openclaw - 💡(How to fix) Fix Performance: Telegram skill invocations (e.g. /weather) pay full workspace bootstrap cost on every turn — 30s+ on Haiku 4.5 [1 comments, 2 participants]

Summary

Trivial Telegram slash-command skill invocations (e.g. /weather) pay the full workspace bootstrap injection cost on every turn even though they don't need most of it. With default config on a Pi 5 + Anthropic Haiku 4.5, end-to-end latency for /weather is ~30 seconds. The only way to bring it under 5 seconds is to set agents.defaults.contextInjection: "never" globally — which then strips USER.md / SOUL.md / IDENTITY.md from all sessions, including the multi-turn DM where they actually carry useful context.

There's no documented way to apply minimal-context only to skill-dispatched turns, the way subagents do (subagents inject only AGENTS.md + TOOLS.md per dist/.../resolveBootstrapContextForRun codepath).

Environment

• OpenClaw 2026.4.15 (041266a)
• Raspberry Pi 5, Debian 13 (trixie), aarch64
• Anthropic provider, primary model claude-haiku-4-5
• Telegram channel via long polling
• Workspace files default-sized: AGENTS.md 7.7k, SOUL.md 1.7k, USER.md ~1.5k, IDENTITY.md 0.6k, TOOLS.md 0.85k, HEARTBEAT.md, plus skills snapshot ≈ 10k → total system prompt ≈ 27k chars per turn

Reproduce

1. Default openclaw onboard --flow quickstart setup with Haiku 4.5 as primary, Telegram channel attached
2. Send /weather from Telegram (built-in weather skill, single-step curl wttr.in tool call + reply formatting)
3. Observe ~25–35 s round trip

Network is verified clean: wttr.in 0.25 s from the Pi, Anthropic API 0.7 s. The cost is OpenClaw’s bootstrap injection + Haiku’s prefill on a 27 k‑char prompt × 2 LLM round trips (decide → tool → format).

Workarounds tried

What's missing

A per‑skill or per‑command context mode, e.g.:

{
  "channels": {
    "telegram": {
      "commands": {
        "weather": { "contextInjection": "never" }
      }
    }
  }
}

…or a reuse of the existing subagent minimal injection (AGENTS.md + TOOLS.md only) for any turn driven by a skill invocation rather than free-form chat. Skills are by design self-contained — their SKILL.md is loaded on dispatch — so the bootstrap stack adds little value but pays full latency cost.

Related

• #9157 — the framework fix that landed contextInjection: never|continuation-skip (closed). Validates the broader concern; this issue is asking for the fine-grained per-skill version.
• #67419 — proposes making continuation-skip the default (open). Helps multi-turn but skill-only invocations still pay the first-turn full cost.

Suggested resolutions

1. Skill-invocation minimal injection — when a turn enters via skill dispatch, inject only AGENTS.md + TOOLS.md (mirror subagent behavior). No new config required, predictable.
2. Per-command override under channels.<provider>.commands.<name>.contextInjection — explicit, opt-in, conservative.
3. Document the latency cost of default bootstrap injection more prominently in docs/concepts/system-prompt so users hitting "my Telegram bot is slow" can find the workaround.