Skip to content

Releases

v0.12.2

Choose a tag to compare

@francoCosimo95 francoCosimo95 released this 21 Oct 15:15
12e39ec

Added

  • Animated replay of two real exchanges at the top of the README (assets/demo.svg, rendered by
    scripts/demo.py): real times from the logs, waits shortened.

Changed

  • The agent's poll may also call the read-only RemoteTrigger get action (to read one routine);
    any action outside list, get, list_runs and get_run_log still fails the run.
  • Every claude CLI call blocks the tools it does not need (Bash, Edit, Write,
    NotebookEdit, WebFetch, WebSearch, Task) unless the call allows them explicitly.

Fixed

  • When a CLI run stops with an ALERT (poll or delivery), its transcript is kept for debugging in
    alert-transcripts/<time>.jsonl (tokens redacted, readable by the owner only, gitignored).

v0.12.1

Choose a tag to compare

@francoCosimo95 francoCosimo95 released this 19 Oct 14:07

Added

  • bridge.py remove-peer <peer> forgets an agent registered by mistake and updates the agents
    block in ~/.claude/CLAUDE.md (the block goes away with the last agent).

v0.12.0

Choose a tag to compare

@francoCosimo95 francoCosimo95 released this 19 Oct 13:54

Changed

  • Faster and cheaper polling: the agent's poll runs on the small model (haiku, override with
    CLAUDE_BRIDGE_POLL_MODEL), since it only lists runs and reads logs.
  • Setup never recreates an existing webhook routine (a new one breaks the owner's fast lane), and
    relaxes the poll to every 15 minutes when the fast lane works: the poll is then a safety net.
  • When the fast lane is off, the agent's final report and bridge.py check say how to turn it on.

v0.11.0

Choose a tag to compare

@francoCosimo95 francoCosimo95 released this 19 Oct 13:45

Changed

  • One protocol.py holds what both sides share (protocol name, types, statuses, A2A states, size
    and id rules, validation); scripts/sync.sh copies it next to each bridge.py.
  • Both sides validate envelopes before sending, and the agent's poll keeps only valid ones; ids
    must be plain (letters, digits, ., -, _), since they name triggers and files.

Fixed

  • Agent envelopes now always carry to (claude-code); answers from the agent had none.

Security

  • SECURITY.md states the trust boundary for relayed owner requests and why they are not signed.

v0.10.0

Choose a tag to compare

@francoCosimo95 francoCosimo95 released this 19 Oct 13:40

Added

  • add-peer --no-claude-md skips the agents block in ~/.claude/CLAUDE.md; when it writes it,
    it shows the line and how to undo it.
  • bridge.py peers prints a FOR_REPO_CLAUDE_MD line for a repository's CLAUDE.md, so cloud
    sessions know the agent's name too.

Changed

  • README: says which Grok Bot this is, what the bridge relies on (preview or undocumented), and
    what stays compatible before 1.0.

v0.9.2

Choose a tag to compare

@francoCosimo95 francoCosimo95 released this 19 Oct 13:16

Fixed

  • Agent side: the transcript audit crashed when an event's message was text, not an object
    (reported by CoS during the 0.9.1 setup).
  • Claude side: if the fast lane fails (for example its webhook routine was deleted), the message
    goes through the inbox instead of failing (FAST_LANE_FAILED ... then SEND_VIA_INBOX).

Added

  • add-peer --inbox-mode --no-fast-lane forgets a webhook that no longer exists; the setup tells
    the agent to use it when it recreated its webhook routine and cannot read the new values.

Verified

  • The channel works live: with the development flag, a message left by bridge.py drop
    appears in an open terminal session within seconds (Claude Team org, Claude Code 2.1.303).

v0.9.1

Choose a tag to compare

@francoCosimo95 francoCosimo95 released this 19 Oct 12:59

Fixed

  • The inbox lists triggers with include_completed and pages back 24 hours, so dedupe and
    cleanup also see ended triggers whatever order the listing returns.

v0.9.0

Choose a tag to compare

@francoCosimo95 francoCosimo95 released this 19 Oct 12:52

Added

  • A2A vocabulary. Answers carry state, the status as an A2A TaskState
    (TASK_STATE_COMPLETED, TASK_STATE_WORKING, TASK_STATE_INPUT_REQUIRED,
    TASK_STATE_REJECTED); thread is the A2A contextId. Mapping in PROTOCOL.md.
  • No double delivery. The inbox names push and ask triggers by the message's own id and
    skips a message it already delivered (BRIDGE_PUSH: duplicate ...), so sender retries are safe.
    Two answers to one question (accepted, then done) both arrive.
  • Agents by any name. add-peer keeps a marked block in ~/.claude/CLAUDE.md naming your
    agents, so "ask Ava" (whatever you named it) reaches the skill in every session. Examples are generic.
  • Channel (experimental). A Claude Code channel in the plugin: the agent drops its answer on
    your Mac (bridge.py drop, through local exec) and an open terminal session shows it at once.
    Needs the channels development flag during the research preview; not yet verified live.

Changed

  • README use cases are chat images (Grok black, Claude orange), generated from
    scripts/chats.py; a test keeps them in sync.

v0.8.0

Choose a tag to compare

@francoCosimo95 francoCosimo95 released this 19 Oct 12:40

Added

  • Agents can ask Claude Code. grok/bridge.py ask (connector tool ask_claude_code) sends a
    question to: "claude-code"; the inbox schedules a one-off routine (grok-bridge ask <id>)
    that investigates read-only and writes the answer as its final reply; the agent's poll reads it
    there (verified live: question in, answer read back). With --session, the
    question goes into a running session instead, which replies with the new Claude side command
    bridge.py answer.
  • Both sides know the bridge. The agent's "Claude Code" skill triggers on "ask Claude Code" /
    "chiedi a Claude Code"; setup asks the agent to remember the channel; the Claude skill knows how
    to answer an agent.

Changed

  • Answering sessions call no tool to reply: in scheduled sessions every connector call waits for
    a human approval, so the answer is their final reply and the agent's poll (now also allowed the
    read-only list action) collects it. The inbox keeps grok-bridge ask routines for 24 hours.

Fixed

  • Safety: the inbox routine's allowed_tools was empty, which means the full default tool
    set. It is now ["TodoWrite"] plus an explicit deny list (shell, writes, web, sub-agents).
    Reading stays allowed, so answering sessions, which inherit the list, can look at code.

v0.7.0

Choose a tag to compare

@francoCosimo95 francoCosimo95 released this 19 Oct 12:18

Added

  • Native Claude Code plugin grok-bridge: the repo is a Claude Code marketplace
    (/plugin install grok-bridge --marketplace francoCosimo95/claude-grok-bridge). Cloud sessions
    get it through the repository's .claude/settings.json, no files to copy. Validated with
    claude plugin validate --strict.
  • Threads. Optional thread field (id of the first message): ask/tell/note --thread <id>;
    answers keep it; the agent archives messages both ways and reads bridge.py thread <id> before
    answering a follow-up.
  • Any agent platform. The inbox pushes answers from any from.kind other than claude-code;
    CLAUDE_BRIDGE_AGENT_KIND names the agent's platform (default grok-bot).

Changed

  • The skill finds bridge.py next to its own SKILL.md, so it works as a plugin, a linked skill or
    a repo skill.