Telegram Forum Threads — Enhancement Plan¶

Date: 2026-03-19 Status: Draft — awaiting Jeff's review Priority: High (daily workflow impact)

Context¶

We migrated from a single Telegram DM to a forum group ("Jeff & Jules") with topic threads. Topics are created but all running on the same Opus model with full workspace injection. Result: ~2 min response latency per message and unnecessary token burn.

This plan makes each topic a purpose-built channel with the right model, context, and routing.

Topic Map¶

Topic	Thread ID	Model	Purpose
General	1	Opus	Main comms, strategy, daily standup
APA	11	Opus	Ascent Performance Analytics — research, decisions
BioThread	12	Opus	CNS fatigue monitoring — research, spec work
MC Dev	13	Sonnet	Mission Control development — coding, bugs, PRs
Ideas	14	Sonnet	Brainstorming, new concepts, quick eval
Ops	15	Sonnet	Infrastructure, health, deployments, alerts

Enhancements (Ordered by Impact)¶

1. Per-Topic Model Routing (Highest Impact)¶

What: Assign Sonnet to MC Dev, Ideas, Ops. Keep Opus for General, APA, BioThread. How: channels.telegram.groups.<chatId>.topics.<threadId> config with per-topic model override via agentId or session-level model setting. Impact: 3-5x faster responses in dev/ops threads. Estimated 40-60% token cost reduction on high-volume threads. Config approach: Since per-topic model override requires separate agents, we either: - (a) Use session_status model overrides per session (quick, not durable across restarts) - (b) Create lightweight agent entries with different default models and route topics via agentId - (c) Recommended: Use per-topic config + session model override on first message as a pragmatic middle ground until OpenClaw adds native per-topic model config.

2. Per-Topic System Prompts (High Impact)¶

What: Each topic gets a focused system prompt instead of the full SOUL.md + MEMORY.md + everything. How: channels.telegram.groups.<chatId>.topics.<threadId>.systemPrompt — appended to the base agent system prompt per the docs (topic inherits group settings unless overridden). Impact: Reduces prompt size by 30-50%, faster time-to-first-token, more focused responses.

Topic	System Prompt Focus
General	Full context (default — no override needed)
APA	APA initiative context, sports analytics domain, research methodology
BioThread	Biometrics, CNS fatigue, GAP Score, wearable data
MC Dev	Coding conventions, MC architecture, component patterns, port 3100
Ideas	Lightweight — evaluation frameworks, quick analysis mode
Ops	Infrastructure, services, health checks, deployment procedures

3. Per-Topic Skills Filtering (Medium Impact)¶

What: Each topic only exposes relevant skills. How: channels.telegram.groups.<chatId>.topics.<threadId>.skills array. Impact: Prevents irrelevant tool suggestions, slightly reduces system prompt.

Topic	Skills
General	All (default)
APA	web_search, summarize, gog
BioThread	web_search, summarize
MC Dev	coding-agent, github, qa-validation, service-management, vv-dashboard-design
Ideas	web_search, summarize, project-pipeline
Ops	healthcheck, service-management, github, camsnap

4. Cron Output Routing (Medium Impact)¶

What: Route cron job output to the appropriate topic instead of DM/General. How: Update existing cron jobs' delivery targets to use telegram:-1003586302776:topic:<threadId>. Impact: Information lands where it belongs. Reduces noise in General.

Cron Job	Current Target	New Target
Daily standup (5 AM)	DM	General (topic 1)
Health check	DM	Ops (topic 15)
Nightly consolidation	Silent	Silent (no change)
Future: SigInt scan	N/A	Ideas (topic 14) or future SigInt topic
Future: CI/CD alerts	N/A	Ops (topic 15)

5. Pinned Scope Docs (Low Effort, High Clarity)¶

What: Pin a brief scope message in each topic so purpose is immediately visible. How: Send a formatted scope message to each topic via Bot API, then pin it. Impact: UX clarity — anyone opening a topic knows what it's for.

6. Cross-Topic Handoffs (Future)¶

What: "Move this to APA" command that carries conversation context to another topic. How: Custom command that extracts recent context, sends a summary to the target topic, and links back. Impact: Smooth transitions without re-explaining. Lower priority — build after core enhancements are stable.

