Skip to content

fix(tg-gateway): dedupe skill bundle, unpin engine model, add claude … - #8

Merged
mnemonik-dev merged 1 commit into
developfrom
claude/tg-claude-slash-commands-irtd4k
Jul 27, 2026
Merged

fix(tg-gateway): dedupe skill bundle, unpin engine model, add claude …#8
mnemonik-dev merged 1 commit into
developfrom
claude/tg-claude-slash-commands-irtd4k

Conversation

@mnemonik-dev

Copy link
Copy Markdown
Owner

…reset runbook

Operator report 2026-07-27: molyanov commands unreliable from Telegram (alias mismatch, agent doesn't recognize them at once), no way to change engine or reset/relogin Claude after quota exhaustion.

Fixes in this repo's scope:

  • claude-skills bundle: remove the 17 underscore-duplicate skill dirs (code_reviewing etc.) — they carried IDENTICAL frontmatter name: values as their dash twins, so every deployed session saw duplicate-skill conflicts and resolution became unreliable. Dash dirs are canonical; command files keep both spellings (Telegram only allows [a-z0-9_] in bot commands). Convention documented in files/claude-skills/README.md.
  • telegram-ai-agent role: purge already-deployed underscore skill dirs from both engine HOMEs (skills sync is additive), driven by telegram_ai_agent_legacy_underscore_skills in defaults.
  • molyanov role: stop symlinking the raw upstream bundle as nested ~/.claude/skills/molyanov-ai-dev — Claude Code resolves personal skills at ~/.claude/skills//SKILL.md, so the nested tree was either undiscovered or a second duplicate source; remove the legacy link on deploy. Clone at /opt/molyanov-ai-dev stays for reference/diffing.
  • config.yml.j2: engine model no longer hardcoded (claude-opus-4-1) — telegram_ai_agent_claude_model / _codex_model vars + per-topic model override, so a quota-driven model switch is a one-var redeploy. Also fix latent trim_blocks newline bug that glued 'topics:' and 'model:' onto preceding lines whenever the topic matrix was non-empty.
  • new bot-claude-reset helper (/usr/local/bin): stop bot, kill stray claude/MCP processes, optionally wipe session-resume state and rotate CLAUDE_CODE_OAUTH_TOKEN / ANTHROPIC_API_KEY in .env, restart. This IS the relogin path — bot auth comes from .env, not claude /login. README runbook added.
  • work/tg-slash-command-gateway/user-spec.md: spec for the bot-fork side (mnemonik-dev/telegram-ai-agent): command discovery + setMyCommands registration, dash/underscore alias normalization, native command mapping (/clear /model /status), /engine switching incl. General chat, in-chat /relogin. Fork is not reachable from this session.

Claude-Session: https://claude.ai/code/session_01HEd1EGpGfDERw1D4QJdSvM

…reset runbook

Operator report 2026-07-27: molyanov commands unreliable from Telegram
(alias mismatch, agent doesn't recognize them at once), no way to change
engine or reset/relogin Claude after quota exhaustion.

Fixes in this repo's scope:

- claude-skills bundle: remove the 17 underscore-duplicate skill dirs
  (code_reviewing etc.) — they carried IDENTICAL frontmatter name: values
  as their dash twins, so every deployed session saw duplicate-skill
  conflicts and resolution became unreliable. Dash dirs are canonical;
  command files keep both spellings (Telegram only allows [a-z0-9_] in
  bot commands). Convention documented in files/claude-skills/README.md.
- telegram-ai-agent role: purge already-deployed underscore skill dirs
  from both engine HOMEs (skills sync is additive), driven by
  telegram_ai_agent_legacy_underscore_skills in defaults.
- molyanov role: stop symlinking the raw upstream bundle as nested
  ~/.claude/skills/molyanov-ai-dev — Claude Code resolves personal skills
  at ~/.claude/skills/<name>/SKILL.md, so the nested tree was either
  undiscovered or a second duplicate source; remove the legacy link on
  deploy. Clone at /opt/molyanov-ai-dev stays for reference/diffing.
- config.yml.j2: engine model no longer hardcoded (claude-opus-4-1) —
  telegram_ai_agent_claude_model / _codex_model vars + per-topic model
  override, so a quota-driven model switch is a one-var redeploy. Also
  fix latent trim_blocks newline bug that glued 'topics:' and 'model:'
  onto preceding lines whenever the topic matrix was non-empty.
- new bot-claude-reset helper (/usr/local/bin): stop bot, kill stray
  claude/MCP processes, optionally wipe session-resume state and rotate
  CLAUDE_CODE_OAUTH_TOKEN / ANTHROPIC_API_KEY in .env, restart. This IS
  the relogin path — bot auth comes from .env, not claude /login. README
  runbook added.
- work/tg-slash-command-gateway/user-spec.md: spec for the bot-fork side
  (mnemonik-dev/telegram-ai-agent): command discovery + setMyCommands
  registration, dash/underscore alias normalization, native command
  mapping (/clear /model /status), /engine switching incl. General chat,
  in-chat /relogin. Fork is not reachable from this session.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HEd1EGpGfDERw1D4QJdSvM
@mnemonik-dev
mnemonik-dev merged commit 251995d into develop Jul 27, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants