Harness support
cones discovers native sessions, reads what each harness reports and runs the harness's own operations. Missing state, context and usage stay absent. It installs no global session hooks. Five harnesses have report readers; six more have experimental terminal adapters. Full reference, with the versions each path was checked against: harness.md.
Capabilities
"Owned" means a client in a terminal cones hosts. "Others" are Gemini, Cursor Agent, Copilot, Amp, Droid and Kimi.
| Verb | Claude Code | Codex | pi | OpenCode | Grok | Others |
|---|---|---|---|---|---|---|
ls | ✓ | ✓ | ✓ | ✓ | ✓ | Process only |
show, search | ✓ | ✓ | ✓ | ✓ | ✓ | Copilot, Cursor; owned screens |
launch | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
resume | ✓ | ✓ | ✓ | ✓ | ✓ | Copilot, Cursor |
fork | ✓ | ✓ | ✓ | ✓ | ✓ | – |
send | ✓ | Daemon | Owned | Owned | – | – |
wait | ✓ | ✓ | ✓ | Owned | ✓ | – |
rename | ✓ | ✓ | Owned | Owned | – | – |
btw | ✓ | ✓ | ✓ | ✓ | – | – |
compact | Background | ✓ | Owned | Owned | – | – |
stop | Background | Daemon | Owned | Owned | Owned | Owned |
rm | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
| Join from the dashboard | Background | Daemon | Owned | Owned | Owned | Owned |
wait needs a reported state, so a session with none never settles. OpenCode 2 native forks are refused.
Discovery
| Harness | Native home | Live source |
|---|---|---|
| Claude Code | $CLAUDE_CONFIG_DIR, otherwise ~/.claude | Session registry |
| Codex | $CODEX_HOME, otherwise .codex beside the Claude home | Processes, writer locks, saved threads |
| pi | $PI_CODING_AGENT_DIR, otherwise .pi/agent beside the Claude home | Processes and session files |
| OpenCode | $XDG_DATA_HOME/opencode, otherwise ~/.local/share/opencode | Processes, owned-client reports, SQLite |
| Grok | $GROK_HOME, otherwise .grok beside the Claude home | Processes and the session folder each holds open |
Extra homes can be listed under homes in jobs.yaml. A headless run such as claude -p or codex exec has no client to join and is read like any other external session.
State
| Harness | Where state comes from |
|---|---|
| Claude Code | The registry's own status: busy, shell, waiting, idle. Background sessions add job records. |
| Codex | Daemon threads: native runtime state, including approvals and questions. Standalone clients: the rollout's task events only, so an approval wait reads as working. |
| pi | The last entry in the session file. Owned clients also report idle, working and dialogs. |
| OpenCode | Owned clients only, through the reporter. External process rows show -. |
| Grok | The last turn event in the session it holds open. |
Quiet output never changes a reported state. Only Claude Code reports a finished turn with a background shell still running.
Context, tokens and cost
Every column is read from what the harness writes: transcripts, rollouts, session files or databases.
- Context window: Codex and Grok report it. Claude Code reports it only to its status line (see below). pi and OpenCode store none.
- Cost: a native session total wins when the harness writes one (Claude Code status line, OpenCode, Grok). Otherwise cones prices the harness's reported usage with a cached models.dev catalog and shows
~$…. Partial coverage shows only the priced subtotal; unavailable values show-. Estimates do not reconcile subscriptions, discounts or invoices. Claude Code replies served through Bedrock or Vertex need the saved status-line cost. - Plan: Claude Code shows the signed-in account's plan or the endpoint (bedrock, vertex); Codex its ChatGPT plan or provider; pi and OpenCode their provider.
Context window for Claude Code
Only Claude Code's status line receives the context window size. To let cones read it, add this after input=$(cat) in your statusLine command:
claude_dir=${CLAUDE_CONFIG_DIR:-$HOME/.claude}
mkdir -p "$claude_dir/statusline"
printf '%s' "$input" > "$claude_dir/statusline/$(jq -r .session_id <<<"$input").json"
Without it, context shows prompt tokens only. cones does not modify Claude settings or your statusLine command.
Message delivery
cones sends a note only through a channel the harness provides, and never types a note into a terminal.
| Harness | Channel |
|---|---|
| Claude Code | The cross-session inbox Claude's own SendMessage uses. A session that bypasses permission prompts holds the note for its user's approval unless "crossSessionInbound": "accept" is set. |
| Codex | queue --thread on the daemon that owns the thread. |
| pi, OpenCode | The reporter cones loads into clients it launched. Clients in other terminals are refused. |
| Grok and others | None found; sending is refused. |
Permissions
Sessions the composer starts skip permission prompts by default, with each harness's own flag. skip_permissions: false starts every harness with its configured permissions; a per-harness key overrides it. Scheduled jobs always run without prompts.
| Harness | Key | Native flag |
|---|---|---|
| Claude Code | claude_skip_permissions | --dangerously-skip-permissions |
| Codex | codex_skip_permissions | --dangerously-bypass-approvals-and-sandbox |
| OpenCode | opencode_skip_permissions | --auto |
| Gemini | gemini_skip_permissions | --yolo |
| Cursor Agent | cursor_skip_permissions | --force |
| Copilot | copilot_skip_permissions | --allow-all |
| Droid | droid_skip_permissions | --auto high |
| Kimi | kimi_skip_permissions | --auto |
| Grok | grok_skip_permissions | --always-approve |
pi has no permission prompts, and Amp has no flag for skipping them.
What cones writes outside its own state
cones keeps its own state in ~/.cones. It installs no global session hooks. These are the other places it writes:
| Write | When |
|---|---|
skills/cones/SKILL.md in the Claude Code home and ~/.agents/skills | Dashboard start. A copy you edited or deleted is left alone. |
Folder trust in Claude's .claude.json and Codex's config | Before a launch into an untrusted folder; the same record the harness's own trust dialog writes. Grok launches pass --trust. |
| A reporting extension loaded into pi and OpenCode clients | Clients cones launches. It changes no tools or permissions. |
~/.codex-<provider> | First launch on a Codex provider other than the default home's. |
~/Library/LaunchAgents/local.cones.*.plist | Saving a scheduled job. |
A custom-title line in a Claude transcript | Renaming an external interactive Claude session by hand. |
Before a start, fork or resume, cones runs your $SHELL -l -i in the folder to pick up its environment, so rc files and direnv apply as they would for a harness you typed there.
Experimental terminal harnesses
Gemini, Cursor Agent, Copilot, Amp, Droid and Kimi get launch, process discovery, ctrl+z return, viewer reuse and stop. They have no external-terminal join, message delivery or scheduled jobs. State, model, context, tokens, cost and last reply stay absent. Copilot and Cursor Agent conversations appear in history and can be resumed.
Gemini, Cursor Agent, Copilot, Amp and Droid were each launched with no login and no network, then named, listed, reopened and stopped. None got past its trust or sign-in screen, so whether it takes the instruction is unverified. Kimi starts only without an instruction. Model-backed turns, permission questions and opening a Copilot or Cursor history row in the dashboard are unverified.