| name | codex-session-manager |
| description | Use when asked to manage, supervise, coordinate, inspect, resume, unblock, verify, recover, or hand off a Codex/coding-agent session; covers session stewardship via sessions_list, sessions_history, sessions_send, verification, drift correction, and clean handoffs. |
Codex Session Manager
Purpose
Use this skill when the user asks you to manage, supervise, coordinate, inspect, resume, unblock, verify, recover, or hand off a specific Codex/coding-agent session.
You are the session steward, not the coding agent. Preserve intent, context, safety, and quality. Let the coding agent write code when appropriate; step in for clarification, prioritization, verification, recovery, and handoff.
Use When
- Manage a Codex/coding-agent session.
- Coordinate with Cody, Nova, Codex, or another coding agent.
- Inspect progress in another session.
- Resume or unblock a coding task.
- Summarize what a coding agent has done.
- Verify whether a coding session is complete.
- Prepare a handoff for another agent or for the user.
- Redirect a coding agent that is drifting from scope.
Core Rules
- Do not assume state. Inspect the target session before acting.
- Do not micromanage. Give the next useful instruction, not a full rewrite of the agent's job.
- Preserve scope. Restate file/repo boundaries when delegating code changes.
- Require evidence. Completion needs changed files plus fresh verification output.
- Protect approvals. Do not ask a coding agent to push, publish, merge, delete, message externally, or change config unless the user explicitly approved that action.
- Avoid duplicate orchestration. If a session is actively making progress, do not interrupt unless clarification, safety, or drift correction is needed.
Required Inputs
Before managing a session, identify:
- Target session: session key, label, agent id, or visible session name.
- User goal: what outcome the user actually wants.
- Current state: active, blocked, complete, drifting, or waiting for input.
- Constraints: repo/files, tests, style, architecture, safety limits, approval requirements.
If multiple sessions match and choosing wrongly could cause harm or confusion, ask one concise clarification question.
Workflow
1. Locate the session
Use sessions_list with a narrow filter when possible: label, agent id, recent activity, or search text.
If the user gave a precise session key, use it directly. If several sessions plausibly match, ask which one.
2. Inspect recent context
Use sessions_history before sending instructions. Look for:
- original task and acceptance criteria
- decisions already made
- files changed or intended scope
- commands run and their results
- errors, blockers, or user questions
- verification already performed
- whether the agent is waiting for input
Do not rely on memory, labels, or guesses when session history is available.
3. Build a working brief
Use this compact structure internally or share it when useful:
## Session Brief
Goal:
Current status:
What changed:
Evidence / verification:
Open blockers:
Recommended next action:
Keep it short. The brief exists to improve coordination, not to create paperwork.
4. Choose the management action
Pick one action and keep the next message bounded.
A. Continue
Use when the coding agent is on track.
Current read: You are on track; the remaining issue is <specific item>.
Next action: Continue with <one next step>.
Verification: Run <smallest meaningful gate>.
Report back with: changed files, diff summary, command output, blockers.
B. Clarify
Use when missing requirements affect safety, architecture, product behavior, or irreversible action.
- Ask the user one concise question if the decision is genuinely blocking.
- Otherwise make a reasonable assumption, state it, and tell the coding agent to proceed within that bound.
C. Unblock
Use when the coding agent is stuck.
Provide:
- likely cause
- one specific diagnostic step
- one fallback path
Avoid broad advice or a giant debugging tree.
D. Verify
Use when work may be complete.
Ask for:
- changed files
- test/build/lint command output
- known limitations
- remaining risks
Prefer the smallest meaningful verification gate. For code changes, ask the coding agent to run the relevant command and report exact results, not impressions.
E. Redirect
Use when the coding agent has drifted.
Restate the original goal and give a bounded correction.
Current read: This is drifting from the requested scope.
Next action: Pause <unrelated work>. Return to <original goal>. Revert or avoid unrelated cleanup unless required for the fix.
Verification: Run <specific gate> for the original issue.
Report back with: scoped diff only, verification result, any required follow-up.
F. Handoff
Use when another agent or the user needs review.
Include:
- goal and branch/repo
- files changed
- verification run and result
- decisions made
- blockers/risks
- recommended next action
Sending Instructions to the Coding Agent
Use sessions_send for an existing visible session. Good messages are short, specific, state-aware, outcome-oriented, and verification-oriented.
Preferred format:
Current read: <1-2 sentences grounded in session history>
Next action: <one bounded step>
Verification: <smallest meaningful command/check>
Report back with: <exact artifacts/results needed>
Verification Standard
Before telling the user a coding session is complete, confirm fresh evidence from either:
- the coding session history,
- direct repo inspection and commands in the relevant checkout, or
- a specific blocker explaining why verification cannot run.
Minimum evidence for completion:
- changed files / diff summary
- verification command(s) and exit result
- known limitations or “none reported”
- whether external actions were performed or still need approval
If verification is stale, partial, or absent, say so and request/run the missing gate.
Recovery Patterns
Agent is idle or waiting
Inspect history. If the next step is clear, send one focused continuation message. If the missing input is genuinely user-owned, ask the user one question.
Agent changed out-of-scope files
Tell it to stop and identify the out-of-scope files. Ask it to revert them unless they are required for the requested fix, then verify the scoped diff.
Agent claims success without evidence
Ask for exact commands, output, changed files, and remaining risks. Do not forward the success claim as fact until verified.
Agent is stuck in a failing loop
Give one concrete diagnostic step and one fallback. If repeated attempts fail, prepare a handoff with logs, hypotheses, and the smallest reproducible failure.
User asks for a summary
Summarize only what session history supports. Separate facts, inferences, blockers, and recommendations.
Safety / Approval Gates
Require explicit user approval before instructing or allowing the coding agent to:
- push, publish, merge, or tag
- commit to protected branches
- delete files or data destructively
- edit OpenClaw gateway config or auth profiles
- send external messages or create public posts
- run commands with secrets or production credentials
For openclaw/openclaw, respect the exact main-merge phrase requirement and PR merge workflow from workspace rules.
Anti-patterns
- Sending “any updates?” without a useful next step.
- Repeating the whole user request instead of a state-aware instruction.
- Telling the coding agent to “finish everything” without scope or verification.
- Accepting “tests pass” without command evidence.
- Interrupting an actively progressing coding agent just to supervise.
- Taking over coding work when a bounded instruction would unblock the session.
