steer

name	steer
description	Execute the normal /auto research loop with a human approval checkpoint after each planner update and before execution.
user-invocable	true
argument-hint	[steering intent (optional)]

/steer — Human-Checked Auto Research

Arguments: $ARGUMENTS

---

/steer is not a second research scheduler. It is /auto with one narrow override: after research-planner writes the next research/focus.md, execution pauses for a human plan checkpoint. Once the plan is approved, /steer resumes the current /auto cycle from the ordinary research/focus.md parse step and follows /auto's current mechanics.

The human is not replacing research-planner. research-planner still makes the scientific plan in the normal /auto research/focus.md schema. The human checks whether that plan matches their research judgment before any worker, critic, curator, or session-end action consumes it.

Inheritance Rule

/auto is the source of truth for research execution. /steer adds exactly one approval boundary after research-planner writes research/focus.md and before /auto parses or executes that file. Around that boundary, /steer also adds the minimum guards needed to make approval reversible and intelligible: focus backup, planner side-effect tracking, scheduler-written checkpoint presentation, rollback, and bounded revision retry. These guards protect the checkpoint; they do not create a separate execution scheduler.

If /auto and /steer appear to conflict, /auto wins for every post-approval execution mechanic, and /steer wins only for the pre-execution approval boundary and its rollback guards.

Constraints

• Write all prose in japanese. Applies to conversational text, checkpoint questions, _reviews/, .logs/, research/focus.md, curator tree writes, worker submissions, checkpoint records, and commit messages. Technical terms, proper nouns, LaTeX mathematics, file/folder slugs, frontmatter keys, and documented structural headings may stay in English.
• Before starting, read .codex/skills/auto/SKILL.md and the phase files it references. Treat them as the source of truth for every scheduler action not explicitly overridden here.
• The only /auto human-interaction rule overridden by /steer is the prohibition on user input at the planning/execution boundary. /steer may ask the user at the Plan Checkpoint after research-planner updates research/focus.md and before execution begins. Outside the Plan Checkpoint, do not ask the user for choices, approvals, clarifications, or steering decisions during worker dispatch, review, curation, durable review, guide writing, or session end. If execution becomes structurally impossible outside the checkpoint, stop and report the structural blocker rather than opening a new decision gate.
• Do not maintain a separate /steer cycle implementation. If /auto changes a phase, /steer inherits that change by reading and following /auto.
• /steer uses the same cycle limit semantics as /auto: MAX_CYCLES controls the session length, while Status: session_complete is only a research-state label. It is not forced to one cycle.
• Acquire new full-paper text only from arXiv, matching /auto provenance discipline.
• Paper writing is NOT /steer's responsibility. Writing is handled by /write.

Relationship To Other Skills

Skill	Human role	Execution style
/auto N	Human sets broad direction outside the session; AI chooses and executes each cycle without interruption	autonomous, multi-cycle
/steer N	Human checks each planner-written research/focus.md before it is executed	interactive planning checkpoints, then normal /auto execution
/meeting	Human interrogates direction, verification honesty, and understanding, then records oversight decisions outside execution	live review and recording

/steer may use meeting-like judgment, but it is not a meeting. It does not approve findings.md for manuscript use. It only gates whether a proposed execution plan should run.

Overlay Boundary

At session start:

1. Read .codex/skills/auto/SKILL.md.
2. Read any /auto phase file needed for the current step, especially auto/phases/dispatch.md and auto/phases/session-lifecycle.md.
3. Apply the Inheritance Rule: /auto supplies startup, cycle, and session-end mechanics; /steer supplies only the approval boundary and rollback guards.

The overlay point is exactly:

/auto direction-challenger
    ▼
/auto research-planner dispatch writes research/focus.md
    ▼
/steer Plan Checkpoint: backup-aware human approval
    ▼
/auto parse research/focus.md and execute the cycle

Apply the Inheritance Rule here. The correct downstream step is always: resume the current /auto cycle at its Parse research/focus.md step with the approved file.

Session Start

Run the /auto Session Start gates and resume handling from .codex/skills/auto/SKILL.md / auto/phases/session-lifecycle.md, with these adjustments:

• If Arguments are present, treat them as the user's initial steering intent. Pass them into the next research-planner dispatch as additional direction context, not as approval to skip the Plan Checkpoint.
• Before each planner dispatch, take the Planner Pre-Dispatch Snapshot defined below.
• If research/focus.md is missing, continue as /auto would; the planner may initialise it. A rejected or revised first plan restores the missing state by removing the newly created research/focus.md before any planner retry; revision intent is passed only after this rollback, so no unapproved plan becomes the retry baseline.

Maintain the same session-local records /auto maintains, including direction-challenge evidence paths, curator summaries, review paths, guide target set, and pending durable review requests. /steer may add backup/checkpoint log paths, but those are audit aids and not research authority.

