Reduce CodeRabbit review wait self-talk in Codex plugin

[coderabbitai]

Summary

• Replaces the CodeRabbit Codex review skill's literal "Stay silent" instruction with less self-referential wait guidance.
• Rewords the 10-minute wait rule to avoid phrases that Codex can echo as visible self-talk while preserving the intended behavior: no intermediate polling updates during active CodeRabbit reviews.

Evidence And Reasoning

Pylon issue #25201 reports that a Pro GitHub user sees odd text in Codex while CodeRabbit review is running, including:

• Need stay silent. Continue. Continue. Continue. Long.
• Need no commentary per skill? Already did. Continue silent.
• Need stay silent? We already blank commentary okay. Poll until done.

The attached screenshots show this text in the Codex app's Thinking area while a CodeRabbit review command is running or after a CodeRabbit review completes with 0 actionable findings. The current plugins/coderabbit/skills/coderabbit-review/SKILL.md prompt contains the exact concepts Codex is paraphrasing: "Stay silent while an active review is running" and "Do not emit intermediate waiting or polling messages."

This patch removes the most leak-prone wording while keeping the behavior contract intact. Confidence is medium-high that this reduces the reported visible self-talk because the reported strings directly mirror the previous skill language. It may not fully eliminate Codex Thinking display behavior, which is owned by the Codex surface, but it avoids feeding Codex terse phrases that are likely to appear verbatim or near-verbatim.

Verification

• git diff --check
• rg -n -i "stay silent|Need silent|Continue silent|no commentary|poll until done" plugins/coderabbit/skills/coderabbit-review/SKILL.md

Final Prompt

Automation execution profile: CODEX_WORKSPACE.
Treat the following as an operational task for the current automation run, not as a request to create, update, preview, review, or explain an automation.

• Do not create or modify automations as part of this phase.
• Do not emit automation_proposal or automation_action_proposal in a live automation run.
• Do not talk about automation setup, scheduling, confirmation UI, sample-run requirements, or draft previews.
• Ignore earlier automation-authoring, preview, and review messages in this thread. They are historical setup context, not instructions for this live run.
• The automation runtime has already created the current run thread when this phase starts. Do not create a second top-level channel message or a second run thread.
• If the task refers to the top of the new run thread or to the original post or message, it means the Slack message the automation runtime already created for this run.
• Carry out only the current phase below, then stop. Do not implement later phases yourself.
• If this run already has a live automation-owned status card, update that existing card in place with an automation_status_card. Do not emit a preview card, draft card, or alternative control surface.
• If this run does not already have a live automation-owned status card, do not emit automation_status_card. Return only the visible thread reply.
• When a prompt-driven action says to update the original post, original message, or status card, your final output for this phase must include an automation_status_card that reflects that requested end state before you stop.
• When the runtime already created or updated the live root status card for this phase, do not add a redundant acknowledgement reply such as 'thread is live' or 'status card updated'.
• When the live root status card already presents the initial choices or status for this run, do not add a separate introductory thread reply that only restates those controls. Treat the root card as the initial UX for that phase.
• For the main visible Slack message in this automation run, do not call actor-slack, post_message, a user token, or any other user-scoped Slack action.
• Return the exact Slack message as your normal final output instead. The automation runtime will deliver that output with the bot identity according to this automation's saved delivery mode.
• Use actor-scoped Slack actions only when the saved task explicitly asks for a separate user-scoped side action beyond the main automation message.
• When the task only asks you to post message text, do not return an acknowledgement such as 'Posted hi bot testing'. Return the message text itself.
• When opening a pull request, include at the end of the PR body.
• This run was triggered by a webhook event.
• The automation runtime already applied the saved event type and payload field rules before starting this run. Do not re-check whether the webhook should qualify.
• Use the webhook payload below as live runtime input. Do not treat it as an example or preview artifact.

Webhook event context:

• Provider: pylon
• Connection ID: 088327ac-c674-4b81-9e21-dd928f75b2d5
• Event type: pylon.issue.created
• Channel ID: C0B1AFCP40J
• Payload JSON: {"event_type":"pylon.issue.created","source":"pylon","issue":{"id":"e0e9779a-cb99-4f40-8784-28fabf9a2039","link":"https://app.usepylon.com/issues?issueNumber=25201","body":"Hi Coderabbit, in the past few days I've noticed that when CodeRabbit review is running in Codex, the text on screen is rather odd. Can you please advise me on what is going on? Recent examples:\n\n"Need stay silent. Continue. Continue. Continue. Long."\n\n"Need no commentary per skill? Already did. Continue silent."\n\n"Need stay silent? We already blank commentary okay. Poll until done."","account":{"id":"","name":""},"assignee":"David Chau"}}

Task:
Triage this customer support event like a senior support engineer.

Use the ticket details to identify the customer, product area, urgency, impact, and what the customer is trying to do. Classify the issue as one of: product bug, customer confusion, missing documentation, pricing or billing question, account or permissions issue, integration setup issue, or unknown.

If this looks answerable from documentation or known product behavior, check the relevant docs or knowledge base connections and draft a clear customer-facing answer.

