • Events Page: Searchable/filterable event browser with pagination
  • Sessions Page: Session list with expandable event details
  • GitHub Dark Theme: Professional #0d1117 color scheme
  • SSE Connection: Live updates with pause/resume and reconnect

New Handler

Handler Purpose
observability-emitter.ts Fire-and-forget event emission to observability server

Event Types Captured

  • Session lifecycle: session.start, session.end
  • Tool execution: tool.execute, tool.blocked
  • Security: security.block, security.warn
  • Messages: message.user, message.assistant
  • Ratings: rating.explicit, rating.implicit
  • Agents: agent.spawn, agent.complete
  • Voice: voice.sent
  • Learning: learning.captured
  • Validation: isc.validated, context.loaded

Technical Details

  • Server Stack: Bun HTTP + SQLite (bun:sqlite)
  • Dashboard Stack: Vue 3.4 + Vite 5 + Tailwind CSS 3.4 + TypeScript
  • Plugin Integration: 82 new lines in pai-unified.ts
  • Event Emission: 1s timeout, fail silently (non-blocking)

File Structure

.opencode/observability-server/
├── server.ts          # HTTP server (:8889)
├── db.ts              # SQLite operations
├── README.md          # Documentation
└── dashboard/         # Vue 3 SPA
    ├── src/components/  # StatsCards, EventStream, EventList, SessionList
    ├── src/pages/       # Dashboard, Events, Sessions
    └── [config]         # Vite, Tailwind, TypeScript

[1.1.0] - 2026-02-02

Major Upgrade: PAI 2.5 + Voice/Sentiment Handlers

This release brings full PAI 2.5 Algorithm compatibility and adds 5 new handlers for voice notifications, sentiment detection, and observability.

Added

PAI 2.5 Algorithm Core

  • Full 7-phase Algorithm (v0.2.25): OBSERVE, THINK, PLAN, BUILD, EXECUTE, VERIFY, LEARN
  • ISC Validator with TaskCreate/TaskList for verifiable criteria
  • Capability Selection with Thinking Tools Assessment in THINK phase
  • Two-Pass capability selection: Hook hints (Pass 1) + THINK validation (Pass 2)
  • Parallel-by-default execution: Independent tasks run concurrently
  • Justify-exclusion principle: Thinking tools are opt-OUT, not opt-IN

New Handlers (v1.1)

Handler Purpose
voice-notification.ts TTS via ElevenLabs Voice Server, Google Cloud TTS, or macOS say fallback
implicit-sentiment.ts Automatic satisfaction detection from natural language (uses Haiku inference)
tab-state.ts Updates Kitty terminal tab title and color based on task context
update-counts.ts Counts skills, workflows, plugins, signals at session end
response-capture.ts ISC extraction, satisfaction tracking, learning capture

Support Libraries

  • lib/time.ts - ISO timestamps, PST timestamps, year-month formatting

Changed

  • Upgraded from PAI 2.4 to PAI 2.5 Algorithm
  • Plugin system now has 13 handlers (up from 8)
  • Enhanced SKILL.md with full Algorithm v0.2.25 documentation

Technical Details

  • Build: 21 modules, 85.77 KB total
  • All handlers integrated in pai-unified.ts
  • Graceful fallbacks: Voice handlers fail silently if services unavailable

[1.0.1] - 2026-02-01

Fixed

  • Anthropic Max Subscription API blocking workaround
  • ISCValidator integration improvements

[1.0.0] - 2026-01-24

Initial Release: Core PAI on OpenCode

The complete port of Daniel Miessler's PAI to OpenCode.

Added

Core Systems

  • Skills System: 29 skills (CORE, Algorithm, Fabric, Research, Art, etc.)
  • Agent System: 14 agents with PascalCase naming
  • Memory System: Projects, sessions, learning loops
  • Plugin System: Security validator, context loader

Plugin Handlers (v1.0)

Handler Purpose
context-loader.ts Loads CORE context at session start
security-validator.ts Blocks dangerous commands
rating-capture.ts Captures user ratings (1-10)
isc-validator.ts Validates ISC criteria
learning-capture.ts Saves learnings to MEMORY
work-tracker.ts Tracks work sessions
skill-restore.ts Restores skill context
agent-capture.ts Captures agent outputs

Installation

  • Legacy interactive installer (removed; CLI installer is canonical)
  • 8 AI providers supported (Anthropic, OpenAI, Google, Groq, AWS Bedrock, Azure, ZEN, Ollama)
  • TELOS personalization framework

Documentation

  • 7 Architecture Decision Records (ADRs)
  • Complete migration guide from Claude Code PAI
  • Plugin development documentation

Architecture Decisions

ADR Decision
ADR-001 Hooks → Plugins architecture
ADR-002 .claude/ → .opencode/ directory
ADR-003 Skills system unchanged
ADR-004 File-based plugin logging
ADR-005 Dual config files approach
ADR-006 Security patterns preserved
ADR-007 Memory structure preserved

Version Comparison

Feature v1.0.0 v1.1.0 v1.2.0 v1.2.1 v1.3.0 v2.0.0 v3.0.0
PAI Version 2.4 2.5 2.5 2.5 2.5 3.0 3.0
Algorithm Basic Full 7-phase Full 7-phase Full 7-phase Full 7-phase v1.8.0 v1.8.0
Handlers 8 13 13 13 13 13 16
Agents 14 14 14 18 15 (cleaned) 15 16
Dynamic Tier Routing No No No No Yes Yes Yes
Provider Profiles No No No Yes (5) Yes (6) Yes (6) Yes (6)
Multi-Provider Research No No No Yes Yes Yes Yes
Observability Dashboard No No Yes Yes Yes Yes Yes
Voice Notifications No Yes Yes Yes Yes Yes Yes
Sentiment Detection No Yes Yes Yes Yes Yes Yes
Image Optimization No No No No 79% reduction 79% reduction 79% reduction
Wisdom Frames No No No No No Yes (5 domains) Yes (5 domains)
Verify Completion Gate No No No No No Yes Yes
Effort-Scaled Gates No No No No No Yes Yes
DB Health Tooling No No No No No No Yes
Installer (CLI-only) No No No No No No Yes
v2→v3 Migration No No No No No No Yes
Security Hardening No No No No No No Full

Upgrade Path

From v2.x to v3.0.0 (Breaking Changes)

Before you start: The v3.0.0 release has significant breaking changes:

  • Skills structure: flat → hierarchical (Category/Skill)
  • Config: single-file → dual-file (opencode.json + settings.json)
  • Paths: .claude/ → .opencode/
  • New CLI-only installer

Recommended upgrade process:

  1. Backup your existing installation:

    cp -r ~/.opencode ~/.opencode-backup-$(date +%Y%m%d)

  2. Run the migration tool (dry-run first):

    bun Tools/migration-v2-to-v3.ts --dry-run

  3. Review the migration report, then execute:

    bun Tools/migration-v2-to-v3.ts

  4. Alternative: Fresh install (recommended for a clean setup):

     bash PAI-Install/install.sh

See UPGRADE.md for detailed step-by-step instructions.

From v1.2.x to v1.3.0

git fetch origin
git checkout main
git pull origin main
 bash PAI-Install/install.sh --update

Re-running the installer is recommended — it generates the new profile format with dynamic tier routing.

Manual alternative: bun run .opencode/tools/switch-provider.ts zen-paid

Voice Server Setup (Optional)

To enable voice notifications:

  1. Start the included voice server: 
    cd .opencode/voice-server && bun run server.ts
  2. Configure TTS provider in .opencode/.env:
    • ElevenLabs: ELEVENLABS_API_KEY=your_key
    • Google Cloud TTS: GOOGLE_API_KEY=your_key
  3. Fallback: macOS say command works automatically

See .opencode/voice-server/README.md for full documentation.

Links: