Coding Agent Delegation

Coding Agent Delegation

Class-level guide for using external autonomous coding agents from Hermes while Hermes keeps ownership of scope, verification, and final claims.

Core rule

External agents are implementation lanes, not authorities. Give them bounded prompts, collect concrete artifacts, then verify with Hermes tools before telling the user work is complete.

Agent choices

Claude Code

Use when the user explicitly requests Claude Code, when Anthropic's coding agent is already installed/authenticated, or when a task benefits from a strong autonomous repo-editing loop.

Pattern:

1. Create a self-contained prompt with goal, files, constraints, and verification command.
2. Run via terminal in PTY/background when interactive or long-running.
3. Require a final summary with changed paths, commands run, and blockers.
4. Verify diffs and tests yourself.

Codex

Use when the user asks for Codex/OpenAI coding agent work, when you want an independent implementation lane, or when a Kanban worker needs a lightweight CLI implementer.

Pattern mirrors Claude Code: bounded prompt, isolated cwd/worktree when possible, final artifact handle, then independent verification.

OpenCode

Use when the user explicitly wants OpenCode or wants an alternate coding agent for implementation/refactor/review. Treat it as a peer lane: never assume its self-report is true without reading files and running verification.

Kanban Codex lane

Use when a Hermes Kanban worker wants Codex as a narrowly scoped implementation lane while Hermes keeps task lifecycle, reconciliation, testing, and handoff. The lane prompt should preserve Kanban ownership boundaries: Codex implements; Hermes reconciles, tests, and updates state.

Template: templates/kanban-codex-lane-prompt.md.

Prompt contract

Every delegated coding prompt should include:

• repository path and branch/worktree context
• exact objective and non-goals
• allowed files or directories
• expected verification command(s)
• output format: changed files, commands run, failures/blockers, handoff notes

Verification checklist

• Inspect git diff and touched files.
• Run the smallest meaningful test/build/lint command.
• If the agent claims an external side effect, verify the URL/ID/status yourself.
• If verification fails, either fix directly or send a precise follow-up prompt with the failing output.

Common pitfalls

• Do not let an external agent choose broad architecture without user-approved scope.
• Do not run long bounded jobs without notify_on_complete=true.
• Do not archive the external agent's failure as success; report blockers directly.
• Do not mix multiple agent lanes in the same dirty worktree unless changes are trivially separable.