发布

  • Feat/gateway slack data dependent turns (#3350)

    frostbyte_neo 发布于 2026-06-30 20:37:24 +00:00

    • feat(gateway): send Slack messages with prior tool output in compound turns

    Expose the slack_send_message action tool to gateway (Telegram) turns and make
    data-dependent compound requests ("check the weather and then send it to slack")
    actually carry the looked-up value into the second tool.

    • gateway_action_tools: expose available registry tools (e.g. Slack) and strip
      interactive approval for headless turns; return shell_run output to the model
      so a later tool can consume it instead of only seeing {"ok": true}.
    • action prompt: split the COMPOUND TURN RULE into independent (emit both at
      once) vs data-dependent (run A, observe its result, then send B with the real
      value) and align the slack_send_message guidance.
    • raise the action loop budget (3 -> 6) and mirror it in the live planning test
      so multi-step data-dependent chains can complete.
    • slack sent_result no longer carries an "error" key, which the shared tool
      runtime would otherwise treat as a failed call.
    • add an end-to-end gateway test proving the shell output flows into the Slack
      message, plus Slack tool unit tests.

    Co-authored-by: Cursor cursoragent@cursor.com

    • feat(gateway): send Slack messages with prior tool output in compound turns

    Expose the slack_send_message action tool to gateway (Telegram) turns and make
    data-dependent compound requests ("check the weather and then send it to slack")
    actually carry the looked-up value into the second tool.

    • gateway_action_tools: expose available registry tools (e.g. Slack) and strip
      interactive approval for headless turns; return shell_run output to the model
      so a later tool can consume it instead of only seeing {"ok": true}.
    • action prompt: split the COMPOUND TURN RULE into independent (emit both at
      once) vs data-dependent (run A, observe its result, then send B with the real
      value) and align the slack_send_message guidance.
    • raise the action loop budget (3 -> 6) and mirror it in the live planning test
      so multi-step data-dependent chains can complete.
    • slack sent_result no longer carries an "error" key, which the shared tool
      runtime would otherwise treat as a failed call.
    • add an end-to-end gateway test proving the shell output flows into the Slack
      message, plus Slack tool unit tests.

    Co-authored-by: Cursor cursoragent@cursor.com

    • modularized integrations resolution

    • new pass down flow

    • gateway in progress, new changes with agent up top

    • gateway in progress, new changes with agent up top

    • restructuring and grouping of the core/agent_harness repository

    • Fix gateway follow-up action context

    • Fix PR CI test isolation failures


    Co-authored-by: Cursor cursoragent@cursor.com

    下载附件