telegram-agent-loop

Stars14

Forks5

UpdatedJuly 22, 2026 at 18:00

Drives Gantry tests through the visible Telegram UI with Chrome. Use only when testing Telegram rendering, Telegram buttons, topic routing, or visible Telegram conversation behavior; use agent-chat-test for direct runtime agent-chat tests.

Installation

Install with Codex or Claude Copy this prompt, paste it into Codex, Claude, or another assistant, and let it review the skill page and install it for you.

Source

knacklabs

knacklabs/gantry

View GitHub Repository View Creator Repositories

Download

Related occupationsSOC

Based on SOC occupation classification

Software Quality Assurance Analysts and TestersComputer and Mathematical Occupations·SOC 15-1253

SKILL.md

readonly

name	telegram-agent-loop
description	Drives Gantry tests through the visible Telegram UI with Chrome. Use only when testing Telegram rendering, Telegram buttons, topic routing, or visible Telegram conversation behavior; use agent-chat-test for direct runtime agent-chat tests.

Telegram Agent Loop

Use this skill for operator-style verification loops where the Telegram UI itself is part of the behavior under test. For direct runtime agent-chat tests, use agent-chat-test instead.

Ground Rules

Use the chrome:Chrome skill and the existing Chrome/Telegram tab for Telegram only.
Do not use Chrome, CDP, or the Telegram browser profile as the Gantry job Browser. If the tested job needs Browser, instruct the agent to use Gantry Browser tools and the job's Gantry profile.
Interact with Telegram like a user: read the visible thread, focus the message box, paste or type the prompt, send it, and watch the visible status.
Keep messages concise and action-oriented. Include exact job ids, run ids, expected tools, and acceptance criteria.
Treat the Telegram agent's report as a claim. Cross-check it with local Gantry CLI/runtime evidence before calling the issue fixed.

Loop

Open and orient
- Claim the existing Telegram tab through Chrome.
- Verify the visible topic is the intended Kai group/topic before sending.
- Read the latest visible messages so the follow-up matches current state.
Send the instruction
- Use the visible Telegram composer, not hidden APIs.
- State what the agent must do, how to avoid browser/profile mixing, and what proof to report.
- Include pass/fail criteria: terminal status, run id, event ids, required tools, tool activity, browser activity, and exact blocker text if failed.
Monitor
- Wait for visible status changes such as queued, working, running, completed, or failed.
- Poll local evidence in parallel with CLI commands such as:
  - node dist/cli/index.js jobs show <job_id>
  - node dist/cli/index.js jobs events <job_id> --run <run_id> --limit 30
- Prefer local runtime events for factual state; use Telegram for agent reasoning and certification.
Intervene only on concrete gaps
- If the job is ready but paused, tell the agent to resume and run now.
- If tool activity is missing, ask the agent to exercise the required tool and report the event id.
- If Browser setup is required, tell the agent to use the Gantry Browser profile shown by readiness, never the Chrome/Telegram tab.
- If credentials are needed, tell the agent to use the approved credential path without printing secrets.
Certify
- Accept certification only when the agent report and local evidence agree.
- Required proof for scheduled jobs: job visible, run produced, terminal status known, required tool preflight persisted, Browser either used or explicitly blocked, and no generic timeout without diagnostics.
- If the agent reports success but local events disagree, send a corrective follow-up with the mismatch.

Follow-Up Template

Current local evidence:
- Job status: <status>
- Run id: <run_id or none>
- Latest event ids: <ids>
- Missing proof: <gap>

Continue from here. Do not restart from scratch unless the current run is terminal. Use the visible Gantry/Kai workflow, report exact event ids, and stop only when the job is terminal and certified.

More from this repository

same repository

forge

knacklabs/gantry

Operate the Symphony Forge harness: start tasks, save approved plans, record decisions, assumptions, and client sign-off, harvest the docs/context inbox, check gate status, and review proposed skills. Invoke when the user says "start a task", "save this plan", "record a decision", "harvest context", "is this PR ready", "harness status", "what now", "create a new app", "set up a new project", "show me progress", "show the board", "review the plan", or "/forge".

2026-07-3114

forge

knacklabs/gantry

Operate the Symphony Forge harness: start tasks, save approved plans, record decisions, assumptions, and client sign-off, harvest the docs/context inbox, check gate status, and review proposed skills. Invoke when the user says "start a task", "save this plan", "record a decision", "harvest context", "is this PR ready", "harness status", "what now", "show me progress", "show the board", "review the plan", "create a new app", "set up a new project", or "/forge".

2026-07-3114

gantry-admin

knacklabs/gantry

Gantry self-administration reference for current host runtime operations: CLI commands, settings.yaml runtime settings, .env credential keys, agent management, scheduler tools, memory tools, browser tools, diagnostics, and service control. Use when asked to manage Gantry itself.

2026-07-3114

agent-chat-test

knacklabs/gantry

Drives Gantry agent-chat tests through the local Control API session path and captures assistant responses. Use when testing an agent response, permission prompt, skill behavior, or runtime chat flow without needing Telegram UI rendering.

2026-07-2214

architecture-refactor

knacklabs/gantry

Use for folder movement, layer boundaries, and domain/application/runtime/adapters refactors in Gantry.

2026-07-2214

permission-safety

knacklabs/gantry

Use for tools, permissions, risky actions, sandbox, browser, and execution safety changes in Gantry.

2026-07-2214

name	telegram-agent-loop
description	Drives Gantry tests through the visible Telegram UI with Chrome. Use only when testing Telegram rendering, Telegram buttons, topic routing, or visible Telegram conversation behavior; use agent-chat-test for direct runtime agent-chat tests.

Telegram Agent Loop

Use this skill for operator-style verification loops where the Telegram UI itself is part of the behavior under test. For direct runtime agent-chat tests, use agent-chat-test instead.

Ground Rules

Use the chrome:Chrome skill and the existing Chrome/Telegram tab for Telegram only.
Do not use Chrome, CDP, or the Telegram browser profile as the Gantry job Browser. If the tested job needs Browser, instruct the agent to use Gantry Browser tools and the job's Gantry profile.
Interact with Telegram like a user: read the visible thread, focus the message box, paste or type the prompt, send it, and watch the visible status.
Keep messages concise and action-oriented. Include exact job ids, run ids, expected tools, and acceptance criteria.
Treat the Telegram agent's report as a claim. Cross-check it with local Gantry CLI/runtime evidence before calling the issue fixed.

Loop

Open and orient
- Claim the existing Telegram tab through Chrome.
- Verify the visible topic is the intended Kai group/topic before sending.
- Read the latest visible messages so the follow-up matches current state.
Send the instruction
- Use the visible Telegram composer, not hidden APIs.
- State what the agent must do, how to avoid browser/profile mixing, and what proof to report.
- Include pass/fail criteria: terminal status, run id, event ids, required tools, tool activity, browser activity, and exact blocker text if failed.
Monitor
- Wait for visible status changes such as queued, working, running, completed, or failed.
- Poll local evidence in parallel with CLI commands such as:
  - node dist/cli/index.js jobs show <job_id>
  - node dist/cli/index.js jobs events <job_id> --run <run_id> --limit 30
- Prefer local runtime events for factual state; use Telegram for agent reasoning and certification.
Intervene only on concrete gaps
- If the job is ready but paused, tell the agent to resume and run now.
- If tool activity is missing, ask the agent to exercise the required tool and report the event id.
- If Browser setup is required, tell the agent to use the Gantry Browser profile shown by readiness, never the Chrome/Telegram tab.
- If credentials are needed, tell the agent to use the approved credential path without printing secrets.
Certify
- Accept certification only when the agent report and local evidence agree.
- Required proof for scheduled jobs: job visible, run produced, terminal status known, required tool preflight persisted, Browser either used or explicitly blocked, and no generic timeout without diagnostics.
- If the agent reports success but local events disagree, send a corrective follow-up with the mismatch.

Follow-Up Template

Current local evidence:
- Job status: <status>
- Run id: <run_id or none>
- Latest event ids: <ids>
- Missing proof: <gap>

Continue from here. Do not restart from scratch unless the current run is terminal. Use the visible Gantry/Kai workflow, report exact event ids, and stop only when the job is terminal and certified.