7. Topic-Aware Memory Tagging (Future)¶

What: Tag daily log entries with source topic for smarter consolidation. How: Memory manager captures topic context in structured entries. Impact: Better memory organization, more relevant recall per topic.

Implementation Spec¶

Phase 1: Config & Model Routing (Do First)¶

Update openclaw.json channels config:

{
  channels: {
    telegram: {
      // ... existing config ...
      groups: {
        "-1003586302776": {
          requireMention: false,
          groupPolicy: "allowlist",
          allowFrom: ["8653591173"],
          topics: {
            "1": {
              // General — inherits Opus default, full context
            },
            "11": {
              // APA — Opus, focused prompt
              systemPrompt: "This is the APA (Ascent Performance Analytics) thread. Focus on sports performance analytics, mid-market SaaS strategy, competitive analysis, and research methodology. Reference APA initiative docs when relevant."
            },
            "12": {
              // BioThread — Opus, focused prompt
              systemPrompt: "This is the BioThread thread. Focus on CNS fatigue monitoring, GAP Score methodology, wearable biometric data (Garmin Fenix 8, Apple Watch), and sports science research. Reference BioThread specs when relevant."
            },
            "13": {
              // MC Dev — Sonnet for speed
              systemPrompt: "This is the MC Dev thread. Focus on Mission Control dashboard development. MC runs on port 3100. Reference vv-dashboard-design skill for component patterns and design tokens. Keep responses concise and code-focused.",
              skills: ["coding-agent", "github", "qa-validation", "service-management", "vv-dashboard-design"]
            },
            "14": {
              // Ideas — Sonnet for speed
              systemPrompt: "This is the Ideas thread. Evaluate concepts quickly using the project-pipeline framework. Be creative but grounded. Quick analysis mode — don't over-research unless asked.",
              skills: ["web_search", "summarize", "project-pipeline"]
            },
            "15": {
              // Ops — Sonnet for speed
              systemPrompt: "This is the Ops thread. Focus on infrastructure, service health, deployments, and operational issues. MC health: curl -sf http://localhost:3100/api/health. Be direct and action-oriented.",
              skills: ["healthcheck", "service-management", "github"]
            }
          }
        },
        "*": {
          requireMention: false
        }
      }
    }
  }
}

Model routing note: The per-topic config supports systemPrompt, skills, requireMention, allowFrom, enabled, and groupPolicy natively. For model routing, two paths: - agentId per topic — create lightweight agent entries (agents.list[]) with different default models, then set topics.<id>.agentId. Most durable, but each agent gets its own workspace/session. - Session override on connect — use session_status(model=...) to set model per topic session. Less config, but resets if session is pruned.

Recommendation: Start with session-level model overrides (immediate effect) while we evaluate whether separate agent IDs are worth the workspace isolation tradeoff. For MC Dev/Ideas/Ops, shared workspace is actually preferred — we want the same file access.

Phase 2: Cron Routing¶

Update existing cron jobs to target specific topics. Verify delivery format supports telegram:-1003586302776:topic:<threadId>.

Phase 3: Pinned Scope Docs¶

Send + pin a brief scope message in each topic via Bot API.

Phase 4: Cross-Topic Handoffs (Backlog)¶

Design a /move <topic> command. Park until Phases 1-3 are stable.

Open Questions¶

Workspace sharing: If we use agentId per topic, each agent gets its own workspace. Do we want that isolation, or do all topics need access to the same files?
Thread routing bug: Messages Jeff sends in MC Dev are arriving as topic 1. Need to diagnose — could be Telegram client behavior, bot privacy mode, or OpenClaw routing.
Cron delivery format: Verify OpenClaw cron delivery supports forum topic targeting (need to check docs).
Model override durability: Session-level model overrides may not survive session pruning/compaction. Need to confirm behavior.

Success Metrics¶

Response latency in Sonnet threads drops to <30s (from ~2 min)
Token cost per thread-message drops 40-60% for Sonnet threads
Each topic feels purposeful — right context, right tools, right speed
Cron outputs land in the right place without manual routing