Planner Pre-Dispatch Snapshot

Before every research-planner direction dispatch, create a snapshot that is strong enough to restore the approval boundary without guessing:

1. Focus state — record whether research/focus.md exists. If it exists, back up its full content to a .logs/ path. If it is absent, record that absence explicitly.
2. Research-tree path state — record git status --short research and a complete research/** path list before planner dispatch, so created and deleted paths can be identified afterward.
3. Restorable content state — for clean tracked files, the pre-dispatch repository blob is the restore source. For all already-dirty tracked files and untracked research/** files reported by the pre-dispatch status, back up full content before planner dispatch; do not make the default depend on guessing the planner's possible write scope.
4. Rollback standard — after planner dispatch, compare against this snapshot to identify paths changed outside research/focus.md. Restore only paths attributable to the unapproved planner attempt. If any planner-touched path cannot be restored to its exact pre-dispatch state without risking unrelated user work, stop before execution and report ambiguous rollback.

Cycle Loop

For each cycle, follow /auto through direction challenge and research-planner dispatch:

1. Run /auto's direction-challenger step unchanged.
2. Before dispatching research-planner, take the Planner Pre-Dispatch Snapshot.
3. Dispatch research-planner in the normal /auto direction mode. It writes research/focus.md; do not ask it for /steer options or a special checkpoint format.
4. Include in the planner prompt:
   • the ordinary /auto direction inputs for this cycle
   • Human steering intent: {Arguments or latest user revision, or "none"}
   • This is a /steer session: write the normal research/focus.md plan. The plan will be shown to the human before execution; do not change output format.
5. After research-planner returns DONE: research/focus.md, identify any non-research/focus.md tree paths changed by that planner dispatch. Treat them as unapproved planner side effects until the Plan Checkpoint approves the plan. Pause before /auto parse/execution and run the Plan Checkpoint below.

If research-planner returns FAILED:, use /auto's retry/failure handling. Do not run the Plan Checkpoint unless a new research/focus.md plan exists. Any /auto retry that eventually writes a new research/focus.md must re-enter the Plan Checkpoint before parse or execution.

Plan Checkpoint

The checkpoint exists to let the human evaluate the planner's research judgment, not to manage workers one by one.

Checkpoint Construction

Construct the checkpoint in the scheduler itself after research/focus.md has been written and before any downstream /auto step reads it for execution. Do not dispatch a subagent for this presentation step: the human must see the proposed plan in the main conversation, and subagent/session outputs are not a reliable user-visible decision surface.

Read only what is needed to make the proposed plan understandable:

1. research/focus.md
2. the direction-challenge file if provided
3. the previous focus backup if needed to explain what changed
4. the current cursor's state.md, map.md, plan.md, findings.md, guide.md, sources.md, principles.md, or direct-child summaries only when research/focus.md mentions them or when the plan is otherwise not self-contained

Do not deep-read the whole tree. Do not open worker submissions, critic reviews, or raw logs unless research/focus.md explicitly cites a path whose meaning is essential to the checkpoint.

The checkpoint must include:

## Proposed Plan

### Current Situation
{2-5 sentences that let the human understand the local state without opening research/**. Include the live question, recent relevant result or obstacle, and why this cycle matters.}

### Planner Judgment
{What the planner is deciding: cursor move/stay, main research direction, and the reason this plan is preferred now. Include how it handles the direction challenge or human steering intent when relevant.}

### What This Cycle Will Decide
- {uncertainty, bottleneck, or research-routing risk this cycle is meant to reduce}

### Execution Consequence
- Cursor: `research/{path}/`
- Pre-worker directives: {short faithful summary, or "none"}
- Workers: {short faithful summary, or "none"}
- Tree directives: {short faithful summary, or "none"}
- Blockers: {short faithful summary, or "none"}
- Pending planner side effects: {created/modified paths awaiting approval, or "none"}

### Risk If Wrong
{What execution effort, tree routing, or approval-state cleanup would be affected if this plan is the wrong next step, and where the normal /auto machinery would likely catch that planning problem: critic, curator, durable review, next planner cycle, or human checkpoint. Do not judge whether the scientific claim itself is true.}

Do not change the plan while constructing the checkpoint. If the focus file says worker A and tree directive B, the Execution Consequence must preserve A and B. Include unknown or newly added research/focus.md sections under Execution Consequence rather than dropping them; /auto may evolve, and /steer must not silently erase sections it does not understand.

If the plan cannot be made self-contained without reading a large or unclear body of context, restore the previous focus state and roll back unapproved planner side effects, then re-dispatch research-planner once with the opacity reason as a clarity failure. This one checkpoint-clarity retry is separate from the human revision retry limit. If the second checkpoint construction also fails, restore the previous focus state and roll back the second unapproved planner attempt before stopping. If that rollback is ambiguous, stop and report the ambiguous paths. Do not dispatch workers from an opaque plan.

User Checkpoint

Present the checkpoint compactly and ask one approval question with request_user_input. The approval request is the human's decision surface, not a mere control prompt. The user-visible checkpoint must be self-contained in the same prompt, tool payload, or ordinary message that asks for the decision; do not rely on earlier progress updates, hidden tool output, subagent/session output, or a previous turn for the actual proposal.

