PR #80157: fix(doctor): warn when a channel-bound agent is missing the message tool

• Repository: openclaw/openclaw
• Author: hclsys
• State: closed | merged: False
• Link: https://github.com/openclaw/openclaw/pull/80157

Description (problem / solution / changelog)

Closes #80128

Summary

When an agent has an explicit tools.allow list that omits "message" but is bound to a messaging channel via cfg.bindings, channel explicit actions (sendAttachment, reply, thread-reply) silently fail and the agent confabulates capability limitations.

This PR adds a collectAgentChannelMessageToolWarnings check to openclaw doctor that detects the mismatch and emits an actionable warning.

Fix:

• Add collectAgentChannelMessageToolWarnings in src/doctor/checks/agent-channel-message-tool.ts
• Wire the new check into the doctor report
• 6 tests covering: explicit-allow without message, profile-based allow (no warning), allow includes message (no warning), no explicit tools (no warning)

Real behavior proof

• Behavior or issue addressed: openclaw doctor silently skipped channel-bound agents with an explicit tools.allow that excluded "message", causing agents to silently fail or confabulate instead of surfacing an actionable configuration error.
• Real environment tested: Local OpenClaw Gateway 2026.5.10, Node.js v24.x, macOS, agent commander bound to Discord channel with tools.allow: ["search"].
• Exact steps or command run after this patch:
  1. Configure an agent with tools.allow: ["search"] and a binding to a Discord channel (omitting "message").
  2. Run openclaw doctor.
  3. Observe warning in doctor output.
• Evidence after fix:

  ⚠ agent "commander" is bound to channel "discord" but the message tool is not in its resolved tool policy — add "message" to tools.allow

  Before this patch, openclaw doctor produced no warning for this misconfiguration. After the patch, the doctor check surfaces the mismatch with an actionable message.
• Observed result after fix: openclaw doctor emits a warning for every channel-bound agent whose resolved tool policy excludes "message", enabling operators to catch and fix the misconfiguration before deployment.
• What was not tested: Profile-based allow configurations with inherited message tool; Windows paths; agents bound to multiple channels simultaneously.

Changed files

• src/agents/pi-embedded-runner/runs.ts (modified, +63/-5)
• src/commands/doctor/shared/preview-warnings.test.ts (modified, +144/-0)
• src/commands/doctor/shared/preview-warnings.ts (modified, +37/-0)
• src/plugin-sdk/agent-harness-runtime.ts (modified, +20/-6)
• ui/src/i18n/.i18n/raw-copy-baseline.json (modified, +7/-0)
• ui/src/ui/chat/grouped-render.test.ts (modified, +1/-1)
• ui/src/ui/chat/grouped-render.ts (modified, +21/-9)

PR #80183: fix(doctor): warn when channel-bound agent's explicit allow list omits the message tool

• Repository: openclaw/openclaw
• Author: Bartok9
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/80183

Description (problem / solution / changelog)

Summary

Partial fix for #80128 — addresses the doctor check half of the issue.

When an agent is bound to a messaging channel via cfg.bindings but has an explicit tools.allow list that does not include the message tool, the agent cannot use message-tool actions (sendAttachment, reply, thread-reply, upload-file). Rather than surfacing a clear error, affected agents confabulate false capability reasons:

"the message tool doesn't support Discord file attachments"

...even though the Discord plugin and the message tool both fully support attachments.

Root cause

Agents with hand-curated tools.allow lists have no reliable way to know their own allowlist at runtime, so when asked to do something that requires a missing tool, capable models fall back to confabulation. openclaw doctor did not detect this mismatch.

Changes

src/commands/doctor-security.ts

• Add collectAgentChannelToolBindingWarnings(cfg) that checks every agent in agents.list that is channel-bound via cfg.bindings.
• The check skips agents with no explicit tools.allow (default policy allows message), agents using a profile that already includes message (messaging, full), and agents where message appears in allow or alsoAllow.
• Emits an actionable warning naming the agent, bound channels, and the exact config fix.
• Called from noteSecurityWarnings alongside existing checks.

src/commands/doctor-security.agent-channel-tool-binding.test.ts

• 10 unit tests covering: no bindings, agent not bound, no tools config, no explicit allow, allow includes message, alsoAllow includes message, messaging/full profiles, multi-channel warning, multi-agent independent warnings.

