Restart OpenCode. The plugin hooks and tools are auto-loaded.
The plugin stores runtime state at ~/.parallax/ (plugin process runs from user home directory).
What It Does
Two Agent Modes
The plugin provides two agent tabs, switch between them with [Tab]:
Agent
Mode
When to Use
Parallax
Plan / Build / Debug
Interactive reasoning with protocol enforcement
Horizon
Autonomous Supervisor
Multi-hour/multi-day unattended execution
PARALLAX AGENT -- Protocol Enforcement (6 steps)
The Parallax agent follows a structured reasoning protocol before writing code. The plugin enforces this via the tool.execute.before hook:
Step
Checkin
What it blocks
1. Ambiguity Check
parallax_checkin("ambiguity")
All writes until done
2. 4 Invariants
parallax_checkin("invariants")
Warns after 3 writes without checkin
3. Verification Gate
parallax_checkin("gate")
Requires invariants first
4. Design Doc (opt-in)
parallax_checkin("design")
Per project via config
5. Commit Decision
parallax_checkin("commit")
Any time after gate
6. Summary
parallax_checkin("summary")
Generates retrospective
Mode Switching
Mode
Tool
When
PLAN
parallax_plan
Vague requirements, before writing code
BUILD
parallax_build
Executing, writing code (default)
DEBUG
parallax_debug
Post-build quality and security audit
Friction Loop
After every write, the plugin auto-runs project verification. On failure, it decrements a retry counter. After 3 consecutive failures, it blocks further writes until the issue is resolved.
Trace Recording
Every session produces a structured reasoning trace -- phases, writes, verification results. The AI can export it to JSON or generate a PR-ready markdown summary.
The Horizon agent is a self-driving project supervisor for multi-hour/multi-day tasks. It plans, researches, executes, self-tests, and self-iterates until a goal is 100% complete without needing mid-execution user input.
Workflow
Phase
What Happens
1. RESEARCH
Parses goal, web searches best practices, analyzes codebase, synthesizes findings
2. PLAN
Decomposes goal into milestones + features with acceptance criteria and complexity estimates
3. EXECUTE LOOP
Dispatches sub-agents, runs tests, self-evaluates across 6 dimensions, retries up to 3x per feature
4. FINAL AUDIT
Runs Universal Auditor, full test suite, exports traces, generates completion report
Self-Check Evaluation (6 Dimensions)
After every sub-agent, Horizon scores output across 6 weighted dimensions:
Dimension
Weight
Check
Protocol Integrity
15%
All Parallax steps completed?
Verification
25%
Tests pass? No lint errors?
Correctness
25%
Matches acceptance criteria?
Design Quality
15%
AI slop detected?
Edge Case Coverage
10%
Error paths handled?
User Perspective
10%
Intuitive for all skill levels?
Pass threshold: >= 75% weighted score. Automatic retry on failure (up to 3 cycles).
Autonomous Decision Engine
When Horizon encounters ambiguity, it autonomously: