| name | telegram-support |
| description | Use the local Telegram support corpus to answer support analytics questions, prepare evidence-backed reply drafts, and post only after explicit operator confirmation. |
Telegram Support
Use this skill when the operator asks about a configured Telegram support chat, wants a draft answer for a user or message, or wants to post a previously approved reply.
CLI
Run commands through the bundled local helper. Resolve the helper relative to this skill's plugin root; from this file that is ../../scripts/tg-support. If the operator is developing from a source checkout, scripts/tg-support from the repo root is equivalent.
<plugin-root>/scripts/tg-support --profile default <command>
The first run bootstraps the Python runtime under the operator's local data directory unless TG_SUPPORT_VENV overrides it.
The CLI prints JSON. Treat that JSON as the source of truth for local corpus state, evidence, conflict checks, draft IDs, and confirmation tokens.
Setup Preflight
Before answering analytics questions, searching, drafting, or posting, run:
<plugin-root>/scripts/tg-support --profile default status
If ok is false, do not continue into the normal workflow. Use next_action to guide the operator through the missing setup step:
setup: ask for a Telegram chat identifier, whether to add optional website or blog seed URLs, whether to add optional GitHub repository evidence configuration, and whether to add optional operator identities for authoritative support replies. Tell the operator they can skip seed setup, skip repository setup, and skip operator identity setup. If repository evidence is provided, pass the repository as --repository <owner/repo-or-url-or-path> and the branch as --repository-branch <branch> only when the operator provides one; the CLI defaults the branch to main when omitted. If operator identities are provided, pass each as a repeated --operator <identity>. Then run setup.credentials: explain that Telegram API ID and API hash are needed from https://my.telegram.org, then run credentials --api-id <id> --api-hash-stdin and provide the hash on stdin after the operator provides it.login: run login so the operator can complete the local Telethon user-session login.sync: run sync to pull Telegram history.crawl: run crawl to fetch configured website or blog resources.index: run index to build the local retrieval index.ready: proceed with the requested support workflow.
After each setup command, run status again and continue from the new next_action. Never print or repeat the API hash back to the operator.
Repository Evidence is optional. If status includes configured repository details, use them only for product-behavior, capability, API-behavior, or debugging questions. If no repository is configured, normal Telegram, web, and Manual Knowledge workflows remain ready.
Support Exchanges
index derives Support Exchange records from Telegram history. A Support Exchange is a role-labeled question/answer unit with requester messages, configured operator replies, peer/community replies, and unanswered state separated from the surrounding chat window.
When search or draft-context returns Support Exchange evidence, show the roles in the evidence summary. Operator replies can support a direct answer. Peer/community replies are useful context but are not authoritative support truth by themselves. Unanswered exchanges are unresolved user needs. If evidence_sufficiency includes peer_exchange_only, ambiguous_exchange_authority, or unanswered_exchange, show the reason and use the same cautious fallback posture as other insufficient evidence.
Manual Knowledge
When the operator asks to save durable support knowledge, extract the note text, effective date, optional expiry date, and caveats from the conversation. Show those parsed fields back to the operator and ask for confirmation before saving.
After confirmation, run:
<plugin-root>/scripts/tg-support --profile default knowledge-add --text <text> --effective-date <YYYY-MM-DD>
Add --expires-date <YYYY-MM-DD> and --caveats <text> when present. If the operator revises or cancels the parsed fields, do not run knowledge-add until they confirm the corrected note.
When search or draft-context returns Manual Knowledge Note evidence, include it in the evidence summary. If the JSON includes non-empty conflicts, show the manual note, older evidence, fresher evidence when present, and ask the operator which source should guide the response before treating the note as settled truth.
When draft-context returns evidence_sufficiency, show the sufficiency state and reasons in the evidence summary. If direct_answer_supported is false, present two draft options before creating a draft: a cautious evidence-limited answer and a DM follow-up asking only for the missing support-blocking information. Do not save the DM follow-up wording as Manual Knowledge and do not treat it as evidence.
When draft-context returns fuzzy_author_candidates, treat them as disambiguation options only. Do not use fuzzy candidates as target history; ask the operator to choose or rerun with an exact username/display name before drafting as that user.
When draft-context returns target.language, draft the reply in that language unless the operator asks otherwise. Treat translated_text fields as helper context for understanding evidence, not as support truth and not as a reason to answer in English. If the target language is uncertain, show the language assumption before creating or posting the draft.
Repository Evidence
For questions about what the product supports, how an API behaves, why a user can or cannot do something, or likely debugging causes, run:
<plugin-root>/scripts/tg-support --profile default repo-evidence "<query>"
Use Repository Evidence from the configured branch as the highest-priority support truth. It outranks Manual Knowledge Notes, Telegram evidence, and web evidence when sources disagree.
If repository_evidence.warning is present, show it before relying on the code evidence. If the warning says stale evidence was used, say that the checkout could not be refreshed and the cited code may be outdated.
For reply drafting, a stale Repository Evidence warning should trigger the same two-option fallback posture as insufficient draft-context evidence: offer a cautious evidence-limited answer and a DM follow-up option before creating the selected draft.
Do not run repository lookup for normal analytics or drafting questions that do not depend on product behavior. Do not create Manual Knowledge Notes, modify code, create branches, or open pull requests from repository findings unless the operator starts a separate explicit workflow.
Safety Boundary
Never post to Telegram directly from agent reasoning. A Telegram write is allowed only after:
draft-context, repo-evidence, or another retrieval command has produced needed evidence and any returned conflicts or stale repository warnings are shown to the operator.- The operator has seen the exact target chat, target user or message ID when available, exact message text, and evidence summary.
- The operator explicitly replies with
post for that exact message and target.
- The CLI confirmation command succeeds.
If the operator says cancel, use the cancel confirmation token or leave the draft unposted. Do not silently retry duplicate posts.
Workflows
- For reply drafting, follow
references/reply-workflow.md.
- For analytics questions, follow
references/analytics-workflow.md.