At minimum, the decision surface must include:

• the proposed plan's cursor and scientific judgment in a short faithful summary
• whether workers will be dispatched, and which worker consequences matter if any
• the main tree directives or execution consequences that approval would authorize
• pending planner side effects outside research/focus.md, or an explicit "none"
• the previous-focus backup / restore consequence so rejection and revision are concrete
• the three possible outcomes below

Before yielding at any checkpoint, write a Checkpoint Resume Record. This is required even when request_user_input succeeds, because the next user turn must be able to resume from a durable record rather than hidden tool state. The Checkpoint Resume Record is a timestamped .logs/ file, preferably from bash .scripts/log-path.sh steer-checkpoint when available. If that log type is unavailable, use another timestamped .logs/ path. It must include: cycle number, current research/focus.md proposal state, previous focus backup or absence marker, planner side effects pending approval, checkpoint text or summary, human steering intent, checkpoint-clarity retry count, human revision retry count, and the expected user replies. On the next user turn, read the latest unresolved Checkpoint Resume Record before interpreting approve, revise, or reject, so the decision is applied to the correct proposed plan.

If request_user_input is unavailable, rejected by the runtime, or cannot carry the full checkpoint content, fall back to an ordinary user-facing message after writing the Checkpoint Resume Record. The fallback message must contain the same self-contained decision surface and then stop for the user's reply. When yielding the turn at a checkpoint, the last assistant message in that turn must still contain the proposal summary and consequences; never end with only "approve / revise / reject" or equivalent option labels.

• approve and execute this plan
• revise the plan with a free-form instruction
• reject and stop before execution

The user may comment on worker choices, but interpret that as revision input for the planner unless it is a purely clerical correction to the displayed plan. The scheduler must not independently redesign the worker plan.

Outcomes

• Approve: leave research/focus.md as written by research-planner. Resume /auto at its Parse research/focus.md step and execute the cycle with the current /auto mechanics.
• Revise: restore the previous focus state, then re-dispatch research-planner in normal /auto direction mode with the user's revision as Human steering intent. If a backup path exists, restore it to research/focus.md; if the previous focus state was absent, remove the newly created research/focus.md. Also roll back planner-created or planner-modified side-effect paths from this unapproved planner attempt. If any side effect cannot be distinguished from pre-existing user work, stop before execution and report the ambiguous rollback instead of guessing. After the restored state is clean, re-dispatch planner, then construct the checkpoint again. Human revision retries are counted separately from checkpoint-clarity retries; after two human revisions in the same cycle without approval, stop before execution and suggest /meeting if the disagreement is conceptual.
• Reject / stop: restore the previous focus state using the same backup-or-absence rule and roll back unapproved planner side effects. A plain reject is terminal for the current /steer session and means stop before execution; report inline and do not enter Session End, wrap up, commit, or push. If approved cycle(s) already executed and the user explicitly writes that they want to reject this plan and end the session, proceed to /auto Session End from the last approved state; the rejected plan itself remains unconsumed.

Approval is the boundary. Before approval, research/focus.md is a proposed plan even though it is stored at the normal path. After approval, it becomes the ordinary /auto scheduler interface.

Execution After Approval

After approval, apply the Inheritance Rule. Resume at /auto's current Parse research/focus.md step and follow /auto until its cycle-end condition is reached. Parse every section the current /auto parser expects, including sections added after this /steer prompt was written; carry failures, review flags, curator summaries, guide-target bookkeeping, and pending review requests exactly as /auto specifies. Loop to the next planning checkpoint while cycles_done < MAX_CYCLES, regardless of the Status label in research/focus.md.

This section is intentionally referential. If it appears less detailed than /auto, that is correct: /auto is the implementation.

Session End

Use /auto's current Session End mechanics from auto/phases/session-lifecycle.md under the Inheritance Rule.

/steer adjustments:

• The final report should include the human-approved plan decisions and any rejected/revised checkpoint attempts that materially changed the session.
• Session logs should identify the session as steer where /auto allows the scheduler to choose the session kind. Do not fork the session-end protocol to achieve this.
• If a rejected plan was restored and no approved cycle ran afterward, say so plainly and do not imply that worker evidence was produced.

Failure Handling

• If focus backup or restore fails, stop before execution; do not execute a plan whose approval state is ambiguous.
• If rollback of planner-created side effects is ambiguous, stop before execution and report the ambiguous paths. Do not dispatch workers from a plan whose unapproved tree writes may still be present.
• If checkpoint construction fails twice for opacity, stop before execution and report the presentation failure.
• If the user revises twice in the same cycle without approval, stop before execution and recommend resolving the disagreement in /meeting.
• Once a plan is approved, use /auto failure handling for workers, critic, curator, durable reviews, guide-writer, the close-session transaction, commit, and push.