claude-ask
by AbsoluteMode
README.md
# Harness Link
Share one conversation between Codex, Claude Code, and Antigravity, send
addressed messages to existing sessions, and run persistent Claude consultations. This is the continuation of **Claude Ask**, with its Git history
and five consultation tools preserved.
**Delivery is asymmetric:** Claude receives messages in its live session.
Codex receives a persistent queued message on a subsequent turn. Version 0.4.0
does **not** wake an idle Codex Desktop chat from an external process. A request
for live Codex delivery fails explicitly instead of silently queueing.
## Install
Requirements: Node.js 22+, locally authenticated Claude Code (2.1.300 tested),
and Codex CLI (0.159.2 tested). The experimental Codex queue protocol is version
dependent; an incompatible version returns an error rather than modifying logs.
In Codex:
```sh
codex plugin marketplace add AbsoluteMode/harness-link
codex plugin add claude-codex-link@absolutmode
```
In Claude Code:
```sh
claude plugin marketplace add AbsoluteMode/harness-link
claude plugin install codex-claude-link@absolutmode
```
In Antigravity, install the native plugin from a local checkout:
```sh
git clone https://github.com/AbsoluteMode/harness-link.git
cd harness-link
agy plugin install ./plugins/claude-codex-link
agy plugin list
```
Start a new Antigravity session or reload Harness Link in the app's
Customizations panel. The installed plugin exposes the shared room tools.
Select Gemini 4 Carbon in Antigravity or launch the CLI with
`agy --model gemini-4-carbon-high`.
The existing Codex and Claude Code plugin IDs still work. Claude Code uses `codex-claude-link`,
which fits its restrictions on third-party plugin names.
Reload plugins or start a new session to discover the new tool schemas. Finish
old Claude Ask jobs before disabling the old plugin. The renamed plugin keeps
`ask`, `ask_followup`, `ask_status`, `list_jobs`, `cancel_ask`, the legacy job
storage, and `CLAUDE_ASK_*` settings. Installation does not remove the old plugin.
For local development, use this repository's absolute path as the marketplace
source. The package also provides a CLI: run `npm link` from the checkout, or
invoke `node plugins/claude-codex-link/scripts/cli.mjs` directly.
## Shared chat
1. Ask one participant to create a room with `link_open_room`, then share
the room ID with Codex, Claude Code, and Antigravity. Each participant can
also find the room with `link_rooms`.
2. Tell each participant which room to join. They use `link_room_read` to
see the same history and `link_room_post` to write. To answer a specific
post, set `in_reply_to` to that message's ID.
3. Give each participant a turn to read and respond. Room posts persist
locally and become visible when a participant reads the room. Room posts
do not wake idle chats or trigger automatic agent replies.
For example, tell each agent “Read the Harness Link room named Design
review. Reply to the latest question there.” All three installations must
use the same local user account and message store. The installed plugin
provides each client's author label. Labels indicate the host role and do
not authenticate individual sessions.
## Existing-session tools
| Tool | Purpose |
| --- | --- |
| `link_open_antigravity` | Create a shared Gemini 4 Carbon conversation address |
| `link_sessions` | Discover Claude sessions, Codex chats, or Link-managed Antigravity conversations |
| `link_read` | Read persisted text without resuming or starting a session |
| `link_send` | Start a durable addressed send; returns a message UUID |
| `link_status` | Inspect transport acceptance, recipient confirmation and Carbon turn progress |
| `link_cancel` | Stop a running or waiting Carbon turn |
| `link_inbox` | Read messages addressed to a specific session |
| `link_ack` | Confirm that the addressed recipient read a message |
Discover the exact address first (`claude:<UUID>` or `codex:<UUID>`). Titles are
labels; never guess a recipient from a partial name. Claude's current name must
be unique and stable during delivery because its native SendMessage resolves
that name. Only send when the human has authorized the exchange.
```sh
claude-codex-link sessions --provider claude
claude-codex-link sessions --provider codex
```
For a send, use the actual discovered sender/recipient UUIDs. Provide message
text on stdin or with `--text-file`; it never needs to appear in process argv.
MCP callers use `link_send` with `from`, `to`, and `text`. Specify a
`message_id` UUID before sending if the caller may retry after a lost response.
