Repository navigation
Tool Arsenal
Dipanshu Kamboj edited this page Aug 31, 2026
·
1 revision
Every tool the agent can call, in one place. Tools mount as plugins — hide or add them
per profile via cordis.patch.yml without touching code.
| Tool | What it does |
|---|---|
bash |
One-shot bash -c with bounded output, spill-to-file, background jobs, timeout + SIGTERM→SIGKILL escalation. Fresh shell per call. |
pwsh |
Same shape for PowerShell (pwsh -NoLogo -NoProfile -NonInteractive). |
terminal_open / terminal_send / terminal_read / terminal_signal / terminal_list / terminal_close |
A persistent PTY session: real interactive TTY (prompts, ctrl-c, vi, ssh…) with scrollback you can re-read. |
Persistent bash/pwsh shells |
The persistent-shell tools keep cwd and exported env across calls — for work that genuinely needs state. |
job_output / job_list / job_kill
|
One controller for every background job: background commands, PTY sends, subagents. Read, list, kill. |
workflow / ralph
|
Multi-round scripted agent loops — one fresh structured child per round, parented by the caller. |
| Tool | What it does |
|---|---|
read / write / edit / str_replace_editor
|
Read, write, and patch files; every op is audited and policy-gated. |
read_image |
Attach an image file for vision models. |
glob / grep
|
Path and content search over the workspace. |
| Tool | What it does |
|---|---|
web_fetch |
Fetch a URL (bounded, robots-aware) with any mounted fetch backend. |
web_search |
Search the web through the mounted provider (DeepSeek/Exa/Perplexity backends are swappable behind the seam). |
| Tool | What it does |
|---|---|
lsp |
Language-server queries (hover, definitions, diagnostics…) through mounted LSP providers. |
ask_user_question |
Ask the human a blocking question from inside the session — answers land as a user message. |
todo_write |
Session-owned plan checklist the UI renders live. |
skill |
Load a skill into context on demand — one reusable, self-contained instruction pack. |
create_goal / get_goal / update_goal
|
Durable goal state: create, pause, resume, block, complete. |
| Tool | What it does |
|---|---|
session_search / session_trace / session_event_read / session_event_search / session_event_trace
|
Five read-only tools to search past sessions, trace events, and mine what an earlier agent did. |
| Tool | What it does |
|---|---|
subagent / subagent_fork
|
Delegate self-contained work to a fresh child agent (continuable or one-shot), foreground or background. |
send_message / interrupt_agent / list_agents
|
Drive live children: message them, interrupt, list. |
report |
The child's channel to hand a result back to its parent. |
spawn_teammate / wait_agent / followup_task
|
Agent Teams (experimental): durable teammates with a shared Lead journal, wake/quiet delivery, mailbox recovery. |
team_task_create / team_task_get / team_task_list / team_task_update
|
Team task board with compare-and-set updates. |
| Tool | What it does |
|---|---|
run_code |
Execute code (TypeScript/Python) with direct typed access to the harness's own tool API. |
cordis_define / cordis_run / cord_inspect_* / cordis_stop / cordis_undefine
|
Hot-load and drive new plugins at runtime — the agent can extend the harness itself. |
schedule_create / schedule_list / schedule_delete
|
Time-based autonomous triggers: after N seconds, at an absolute time, or fixed-rate. |
exit_plan_mode |
Close a plan-mode turn and hand the plan back for approval. |
The pattern: every tool talks to a capability seam (ctx.shell, ctx.fs, ctx.web,
ctx.lsp, ctx.subagents…), never directly to an implementation. Swap the backend —
web-fetch-http → a new provider, bash-local → your executor — and the model-facing
schema stays identical. That's what makes this thing composable.
Browse exact schemas: Tool catalog