CLI agent router
First honor the user's selected agent. Otherwise choose by verified local capability, authentication, tool access and task scope. Use configured model defaults unless the task specifies a model; check current CLI help instead of guessing model IDs, context sizes, prices or flags.
Route the work
| Need | Route |
|---|---|
| Small change within the current agent's abilities | Complete it here |
| Independent review with native host subagents available | Use the host's review role when delegation is authorized |
| Explicit Codex execution | Read codex; check codex --version and codex exec --help |
| Explicit Claude execution | Check claude --version and claude --help; inspect local authentication status |
| Provider/MCP tools configured in OpenCode | Read opencode-cli; verify the actual provider and tools |
| Another installed CLI, such as Gemini | Inspect its local help and authentication; no assumed skill or free quota |
| Repeated repair with acceptance commands | Read bounded-agent-loop |
| Persistent terminal or isolated multi-worktree pilot | Read terminal workflows |
Availability is not readiness: a command can exist but lack credentials, a provider, network access or compatible filesystem paths. Never copy auth files through this repository. Never print credentials during readiness checks.
PowerShell: Get-Command codex,claude,gemini,opencode -ErrorAction SilentlyContinue.
POSIX: command -v codex claude gemini opencode. In WSL, inspect resolved shims:
a Windows executable may require Windows paths and use Windows authentication.
On Debian/Linux the shims are native executables, so no Windows-path
translation applies.
A successful version command does not prove its working directory or tools work.
Delegate explicitly
Give the worker a concrete task, root/branch, instructions, exact file ownership,
acceptance command, deadline and output format. Separate read-only exploration
from implementation. Reuse the portable prompts in bounded-agent-loop or
cavecrew; these prompt roles are not universally registered agent names.
Use separate worktrees for independent writers and one integrator. Do not run multiple writers against the same checkout. Serialize dependent work. Preserve the worker's actual exit, stderr and test evidence; a completion sentence alone is not proof. Review its diff and run acceptance independently.
Keep argument arrays as arrays; do not interpolate user prompts into shell code.
Use stdin or a private prompt file when supported. Give each run a unique local
log directory. Background PowerShell helpers use Start-Process -WindowStyle Hidden (on Debian/Linux, detach with nohup <cmd> & or setsid <cmd>);
interactive windows are only for a user-requested visible session.
Stop and hand off
Apply only the permissions required by the authorized task. Do not silently
enable permission bypass, autopilot, schedules, automatic commits or merging.
A preauthorized edit does not need another confirmation at each iteration.
On unavailable tools, cancellation, repeated failure or exhausted budget, record
the exact limitation and remaining work using session-handoff and
cross-harness-state. Resume a specific task/session after checking its root,
branch and remaining budget; never choose an unrelated "last" session.