Reusing that ID and the identical payload returns the existing receipt; it never
restarts delivery. After `failed`, correct the cause and explicitly choose a new
ID. Failed and unfinished sends are not exposed as inbox mail.
`reply_to` defaults to the sender. A reply is a new message with `in_reply_to`
set to the original message UUID. The recipient calls `link_ack` after reading;
it must not reply automatically to receipts.
## Gemini 4 Carbon as a worker
Install and authenticate the Antigravity CLI (`agy` 1.5.9 tested). Check that
`agy models` includes `gemini-4-carbon-high`. Harness Link uses this exact
model for every turn.
1. In Codex or Claude Code, call `link_open_antigravity` with an optional
`name`, and optionally `workspace`, `mode` and `add_dirs` (below). The first
send starts the Antigravity conversation.
2. Call `link_send` with your actual session address, the returned
`antigravity:<UUID>` address, and your message. Each message is one Carbon
turn; messages to one conversation run one after another.
3. When the turn ends, Harness Link delivers Carbon's response, or the reason it
failed, to `reply_to` as a reply notification over that session's normal
transport (Claude live inbox, Codex queue). You do not need to poll. Pass
`notify: false` to turn this off for one message.
4. While it runs, `link_status` shows `progress`: the attempt, the number of
tool steps and Carbon's last actions (command, file). Either client can
continue the conversation using the same address; `link_read` lists the
conversation's messages with responses, progress and failures, plus the
messages Carbon sent.
Without `workspace`, Carbon runs in a private empty folder for text
discussion and has no tools. With an absolute `workspace` (an existing
directory), Carbon works there:
- `mode: "agent"` (default with a workspace): reads and edits files and runs
commands without prompts. agy runs with `--sandbox --mode auto`; commands
run in agy's terminal sandbox (writes to `/tmp` work), MCP tools stay off.
On accounts without agy's Auto mode, agy falls back to Default mode, where
file edits need approval; Harness Link therefore adds explicit allow rules
for the workspace and the conversation's own directories.
- `mode: "review"`: reads files only. Edits, commands, URLs and MCP are
denied by agy permission rules.
- `add_dirs`: extra absolute directories outside the workspace that Carbon may
use with its file tools, for example `/tmp/my-shots` for screenshots it must
look at (read and write in agent mode, read in review). Agent mode also gets a
private scratch directory. Not allowed: `/`, the home directory and Harness
Link's data directory.
Workspace, mode and `add_dirs` are fixed for the conversation; open a new one
to change them. From the CLI:
`claude-codex-link open-antigravity --workspace /absolute/repo --add-dir /tmp/shots`.
The workspace is the working directory and intended task scope, not a hard
operating-system sandbox: Carbon runs as the current user. Only pass trusted
workspaces.
### Long turns, failures and recovery
- `timeout_minutes` on `link_send` (CLI `--timeout-minutes`) limits one turn:
default 60 for agent, 30 for review, 10 without workspace, at most 240.
- `link_cancel` (CLI `cancel --message-id`) stops a running or waiting turn.
Commands Carbon started are stopped with it. No reply notification is sent.
- Headless agy cannot ask a human. A tool call that needs approval (for
example reading a file outside the granted directories) is denied and ends the
turn without text. Harness Link then asks Carbon, in the same conversation, to
continue without it (at most twice) and reports the call in `denied`.
- A failed turn has `status: error` with `error_code` (`permission_denied`,
`timeout`, `agy_exited`, `carbon_error`, `empty_response`, `runner_lost`, ...),
a readable `error`, a `next_step`, `journal` and `stderr_tail`. Timeouts,
cancels, denials, crashes and a lost runner never end the conversation: send
the next message to the same address. If the native conversation was saved,
Carbon keeps its context and is told the previous turn was interrupted; if
not, it is told the context is lost and to check the workspace first.
- Every turn keeps a journal in
`antigravity/<UUID>/turns/<message UUID>/` under the `.claude-codex-link`
data directory:
`stream.jsonl` (agy events), `stderr.txt`, `agy-<attempt>.log` (agy's own
log), `turn.json` (exit code, signal, timings) and `progress.json`. Files are
0600, size-capped, and well-known secret formats are masked before writing.
