opencode-driven-development

Imported from zaycruz/hermes-opencode-plugin (SKILL.md). Install upstream with npx skills add zaycruz/hermes-opencode-plugin. Copyright stays with the author (MIT).

OpenCode-Driven Development

The Mental Model

Hermes already has delegate_task for spawning subagents. But those are just Hermes clones with basic tools — one agent, one model, sequential execution.

OpenCode is a meta-subagent. One opencode call spawns an entire engineering team:

You (Hermes) — brain, memory, planning, user-facing
  │
  ├── read_file, terminal, web_search     ← do it yourself (quick stuff)
  ├── delegate_task                        ← simple Hermes subagent (basic subtasks)
  └── opencode                             ← OMO meta-subagent (real engineering)
        └── Sisyphus auto-orchestrates:
              ├── Hephaestus  (deep implementation, parallel)
              ├── Oracle      (architecture advice)
              ├── Librarian   (docs research, parallel)
              ├── Explore     (codebase grep, parallel)
              └── ...5+ agents running simultaneously

One opencode call can trigger dozens of parallel tool executions across multiple models. It has LSP, AST-grep, hash-anchored edits, background agents — things you don't have.

This is how you lighten your cognitive load. Don't manage 50 tool calls to implement a feature. Hand it off to OpenCode and focus on the bigger picture.

The Decision Framework

Every task, ask yourself:

Do it yourself (read_file, terminal, web_search)

• Reading a file to understand something
• Running a quick command
• Looking something up
• Answering a question
• Single-line fix you already know

delegate_task (Hermes subagent)

• Simple focused subtask that needs basic file/terminal
• Tasks where you want to control the exact toolset
• Non-coding work (research, analysis, writing)

opencode (OMO meta-subagent)

• Multi-file implementation — feature that touches 3+ files
• Refactoring — rename, restructure, move code across files
• Bug fixing — needs investigation, testing, verification
• Test writing — needs to run tests and iterate
• Any task where you'd make 10+ tool calls — offload it
• Anything where parallel agents help — OMO fires 5 agents at once
• LSP-precision work — workspace rename, find all references

Rule of thumb: If you'd need more than 3-4 tool calls, use opencode. The overhead of one opencode dispatch is cheaper than managing the complexity yourself.

How OMO Amplifies You

When you dispatch to OpenCode, here's what happens inside:

1. Sisyphus receives your task, analyzes it
2. Breaks it into subtasks with dependency ordering
3. Wave 1: fires independent subtasks in parallel as background agents
   • Hephaestus implements component A
   • Librarian researches the API docs
   • Explore greps the codebase for patterns
4. Wave 2: dependent tasks fire when Wave 1 completes
5. Results aggregate back — Sisyphus synthesizes
6. You get structured output: text summary, file diffs, tool call count

What you DON'T have to do:

• Manage parallel execution
• Pick the right model for each subtask (OMO's category routing handles it)
• Handle LSP operations (rename, references, diagnostics)
• Deal with hash-anchored edits (zero stale-line errors)
• Retry failed subtasks (OMO has fallback chains)

Dispatch Patterns

Offload and Move On

The most common pattern. Hand off the work, move to the next thing:

opencode(
    action="run",
    prompt="Implement user authentication with JWT. Files: src/auth/, tests/auth/. Use bcrypt for passwords, 15min token expiry. Run pytest when done.",
    directory="/project"
)

While OpenCode's agents are working, you can:

• Update your todo list
• Check memory for the next task
• Communicate with the user
• Plan the next dispatch

Parallel Offload

Multiple independent tasks — fire them all:

# These hit different parts of the codebase — safe to parallelize
opencode(action="run", prompt="Add rate limiting to API routes in src/api/", directory="/project")
opencode(action="run", prompt="Write migration for new user_preferences table", directory="/project")
opencode(action="run", prompt="Add Stripe webhook handler in src/webhooks/", directory="/project")

Each one spawns its own Sisyphus → parallel agents. You just kicked off potentially 15+ agents with 3 tool calls.

Heavy Lift

For the big stuff — give Hephaestus full autonomy:

opencode(
    action="run",
    agent="hephaestus",
    prompt="Refactor the monolithic OrderService into separate services: OrderCreation, OrderFulfillment, OrderNotification. Maintain all existing tests. Add new tests for service boundaries.",
    directory="/project"
)

Hephaestus will explore the codebase, understand the patterns, plan the refactor, execute it, and verify tests — all autonomously.

Scout Then Strike

Use OpenCode for recon, then for execution:

# First: get a plan from Prometheus
opencode(
    action="run",
    agent="prometheus",
    prompt="Analyze this codebase and create a plan for adding GraphQL support alongside the existing REST API."
)

# Review the plan, adjust if needed, then execute
opencode(
    action="run",
    prompt="Execute the GraphQL integration plan. [paste relevant parts or reference the plan file]"
)

Iterative Session

For work that needs your feedback between rounds:

# Round 1
result = opencode(action="run", prompt="Implement the dashboard page per the wireframe in docs/dashboard.png", directory="/project")

# You review: looks good but needs responsive layout
opencode(
    action="session",
    session_id=result.session.id,
    prompt="Make the dashboard responsive. Mobile: stack cards vertically. Tablet: 2-column grid. Desktop: current 3-column layout."
)

Injecting Your Context

You have memory. OpenCode doesn't. Bridge the gap:

opencode(
    action="run",
    prompt="""Fix the flaky test in tests/api/test_orders.py::test_concurrent_checkout.

    CONTEXT FROM MEMORY:
    - This project uses PostgreSQL with row-level locking for inventory
    - Previous fix attempt (2 weeks ago) tried adding retry logic but it masked the real issue
    - User prefers explicit locks over optimistic concurrency
    - Test environment uses SQLite (this may be the root cause)
    """,
    directory="/project"
)

Before every dispatch, ask: Do I know something from memory that would help OpenCode do this right the first time? If yes, include it.

What You Do After Each Dispatch

1. Check status — completed or error?
2. Read the summary — what did it actually do?
3. Verify file changes — expected files modified?
4. Update memory — anything learned? conventions discovered? decisions made?
5. Update todos — mark complete, add follow-ups if needed
6. Tell the user — concise summary of what was done

Integration with Hermes Systems

Memory + OpenCode

• Inject project conventions into prompts → OpenCode codes your user's way
• After results, save new patterns/decisions to memory → future dispatches are smarter
• Memory compounds: each task teaches Hermes something that makes the next dispatch better

Cron + OpenCode

• every morning: "Run full test suite in /project, report failures"
• every friday: "Check for outdated npm packages, create upgrade PR if safe"
• every 2h: "Check CI status, auto-fix lint errors on failing builds"

Cross-Platform + OpenCode

• User on Telegram: "fix the login bug"
• You: dispatch to opencode, get results, reply on Telegram with summary
• User never knows the implementation details — they just see it's fixed

delegate_task + OpenCode

You can even use both together:

• delegate_task for non-coding work (research, drafting docs, analysis)
• opencode for the actual engineering
• Both running in parallel, you coordinating

Agent Selection Cheat Sheet

Red Flags

• Over-dispatching: Don't use opencode to read a file. Use read_file.
• Under-specifying: "Fix the bug" → bad. "Fix the null pointer in src/api/users.py:47 when email is missing" → good.
• Ignoring results: Always read what came back. Don't fire-and-forget without checking.
• Micromanaging: Don't break a task into 20 tiny opencode calls. Give Sisyphus a meaty task and let it orchestrate.
• Skipping memory: If you know something relevant, inject it. OpenCode is stateless — you're its memory.
• Forgetting verification: Always include "run tests" or equivalent in the prompt.

The Bottom Line

Small task (< 3 tool calls)  →  do it yourself
Medium task (basic subtask)  →  delegate_task
Heavy task (real engineering) →  opencode

You think. OpenCode builds.
You remember. OpenCode executes.
You plan. OpenCode parallelizes.