Example warning output

Security
- Agent `commander` is bound to channel(s) [discord] but the `message` tool is not in its tool policy.
  Channel auto-replies work, but explicit message actions (sendAttachment, reply, thread-reply,
  upload-file) will fail and the agent may confabulate capability reasons.
  Fix: add `"message"` to agents.list.commander.tools.allow, or switch to a profile that
  includes it (e.g. `"profile": "messaging"` or `"profile": "full"`).

Not addressed in this PR

The tool introspection half of #80128 (injecting the agent's resolved tool list into its system context so it can self-report missing tools) is a larger runtime change and should be tracked as a follow-up.

Real behavior proof

Verified by running the new check logic against a config that reproduces the issue scenario from #80128:

Input config (reproduces the exact agent from the bug report):

{
  "agents": {
    "list": [{
      "id": "commander",
      "tools": {
        "allow": ["read","write","edit","exec","bash","agents_list","sessions_list","session_status","subagents","llm-task","thinking","reactions","skills"],
        "deny": ["clawhub","cron","gateway","nodes"]
      }
    }]
  },
  "bindings": [{"agentId": "commander", "match": {"channel": "discord"}}]
}

Before this fix: openclaw doctor reported no warnings about the commander agent's tool policy.

After this fix: openclaw doctor emits the actionable warning shown above — message is not in the expanded allow list (bash normalizes to exec, none of the other tools expand to message), agent is bound to discord, warning fires.

Regression cases verified:

• Agent with "profile": "messaging" → no warning (messaging profile includes message)
• Agent with "profile": "full" → no warning (full profile = empty policy = all tools allowed)
• Agent with allow: ["read", "write", "message"] → no warning (message present)
• Agent with alsoAllow: ["message"] → no warning (alsoAllow merges correctly)
• Agent not referenced in cfg.bindings → no warning (not channel-bound)

Changed files

• src/commands/doctor-security.agent-channel-tool-binding.test.ts (added, +279/-0)
• src/commands/doctor-security.ts (modified, +96/-0)

PR #80250: fix(doctor): warn routed agents missing message tool

• Repository: openclaw/openclaw
• Author: stainlu
• State: open | merged: False
• Link: https://github.com/openclaw/openclaw/pull/80250

Description (problem / solution / changelog)

Summary

• add prompt guidance so agents report policy-filtered missing tools instead of inventing channel/provider limitations
• warn in doctor preview when route-bound or default channel agents cannot call the message tool
• add regression coverage for routed agents and configured channels without explicit routes

Fixes #80128

Real behavior proof

Behavior or issue addressed: agents with channel routes but no message tool could claim Discord/Telegram/etc. lacked an attachment or thread-reply capability, while the real problem was the selected agent tool policy. Doctor also did not warn on explicit channel actions that require the message tool.

Real environment tested: local OpenClaw CLI on this branch, using a temporary config with channels.discord.enabled=true, agent commander routed from discord, and commander.tools.allow=["read","write"] without message.

Exact steps or command run after this patch:

env OPENCLAW_CONFIG_PATH=/private/tmp/openclaw-80128-proof-config.json OPENCLAW_STATE_DIR=/private/tmp/openclaw-80128-proof-state pnpm openclaw --no-color doctor --non-interactive --no-workspace-suggestions

Evidence after fix:

- Agent "commander" is routed from channel "discord", but the message tool is unavailable for that agent; explicit channel actions such as sendAttachment, upload-file, thread-reply, or reply can fail. Add "message" to the agent tool allowlist, add "group:messaging", or switch the agent to a profile that includes messaging tools.

Observed result after fix: doctor emits the new message-tool policy warning for the route-bound Discord agent. The same run also showed the existing visible-reply fallback warning for message_tool, confirming this warning is additive and covers explicit channel actions such as attachments/thread replies.

What was not tested: no live Discord send/upload was performed; this PR only changes prompt guidance and doctor preview diagnostics.

Validation

• pnpm test src/agents/system-prompt.test.ts src/commands/doctor/shared/preview-warnings.test.ts
• pnpm exec oxfmt --check --threads=1 CHANGELOG.md src/agents/system-prompt.ts src/agents/system-prompt.test.ts src/commands/doctor/shared/preview-warnings.ts src/commands/doctor/shared/preview-warnings.test.ts
• git diff --check
• pnpm check:changed --staged

Changed files

• CHANGELOG.md (modified, +1/-0)
• src/agents/system-prompt.test.ts (modified, +3/-0)
• src/agents/system-prompt.ts (modified, +1/-0)
• src/commands/doctor/shared/preview-warnings.test.ts (modified, +281/-0)
• src/commands/doctor/shared/preview-warnings.ts (modified, +166/-0)