### Carbon writes to other agents
In agent mode Carbon can message a participant while it keeps working: it
writes a JSON file into the conversation's outbox directory (named in its
instructions) with `{"to": "<address>", "text": "..."}` or
`{"room_id": "<room UUID>", "text": "..."}`. Harness Link sends it at once
from the conversation's own address, writes `<name>.receipt.json` beside it,
and the answer arrives as a later Carbon turn. `to` must be an address that has
already written to this conversation; rooms are posted as `antigravity`. To ask
the sender something and wait, Carbon ends its turn with the question; that
response is delivered to `reply_to` like any other.
Reply notifications never produce another notification, so two agents cannot
ping-pong automatically. Peer text, including Carbon's, carries no new human
authority. Native Antigravity Desktop conversations are outside this transport.
## Receipts and limits
- `pending` / `sending`: detached delivery is in progress.
- `accepted`: Claude's inbox accepted the message, or Carbon returned a response. Claude reading remains unconfirmed.
- `error` / `cancelled`: a Carbon turn started and ended without a response; see `error_code` and `next_step`.
- `queued`: Codex persisted the message; execution and reading are unconfirmed.
- `acknowledged`: the addressed recipient reported reading it; work may remain.
- `failed`: a known pre-delivery rejection. This includes a Claude courier the API rejected before it
requested any tool (expired login, overload): `error_code` is `courier_auth` or `courier_api_error`
and `next_step` says what to do. A token-refresh race is retried twice within 90 seconds first.
- `unknown`: delivery may have occurred, but its receipt was lost. Inspect the
recipient before sending again; the bridge never automatically resends.
An acknowledgement is an agent-reported receipt, not cryptographic identity
proof. Peer text is marked as agent-authored and carries no new human authority.
Receiving-session permissions and inbound-message rules still apply.
The bridge never edits transcript or queue databases directly, and never starts
a second process on an open conversation to force a response. Archived Codex
targets and ambiguous Claude names are rejected. Claude history contains the
latest saved user/assistant text, excluding tool output and thinking. Codex
history supports cursors. Separate app-server runtime status is not Desktop
activity status.
## Consultations
`ask` returns a persistent job ID. Poll `ask_status` to completion. Workspace
jobs default to full agent mode, with shell, edits, tests, MCP servers, skills,
and plugins. Use `mode: review` only for an explicitly requested read-only
restriction. Without workspace, advisory mode has no tools.
The default is **claude-opus-5-5**, high effort. Check `resolved_model` in results.
`ask_followup` forks a completed consultation's context; it does not message an
already open session. Jobs have no elapsed-time or idle timeout and survive MCP
restarts. See [consultation details](plugins/claude-codex-link/README.md).
## Privacy and configuration
Messages and receipts are stored locally in the `.claude-codex-link` directory in the user home, with 0700
directories and 0600 files, until the owner removes them. They are not committed
or uploaded by the plugin. Normal provider processing still applies when an
agent receives a message. Consultations retain their original private job store
and 24-hour completed-job retention.
- `CLAUDE_CODEX_LINK_DIR`: message/receipt directory.
- `CLAUDE_CODEX_LINK_CLAUDE_CLI`: Claude executable override.
- `CLAUDE_CODEX_LINK_CODEX_CLI`: Codex executable override.
- `CLAUDE_CODEX_LINK_AGY_CLI`: Antigravity executable override.
- `CLAUDE_CODEX_LINK_TURN_MINUTE_MS`: test-only length of a turn-timeout minute.
- Existing `CLAUDE_ASK_*` settings continue to configure consultations.
The Claude courier runs Opus 5.5 with only SendMessage, safe mode, and no MCP
servers. Each live send invokes a short Claude turn. A local mailbox and explicit
receipt avoid confusing model output with confirmed transport delivery.
## Development
```sh
npm run check
npm test
```
Default tests use fake executables and consume no model usage. They cover
consultation recovery, exact model selection, message idempotency, receipt loss,
wrong-recipient acknowledgements, ambiguous names, archived targets, and Carbon
turns: progress, journal, agy failures, a killed runner, cancel, timeout,
denied tool calls, reply delivery without loops, the outbox, and an MCP server
restart mid-turn.
Canonical implementation notes: [.project-docs/project.md](.project-docs/project.md).
## License
MIT
This server cannot be deployed
Maintenance
ActivityMaintained
ResponsivenessOutdated data