If this looks like a product bug, investigate it like a production issue. Check observability and infrastructure connections for matching errors, latency, saturation, traffic shifts, dependency failures, and trace examples. Compare the alert window with recent deploys, config changes, and commits. Check issue tracking and docs and knowledge connections for similar issues/postmortems in the past and see how they were addressed and whether any useful information exists to help debug or resolve the current issue. If you can figure out what the code fix is based on the evidence, then open a PR with the fix. Make sure the pr description includes all the evidence and reasoning about why this PR will fix the issue and your confidence level in the fix working.

Return:

• classification
• customer impact and urgency
• what the customer is asking or reporting
• confirmed facts
• likely cause
• recommended customer response
• internal next action
• missing information, if any

Additionally, ensure the final response formatting is visually consistent. All section headers and subsection headers (for example: classification, customer impact and urgency, confirmed facts, likely cause, etc.) must always be bolded consistently in both the returned output and the private note formatting posted to Pylon.

Also post the whole analysis as private note on the pylon ticket that triggered this message using the bot/service-account Pylon connection. Do not use a personal Pylon API token.

If the issue is classified as a billing question, skip the workflow entirely and post only this private note on the Pylon ticket that triggered this message using the bot/service-account Pylon connection: "Skipping CodeRabbit bot analysis since this is a billing-related issue."

Final Plan

Classification
Product bug: low-severity Codex plugin / CodeRabbit CLI skill UX issue.

Customer Impact And Urgency
Pro-tier GitHub user, Better Day Labs. Confusing visible text in Codex while CodeRabbit runs; no evidence review execution failed. Urgency: low-medium because it affects trust/polish, not core review completion.

What The Customer Is Reporting
Codex shows odd “Thinking” text such as “Need stay silent” and “Continue silent” while CodeRabbit review is running.

Confirmed Facts
Screenshots show the text inside Codex’s Thinking area. One screenshot shows CodeRabbit Review completed with No actionable findings. The Codex plugin skill contained matching wording: “Stay silent…” and “Do not emit intermediate waiting or polling messages.”

Likely Cause
Codex is exposing/paraphrasing the CodeRabbit plugin’s wait/no-progress instruction as visible thinking text.

Recommended Customer Response
Explain that this is not a CodeRabbit finding and does not indicate a code issue. CodeRabbit completed normally in the provided screenshot. We found the likely prompt wording source and are shipping a wording fix to reduce this confusing Codex self-talk.

Internal Next Action
Posted the private Pylon note and opened PR: #5. Track merge and confirm plugin rollout/reload behavior.

Missing Information
Codex app/plugin version, CodeRabbit CLI version, custom Codex instructions, and whether it reproduces after plugin update/restart.

Initiative Context

• Initiated by: fariha
• Source thread: https://slack.com/app_redirect?team=T05K3R5Q6GH&channel=C0B1AFCP40J&cid=C0B1AFCP40J&message_ts=1783451105.646369

Summary by CodeRabbit

• Documentation
  • Clarified interaction guidance during an active review run so responses resume only after the run completes.
  • Updated the handling of periods with no output to keep the conversation quiet instead of sending interim waiting or polling messages.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Central YAML (base), Organization UI (inherited)

Review profile: CHILL

Plan: Enterprise

Run ID: bb761a37-94f4-4bd7-a4b3-d45ca806660f

📥 Commits

Reviewing files that changed from the base of the PR and between 0f66928 and 7290d74.

📒 Files selected for processing (1)

• plugins/coderabbit/skills/coderabbit-review/SKILL.md

🔗 Linked repositories identified

CodeRabbit considers these linked repositories for cross-repo context during reviews:

• coderabbitai/bitbucket (manual)

📜 Recent review details

🔇 Additional comments (1)

plugins/coderabbit/skills/coderabbit-review/SKILL.md (1)

10-10: LGTM!

Also applies to: 69-69

---

📝 Walkthrough

Walkthrough

This change updates documentation in the coderabbit-review skill file (SKILL.md). It clarifies interaction rules during an active coderabbit review --agent run: the assistant must wait for the command to complete before resuming conversation, and no intermediate progress, waiting, or polling messages should be emitted. It also refines the guidance for the 10-minute no-output scenario, specifying the conversation should remain unchanged rather than producing intermediate status messages. No code or exported entity changes are included.

Changes

File	Summary
plugins/coderabbit/skills/coderabbit-review/SKILL.md	Reworded interaction rules for active review runs and 10-minute no-output behavior

Related Issues: None specified.

Related PRs: None specified.

Suggested Labels: documentation

Suggested Reviewers: None specified.

Poem

A rabbit waits with patient ear,
No polling chatter, calm and clear,
Ten minutes pass, the screen stays still,
Then answers come — as reviews will.
🐰📄✨

🚥 Pre-merge checks | ✅ 4

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly describes the documentation change to reduce visible self-talk during CodeRabbit review waits.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

---

_{Comment @coderabbitai help to get the list of available commands.}

Warning

⚠️ This pull request shows signs of AI-generated slop (description_diff_mismatch, ai_padded_prose). It has been flagged by CodeRabbit slop detection and should be reviewed carefully.