feat: Delivery mirroring - let agents remember what they sent by TSavo · Pull Request #1031 · openclaw/openclaw

TSavo · 2026-01-16T17:26:31Z

🦞 The Problem

When Clawdbot sends a message to Discord/Slack/etc., that delivery is a side-effect that never enters the session transcript. On the next turn, the agent has no memory of what it sent — unless it fetches channel history or reads a log file.

This is particularly painful for cron jobs:

An isolated cron run generates a digest and sends it to Discord
The postToMain summary is just a truncated line, not the full output
If you ask "what did you post this morning?", the agent genuinely doesn't know

The agent literally cannot answer "what did I just say?" without extra tooling.

💡 The Solution: Delivery Mirroring

Add a first-class messages.deliveryMirror config option that mirrors delivered messages back into the session transcript as assistant messages.

{
  "messages": {
    "deliveryMirror": {
      "enabled": true
    }
  }
}

When enabled, every successful outbound send also appends an assistant message to the session transcript with the exact text that was delivered. Future turns naturally include "what I said" in context.

🔧 Implementation

Mirroring hooks added at three layers (covers all send paths):

routeReply — normal agent replies after a turn
message tool — when agents use the message tool to send
Gateway send API — raw gateway sends (with optional sessionKey param)

Also adds isolation.postToMainMode: "full" for cron jobs — isolated runs can now post full output (not just summary) back to the main session.

Files Changed

src/config/sessions/transcript.ts — new appendAssistantMessageToSessionTranscript() helper
src/auto-reply/reply/route-reply.ts — mirroring for normal replies
src/agents/tools/message-tool.ts — mirroring for message tool
src/gateway/server-methods/send.ts — mirroring for gateway API
src/cron/* — postToMainMode option for isolated cron runs
Config types + schemas updated

✅ Testing

Unit tests for appendAssistantMessageToSessionTranscript (5 tests, all passing)
Docker build + isolated gateway test
End-to-end: agent sends via message tool → message appears in transcript
Linter passes (0 warnings, 0 errors)

🤖 AI-Assisted

This PR was built with Claude (Opus). The implementation was tested in a Docker container with an isolated config before committing. I understand what the code does and have verified it works end-to-end.

Why this matters: Agents that can't remember what they said aren't really agents — they're stateless responders. This makes Clawdbot's cron jobs, scheduled digests, and multi-turn workflows actually coherent.

…ript Adds messages.deliveryMirror.enabled config option that mirrors delivered messages back into the session transcript as assistant messages, so the agent can 'remember' what it sent without fetching channel history. Mirroring hooks added to: - routeReply (normal agent replies) - message tool (agent tool sends) - gateway send API (with sessionKey param) Also adds cron isolation.postToMainMode='full' option for isolated cron runs to post full output (not just summary) back to main session. Includes unit tests for appendAssistantMessageToSessionTranscript helper.

chatgpt-codex-connector

💡 Codex Review

https://github.com/clawdbot/clawdbot/blob/1713455dd7c5ccd27d3892a12b1caaf658dc0e03/src/cli/cron-cli/register.cron-add.ts#L180-L184
Wire post-mode/max-chars into cron add payload

The CLI adds --post-mode and --post-max-chars, but the isolation object only sets postToMainPrefix before calling cron.add. As a result, postToMainMode and postToMainMaxChars are never sent to the gateway, so users who pass --post-mode=full (or change max chars) will still get the default summary behavior with no warning. This makes the new flags silently ineffective for isolated jobs.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

…yload Addresses Codex review feedback - the CLI options were defined but not actually passed through to the isolation object when calling cron.add.

steipete · 2026-01-17T00:31:46Z

Thanks! Taking over. Slopus missed some stuff :)

Findings

Blocker: src/config/sessions/transcript.ts:45-51 appends { message }
without type/id/parentId/timestamp. Session format requires type:"message"
and IDs; readers filter on type === "message" (see session tests).
Mirrored lines likely ignored → feature doesn’t work. Use
SessionManager.appendMessage(...) or write full SessionMessageEntry with
parentId chain.
High: src/agents/tools/message-tool.ts:192-214 mirrors on bestEffort sends
even if delivery fails (send result can be undefined). This can “remember”
messages never delivered. Gate on result.sendResult?.result (or equivalent
success signal) before mirroring.
Medium: src/config/sessions/transcript.ts:22-24 uses cached session store;
cross‑process updates can be stale, returning “unknown sessionKey” and
skipping mirror for newly created sessions. Use
loadSessionStore(storePath, { skipCache: true }) or updateSessionStore on
this write path.

steipete · 2026-01-17T01:48:31Z

Squashed and landed on main: 6c79964bc6c68edefc4311f05c9e50dfce814c0b. Tests: pnpm lint, pnpm build, pnpm test.

Co-authored-by: T Savo <TSavo@users.noreply.github.com>

steipete · 2026-01-17T02:03:27Z

Rebased main + re-landed squash: fdaeada3e3c72ef9b4c5a2f38cf6d31a1844f765. Tests: pnpm lint, pnpm build, pnpm test.

Heartbeat-initiated outbound messages were not being mirrored back into the session transcript, causing context confusion when users replied to heartbeat messages. The agent had no memory of what it sent. This adds the 'mirror' parameter to the deliverOutboundPayloads call in runHeartbeatOnce, matching the pattern established by openclaw#1031 for other delivery paths (routeReply, message tool, gateway send API). Fixes the case where: 1. Heartbeat triggers agent to send a message (e.g., "was machst du grad?") 2. User receives and replies to that message 3. Agent has no context about the original message it sent 4. Agent is confused about what the user is replying to Now the outbound heartbeat text/media is recorded in the session transcript, giving the agent the context it needs.

Co-authored-by: T Savo <TSavo@users.noreply.github.com>

chatgpt-codex-connector bot reviewed Jan 16, 2026

View reviewed changes

TSavo added 2 commits January 16, 2026 09:45

fix: wire --post-mode and --post-max-chars into cron add isolation pa…

6557497

…yload Addresses Codex review feedback - the CLI options were defined but not actually passed through to the isolation object when calling cron.add.

fix: regenerate protocol bindings and fix formatting

4aa930e

steipete self-assigned this Jan 17, 2026

steipete added a commit that referenced this pull request Jan 17, 2026

feat: mirror delivered outbound messages (#1031)

fdaeada

Co-authored-by: T Savo <TSavo@users.noreply.github.com>

steipete closed this Jan 17, 2026

rohan-ixlayer mentioned this pull request Jan 24, 2026

iMessage self-chat creates echo loop despite v2026.1.22 fixes #1649

Closed

PatrickBauer mentioned this pull request Feb 6, 2026

fix(heartbeat): mirror outbound messages to session transcript #10677

Open

rmarr mentioned this pull request Feb 24, 2026

[Bug]: iMessage + /reasoning on creates infinite agent feedback loop #25897

Closed

zooqueen pushed a commit to hanzoai/bot that referenced this pull request Mar 6, 2026

feat: mirror delivered outbound messages (openclaw#1031)

33d7619

Co-authored-by: T Savo <TSavo@users.noreply.github.com>

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

feat: Delivery mirroring - let agents remember what they sent#1031

feat: Delivery mirroring - let agents remember what they sent#1031
TSavo wants to merge 3 commits intoopenclaw:mainfrom
TSavo:delivery-mirror

TSavo commented Jan 16, 2026

Uh oh!

chatgpt-codex-connector bot left a comment

Uh oh!

steipete commented Jan 17, 2026

Uh oh!

steipete commented Jan 17, 2026

Uh oh!

steipete commented Jan 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Conversation

TSavo commented Jan 16, 2026

🦞 The Problem

💡 The Solution: Delivery Mirroring

🔧 Implementation

Files Changed

✅ Testing

🤖 AI-Assisted

Uh oh!

chatgpt-codex-connector bot left a comment

Choose a reason for hiding this comment

💡 Codex Review

Uh oh!

steipete commented Jan 17, 2026

Uh oh!

steipete commented Jan 17, 2026

Uh oh!

steipete commented Jan 17, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants