Skip to main content
workflow / six steps

corpus

Structured agent work, checked at every step.

Define the work, run agents, verify outputs, preserve evidence.

loop preview

The loop at a glance.

ready
  1. 01Pull
  2. 02Spec
  3. 03Task
  4. 04Run
  5. 05Review
  6. 06Close
pwrcheckevidence
corpus status

# current run

$ corpus review TASK-auth-refresh

PASS AC-001 — output pasted

UNVERIFIED AC-002 — manual resize pending

HUMAN ATTENTION — retry.ts changed outside scope

REVIEW-auth-refresh

review packet / example

review

AC-001 pass

Evidence: `npm test auth-refresh` pasted, exit 0.

AC-002 unverified

Evidence missing. Human attention required before merge.

Don't mark it done without evidence.

common failure modes

The hard part is checking the work.

Agents can produce a plausible diff faster than a team can verify it. corpus makes the handoff, scope, evidence, and memory visible before the change ships.

INTAKE

Vague tickets

A ticket enters as prose. corpus preserves it, then turns it into requirements.

SCOPE

Agent drift

A task packet names what to change, what not to touch, and how to verify.

EVIDENCE

Unbacked completion

A Pass needs pasted output, a CI link, or a named manual observation.

LEDGER

Lost findings

Lessons move from a run summary into findings so the next task can load them.

workflow / six steps

The six points are the loop.

Pull, Spec, Task, Run, Review, Close. Each point creates or checks an artifact the next point can read.

010203040506

workflow / six steps

Six steps, one loop.

The mark maps to the workflow: Pull, Spec, Task, Run, Review, Close. Each pass leaves a file the next step can use.

  1. 01
    Pull

    Capture the ticket and intent in an intake file.

  2. 02
    Spec

    Write requirements, each with a verification method.

  3. 03
    Task

    Hand the agent a bounded packet: scope, do-not-change, verify.

  4. 04
    Run

    The agent implements and pastes real evidence.

  5. 05
    Review

    Check evidence per requirement; human attention where needed.

  6. 06
    Close

    Merge, save findings, update the board.

Close feeds the next Pull — the loop closes

review packet / example

Keep evidence next to the claim.

Specs and reviews are working records. The paper treatment shows up where commands, notes, and evidence live.

SPEC-auth-refresh

spec example / acceptance criterion

spec

AC-003 — Expired refresh token redirects to login

The client must clear local session state and route the user to `/login`.

Verify with: auth-refresh-expired.test

what you get

A few pieces that help.

spec

Spec-first

Write the contract once. Give the agent a task packet, not a drifting chat thread.

review

Review by exception

The review packet shows evidence per requirement and routes the exceptions to a human.

branch

Worktree discipline

One task, one branch, one place to inspect what changed before it joins the main line.

limits

Honesty framework

Every rule says whether it is convention, checklist, toolable, or actually enforced.

plain answers

What corpus does and doesn't do.

Is corpus an agent?

No. Your coding tool writes the code. corpus structures the files around it: specs, tasks, reviews, findings, and the status board.

Does corpus decide whether code ships?

No. Agents can run checks and paste evidence; a human or independent reviewer owns the result.

Do I need the CLI?

No. corpus works as plain markdown. The optional CLI scaffolds, checks, isolates worktrees, and reconciles facts; it does not become a model loop.

Why the seal?

The six points map to the six-step loop: Pull, Spec, Task, Run, Review, Close. It is a diagram before it is a mark.

start / first pass

Start with one spec.

Copy the kit, write the contract, hand the agent a bounded task, and review the evidence. The loop scales because the first pass is small.