research-status

Files (2)

• SKILL.md
• openai.yaml · agents/

SKILL.md (first lines)

# Research Status

> **Not a mainline step.** This reports position; it never advances the project.

Lead with **position**. Answer three questions and nothing else: where am I, what do I have, what is missing.

This skill is **read-only**. It never creates, updates, supersedes, or repairs an artifact, and it never asks the user a decision question. When it finds a defect, it names the defect and the skill that repairs it.

Resolve `.research/config.yaml` per [writeback-protocol.md](../../primitives/research-workspace-contract/references/writeback-protocol.md); apply only its resolution rules. Stop with `blocked: research_workspace_not_initialized` when the config is absent, and recommend `research-setup`.

## Read the current state

Read, at their resolved paths and in this order. All four checks are required before reporting a verified position:

1. the context's single `## Artifact index` table — identity, status, and lineage of every current artifact;
2. the claim registry — counts per declared status;
3. the workpackage index and any claim or receipt state, when the workpackages directory is populated.
4. the current brief, draft, review, or receipt identified by the index, and the direct prerequisites needed for the reported next gate. Inspect their actual status, domain verdict, input identities and acceptance gates. Even when an earlier step is missing from the index, inspect the present current artifacts before reporting their consistency. A missing earlier index row does not waive this check.

Use the index to locate current artifacts; it identifies what is current but does not prove readiness. Limit content reads to the metadata, domain verdict, input identities and acceptance gates needed for this status check. Do not perform a new substantive review. If a pointer is missing, stale, ambiguous, or disagrees with the artifact, report the mismatch and recommend reconciliation; remain read-only. If a referenced artifact cannot be opened, say its content and readiness are unverified rather than repeating the index as established fact.

## Locate the mainline position

Infer the furthest completed task only after matching its index row to the inspected artifact and applicable receipt. The table locates candidates; it does not substitute for their content check:

| Step | Present when the index holds |
|---|---|
| 1 `research-1-grill` | a resolved-idea outcome with linked consequential decisions |
| 2 `research-2-brief` | a current Research Brief |
| 3 `research-3-falsify` | a falsification plan covering the brief's central claims |
| 4 `research-4-tasks` | current workpackages |
| 5 `research-5-run` | at least one sealed workpackage receipt |
| 6 `research-6-review` | a current review report |

A step whose artifact is `partial` or `blocked` is the current position, not a completed one. When artifacts exist out of order, report the gap rather than the highest number.

The furthest completed task is not necessarily the next eligible stage. A review can be complete with a `major-revision` or `reject` verdict; then report its repair gate before recommending a deliverable. A pilot plan does not prove execution, and one sealed package does not prove all packages completed. Prefer the earliest unresolved load-bearing prerequisite and cite the inspected verdict or receipt.

## Report

Give the user, in this order and in their language:

1. **Position** — the current mainline step and whether it is complete, partial, or blocked.
2. **What exists** — one line per current artifact: kind, ID, status, updated. Name superseded artifacts only when a stale consumer still points at one.
3. **What is missing** — unresolved claims by status, blocked gates, workpackages that are ready but unclaimed, and any artifact whose status disagrees with an artifact that depends on it.
4. **Next action** — exactly one recommended skill invocation, with the reason it is the binding constraint.

Report `blocked` and say registration was not read when the config or the context cannot be read. Never infer progress from conversation history, filenames, effort, or an index assertion alone. Cite the inspected artifact verdict or receipt for each readiness judgment; distinguish index state, actual artifact state, and any unverified dependency.

From yhlai0911/research-os, MIT.

More from yhlai0911/research-os