Research Pipeline Runner Agent Skill

Research Pipeline Runner

The runner turns one requested outcome into an auditable Workspace. Its leading

word is continuation: every action must leave enough durable state for the

next command or agent to continue without reconstructing the Run from chat.

Inputs

• The user Goal, including the target deliverable and constraints.
• An existing Workspace, or a path under workspaces/<name>/ for a new one.
• The selected Pipeline when already recorded in PIPELINE.lock.md.

Outputs

• A Workspace containing GOAL.md, PIPELINE.lock.md, UNITS.csv,

STATUS.md, CHECKPOINTS.md, and DECISIONS.md.

• Pipeline-declared intermediate and target Artifacts.
• Run Evidence under .harness/ and output/.

Steps

1. Bind the Goal to one Workflow

Read the Goal and existing Workspace state. When no valid Pipeline lock exists,

use pipeline-router; load docs/PIPELINE_TAXONOMY.md only for that routing

branch. Preserve an existing valid lock.

Completion criterion: PIPELINE.lock.md names exactly one Pipeline and its

Unit template, and UNITS.csv belongs to that contract.

2. Reconcile before execution

Run the Workspace Doctor or the equivalent Harness inspection. Resolve

recoverable projections first. Route integrity errors to an explicit repair;

retain all earlier Attempts and Decisions.

Completion criterion: the Workspace has either one eligible next Unit or one

specific blocking condition with a recorded next action.

3. Execute one Unit

For the next eligible Unit:

• read its owner, dependencies, inputs, outputs, acceptance rule, and Skill;
• open the Attempt through the Pipeline adapter;
• follow that Skill's SKILL.md and selectively load its context pointers;
• verify every required output and acceptance condition;
• commit Completion through the Pipeline adapter.

Completion criterion: the Unit is committed as DONE with matching Attempt,

Manifest, and Artifact evidence, or it is BLOCKED with a diagnosable Failure.

4. Respect the stop line

The stop line is the first HUMAN checkpoint, unresolved Decision, failed

quality contract, missing required Artifact, or terminal execution error.

Summarize what exists, name the evidence to inspect, and leave one concrete

resume command or question.

Completion criterion: the Run never advances beyond a stop line, and another

operator can identify the blocker from Workspace files without chat history.

5. Continue to the declared outcome

Repeat reconciliation and single-Unit execution while an eligible Unit exists.

After the final Unit, run Audit and Artifact Pack generation.

Completion criterion: target Artifacts exist, required Evaluations pass, Audit

contains no blocking issue, and the Artifact Pack indexes the delivery and Run

Evidence.

Branch Pointers

• Workflow selection: read docs/PIPELINE_TAXONOMY.md, then the chosen

pipelines/*.pipeline.md. The taxonomy owns the routing catalog; this Skill

does not duplicate it.

• Survey strict mode: read the locked Survey Pipeline's quality contract and

the active delivery profile before writing. Branch-specific markers and gates

live there.

• Offline retrieval: use the selected retrieval Skill's documented import

path and preserve source provenance.

• Manual Unit execution: use unit-executor while retaining the same

Completion Protocol.

Operator Commands

Use the outcome-first CLI for normal operation:

uv run rh goal create --topic "<topic>" --workflow <workflow> --workspace workspaces/<name>
uv run rh run start --workspace workspaces/<name>
uv run rh run resume --workspace workspaces/<name>
uv run rh run status --workspace workspaces/<name>
uv run rh evidence inspect --workspace workspaces/<name>
uv run rh improve diagnose --workspace workspaces/<name>

Use scripts/pipeline.py for maintainer debugging, explicit approvals, and

manual status transitions. The Pipeline adapter remains the only supported path

for changing Unit state.