lisa-openclaw-connect-repo-topic

name	lisa-openclaw-connect-repo-topic
description	Bind a Telegram forum topic to…

lisa-openclaw-connect-repo-topic

Turn a Telegram forum topic into a repeatable OpenClaw entrypoint for code work. A dispatcher agent receives the request and delegates; a worker agent runs the local coding CLI in a safe target directory and relays the result.

This SKILL.md is complete and runnable under both Claude Code and Codex (no command/$ARGUMENTS layer in Codex). Keep real ids/handles/paths out of committed files — they go only in ~/.openclaw/openclaw.json. For exact config shapes read references/repo-topic-config.md.

Prerequisites

• lisa-openclaw-setup has run and its capability probes passed.
• A Telegram supergroup with forum topics, and a bot that is (or can be promoted to) a group admin — bot presence alone is not enough for topic routing.

Boundaries (apply every time)

• groups = the human trust boundary
• topics = the project-routing boundary
• native-reply roots = the short-term session boundary inside a topic
• agents = the capability boundary
• Do not mix admin-only repos with broader-collaboration repos in one group. If two repos need different member lists, use separate groups, not topic-only separation.

Scope modes (pick exactly one per topic)

• single-repo — the topic is pinned to one repo directory; the dispatcher never asks which repo.
• folder-scoped — the topic is pinned to a parent folder containing multiple repos; the dispatcher must confirm the inferred repo(s) unless the user named them explicitly.

Do not mix both behaviors in one topic.

Required inputs

• trust boundary: admin or developer
• scope mode: single-repo or folder-scoped
• access surface: Telegram group + topic (by name; ids captured during the run)
• the bot handle to bind
• workspace root on disk (repo dir for single-repo; parent folder for folder-scoped)
• allowed Telegram user ids
• completion policy:
  • admin topics may request change + docs + commit + PR + merge
  • developer topics default to change + docs + commit + PR only
• for folder-scoped: a repo catalog (label, absolute path, a few matching keyword hints)

Workflow

1. Resolve the contract

Confirm the trust boundary, scope mode, workspace root, bot handle, and (for folder-scoped) the repo catalog. If a needed agent doesn't exist yet, create it in step 3.

2. Provision / reuse the Telegram surface

Reuse the trust-matched group and the exact topic if they already exist; otherwise create them (enable forum/topics mode first). Ensure the bot is in the group and promoted to admin. Telegram Bot API cannot list groups by name — discover ids from a logged-in Telegram client, an existing route in ~/.openclaw/openclaw.json, or a received Bot API update. If browser/UI automation blocks on a human confirmation step, give the user only the blocked step, wait, then resume.

3. Implement the dispatcher + worker pair

Create or update two agents (exact tool-deny shapes in references/repo-topic-config.md):

• <topic-slug>-dispatch — skill-aware entrypoint; tools restricted to delegation/session tools (+ optional read); denies file writes, shell/exec, browser, gateway, automation; may delegate only to <topic-slug>-codex.
• <topic-slug>-codex — restricted worker; keeps read + exec; denies session spawning, browser, gateway, and direct file-write tools; runs the local coding CLI for the actual work.

4. Repo-selection behavior

• single-repo: treat the repo as decided; pass the fixed repo path to the worker; never ask.
• folder-scoped: infer candidates from the catalog. If the user named repo(s) explicitly, skip confirmation; otherwise ask before spawning:
  • one likely repo: This looks like <repo>. Is that correct?
  • multiple: This may belong to <repo-1> or <repo-2>. Is that correct, or which repo(s)? On confirmation of multiple repos, spawn one worker run per repo (or ask the user to split a too-coupled change). Never skip confirmation in folder-scoped mode unless selection was explicit.

5. Worker safety contract

Every worker task receives an explicit repo path and:

1. inspects git status --short --branch in the selected repo
2. if the primary checkout is dirty or already on a feature branch, creates a temporary worktree from main under /tmp
3. trusts local tool config in a fresh worktree if required (e.g. mise trust <target_dir>)
4. runs the coding CLI in the foreground in the target directory
5. cleans up the temporary worktree after success when safe

The worker prepares the safe target and launches the CLI; it does not edit files directly.

6. Bind the topic

Route the topic to the dispatcher. In single-account Telegram configs, set channels.telegram.groups.<group-id>.topics.<topic-id>.agentId = <topic-slug>-dispatch. In multi-account Telegram configs, set the same route under the owning bot account: channels.telegram.accounts.<account-id>.groups.<group-id>.topics.<topic-id>.agentId = <topic-slug>-dispatch. Account configs do not inherit root channels.telegram.groups routes, so a root-only route can validate while the actual bot still ignores the topic-level requireMention, agentId, and prompt. When in doubt, mirror the route under the owning account and keep any existing root group route only as a fallback/documentation shape.

Account-scoped topic agentId routes also require an OpenClaw runtime that treats the topic agent as an explicit group route for non-default Telegram accounts. If the self-test logs drop non-default account requires explicit binding, update or rebuild OpenClaw before adding config workarounds; do not create persistent ACP bindings solely to bypass that guard.

Keep allowlist policy, and add allowFrom only when membership must be narrower than the group. Leave the topic-level requireMention = false (the default) so the agent activates on any message — the topic is bound 1:1 to this dispatcher, so an @mention carries no routing information and is pure friction. Set it to true only for a shared-workspace topic where humans also coordinate with each other and you don't want every line to spawn a run; the group-level requireMention stays true regardless. See "Mention gating" in references/repo-topic-config.md for the tradeoff. The topic systemPrompt must state the scope mode, treat each native-reply root as an independent request context, confirm repo selection only in folder-scoped mode, spawn the worker with an explicit repo path, and return the worker result to the topic. If the route may return generated files, the prompt must also say to write or copy returnable artifacts under ~/.openclaw/media before calling message(action="send"), because outbound local media delivery rejects arbitrary /tmp paths. Back up ~/.openclaw/openclaw.json before editing and preserve unrelated routes.

7. Validate + self-test

openclaw config validate
openclaw gateway restart
openclaw gateway status
openclaw channels status --probe

Then from the target topic, send a plain message with no @mention (the default requireMention = false means the agent must activate without one) asking for an exact-token reply with no file changes, commits, PRs, or merges, e.g. reply with exactly TELEGRAM-ROUTE-OK. Confirm the visible reply, that the dispatcher spawned the worker, and that the worker ran in the intended repo. If there is no dispatcher session, inspect /tmp/openclaw/openclaw-YYYY-MM-DD.log; drop non-default account requires explicit binding means the OpenClaw runtime is too old for account-scoped topic-agent routing. If the topic was deliberately left at requireMention = true, mention the bot instead (<bot-handle> reply with exactly TELEGRAM-ROUTE-OK) and additionally confirm that an un-mentioned message is ignored. For folder-scoped topics, also send a request that implies but doesn't name a repo and confirm the dispatcher asks for confirmation before proceeding. Do not treat openclaw agent --agent <id> ... as proof a topic route works — use the visible topic reply.

Output standard

Finish with: group id + topic id used; scope mode; workspace root and repo path or catalog; dispatcher + worker agent ids; whether the bot is confirmed group admin; validation results; whether the self-test passed; and any remaining manual follow-up or trust caveat.

Related

lisa-openclaw-setup (run first), lisa-openclaw-connect-staff (facilitator/specialist staff on Telegram or Slack).