Orchestration Workflow

A multi-agent orchestration workflow for OpenCode (v1 & v2): the main session acts as a lead — it plans a task-specific workflow, delegates to specialized subagents running in parallel, adversarially verifies every finding, and applies only what survives. Works with any LLM provider.

Quick Start

git clone https://github.com/itohnobue/orchestrator-opencode
cd orchestrator-opencode
./install.sh /path/to/your/project   # macOS/Linux
# or: .\install.ps1 C:\path\to\project   (Windows)

The installer copies .opencode/ (agents, tools, templates, skills, plugin), AGENTS.md, and a minimal opencode.json into your project — an existing AGENTS.md or opencode.json is never overwritten — and creates tmp/ for agent artifacts. Open the project with OpenCode and give it a task; the workflow activates automatically and runs autonomously: it displays the full plan and proceeds, without approval prompts.

Default allowance

The shipped opencode.json sets only permission: allow — no model pin: the model and provider come from your machine's global OpenCode config (~/.config/opencode/opencode.json). Edit the project opencode.json for a per-machine override. Reasoning effort is configured in the global config (agents do not pin their own).

How it works

• Three-agent planning pipeline (always) — agentic-planner researches the codebase and classifies the task (size, domains, ambiguity, severity, change type), then selects the workflow bricks; volume-splitter resolves file scopes to exact files with line counts and applies mechanical size caps; agent-organizer reviews the plan, fixes structural violations, and redistributes the questions each agent must answer. The final plan is displayed before any stage agent starts.
• Parallel execution from a custom manifest — the workflow is assembled per task from bricks (research, discovery, implementation, review, verification, fix, test), not a fixed skeleton. Independent subtasks run as parallel agents (up to 20 per batch, 25 when the stage fits in one batch); stages are sequential only when one consumes the previous stage's verified output.
• Research-defined specialists — specialist identity comes from research FOCUS angles, not static personas. Each agent receives exactly the research covering its scope: a compact digest in its prompt plus the full report path on demand. At MEDIUM+ severity, discovery and post-implementation review get a second opinion with a complementary standpoint — never the same lens twice.
• Adversarial verification — findings are extracted, deduplicated, and tagged, then routed by severity to falsification agents (CRITICAL 1:1, HIGH 1 per 3, MEDIUM 1 per 10) that try to disprove them with evidence; dismissed suspicions are re-examined too. Every finding ends CONFIRMED (fix list), REJECTED (dropped), or WEAKENED (downgraded) — the verified grid decides what gets fixed.
• Findings-driven convergence — discovery/review stages repeat only when the verified grid holds a triggering (un-folded) confirmed HIGH/CRITICAL finding; a HIGH+ that merely re-states an already-confirmed defect class (extension/corroboration) does not re-fire the loop (ceiling: LOOP default, ONCE opt-in). Fix passes repeat — fix → build-gate → post-fix review — until no confirmed code defects remain; regression tests pinning the fixes are added.
• Continuation handoff — a fixed 8-section carry-over so a fresh agent or later session resumes without re-discovery; the handoff skill retires subagent runs at the reuse threshold and checkpoints sessions on demand.
• Memory that survives — two-tier knowledge/session memory via memory.sh; confirmed findings are harvested into knowledge.md after each run.

The 12 agents

.opencode/agents/ — INDEX.md is the quick reference.

Agent	Role
agentic-planner	Researches the project, classifies the task, selects workflow bricks, produces the plan manifest (Research Coverage Map + Routing Table + per-agent tiers)
volume-splitter	Resolves file scopes to exact paths with line counts; applies mechanical split/merge rules
agent-organizer	Structural plan review: tiers, routing precision, FOCUS complementarity, MUST ANSWER redistribution
executor	The ONE generic executor — DISCOVER, IMPLEMENT, REVIEW, FIX, TEST, TEST-UPDATE, quick-fix, build-gate, final gate. PLAIN (task context as briefing) or researched (digest + full report path)
postfix-reviewer	Post-fix review ONLY (read-only) — verifies applied fixes against their design; verdict APPROVED / NEEDS-FIX
verification-analyst	Extraction + synthesis — dedup, confidence tags, verification grid
knowledge-harvester	Knowledge harvesting from verified findings — PATTERN/INCIDENT classification, dedup, prevention recommendations, supersede-evaluate; writes tmp/knowledge-harvest-report.md
adversarial-reviewer	Falsification gate — the single distinct quality gate; batch sizes CRITICAL (1:1), HIGH (1:3), MEDIUM (1:10) are volume controls
web-searcher	RESEARCH brick — internet research (standards, formats, versions, ecosystems, advisories)
research-analyst	RESEARCH brick — structured analysis/synthesis; mid-execution research
data-researcher	RESEARCH brick — dataset research
prepare-agent	(Single-session suite) research generation — full report + compact digest, FOCUS: defines the specialist

Single-session mode

The suite also ships a single-session-workflow skill that switches out of orchestrator mode on demand: the main session becomes the worker, delegating only when a subtask is big or context-heavy. Invoke it with "switch to single-session mode".

Requirements

• OpenCode CLI — V1 1.18.29+ or V2 2.0+
• At least one LLM provider configured in ~/.config/opencode/opencode.json
• uv — auto-installed repo-locally into tmp/uv/ if missing (never system-wide, per the AGENTS.md tool-use policy); handles Python dependencies for the tools

License

MIT