Preview 0.33

U01: Direct a bounded autonomous sprint

Know before you begin

A sprint helps when you can describe the outcome and let the agent manage the steps. You still own product choices and publication decisions. Complete P08 and keep the terminal and harness at the same fork. The website remains the primary lesson surface.

Confirm the fork-first lesson boundary

You · Browser · All operating systems

Open the reviewed immutable Preview 0.33 U01 boundary manifest before using any command. U01 is public only in your personal fork with origin as the fork, upstream fetch-only, and origin as the effective push remote.

Why

The scenario and sprint can create commits, a fork branch, and a pull request; keeping upstream non-pushable is the exercise's safety boundary.

Reviewed resources for Confirm the fork-first lesson boundary

Expected result

You are in a clean forked checkout that completed P08 and can safely create a U01 attempt.

If that does not happen

If origin can push to arbiterForge or the topology is incomplete, stop and return to P08. Do not rewrite remotes or bypass the safety check for this lesson.

Evidence

The public Preview 0.33 manifest lists U01, while the Prepare preflight independently validates the live fork topology.

What you will prove

Deliver docs/academy-sprint-summary.md explaining the Approval boundary, SMARTS decision trail, and Hard-gate stop.

Teach the current policy: the original request authorizes agreed work. The agent may maintain the spec and plan and choose ordinary in-scope methods. It asks about consequential unresolved choices and respects explicit stops. Merge, publication, destructive cleanup, and overrides retain their confirmations.

This request includes a governed commit and a PR in your fork after verification. It never authorizes merging.

Prepare safely

Prepare a fork-safe U01 attempt

Run one installed Academy Prepare command from the clean forked repository root. It creates a numbered academy/U01-autonomous-sprint attempt with the reviewed scenario and brief.

Why

Prepare installs only reviewed scenario bytes after validating the fork/upstream push boundary, so the sprint begins from an inspectable base.

Academy · Native terminal · Windows

$academy = "$env:LOCALAPPDATA\ArbiterAcademy\preview-0.33\Scripts\arbiter-academy.exe"
& $academy --repository (Get-Location).Path prepare U01-autonomous-sprint

Academy · Native terminal · macOS

academy="${XDG_DATA_HOME:-$HOME/.local/share}/arbiter-academy/preview-0.33/bin/arbiter-academy"
"$academy" --repository "$PWD" prepare U01-autonomous-sprint

Academy · Native terminal · Linux

academy="${XDG_DATA_HOME:-$HOME/.local/share}/arbiter-academy/preview-0.33/bin/arbiter-academy"
"$academy" --repository "$PWD" prepare U01-autonomous-sprint

Expected result

Academy reports the numbered attempt branch and preparation commit; training_scenarios/U01-autonomous-sprint.json and U01-sprint-brief.json are committed on that branch.

If that does not happen

If remote validation or preparation fails, preserve the error and return to P08. Do not copy scenario files or create an academy attempt branch yourself.

Evidence

The Academy output names academy/U01-autonomous-sprint/N and its preparation commit; git status --short is empty afterward.

Inspect the prepared scenario

After Prepare reports the numbered attempt, use one command surface to inspect the installed scenario file. A native terminal command has no !. A harness shell command begins with exactly one !.

Why

The brief defines the intended operator guide and its limits before autonomous work starts.

You · Native terminal · Windows

Get-Content training_scenarios/U01-autonomous-sprint.json

You · Native terminal · macOS

cat training_scenarios/U01-autonomous-sprint.json

You · Native terminal · Linux

cat training_scenarios/U01-autonomous-sprint.json

You · Codex harness · Windows

! Get-Content training_scenarios/U01-autonomous-sprint.json

You · Claude Code harness · All operating systems

! cat training_scenarios/U01-autonomous-sprint.json

Expected result

The prepared attempt shows autonomous_sprint, academy-sprint, new-native-artifacts, and docs/academy-sprint-summary.md.

If that does not happen

If the file is absent, do not create it by hand. Confirm that Prepare succeeded on the U01 attempt branch, then use Reset only after preserving the failed attempt.

Evidence

Retain the scenario file, U01-sprint-brief.json, and the numbered attempt branch name.

Prepare leaves the sprint unplanned. Arbiter creates the native specification and plan, then keeps them current as it works. Open the rendered pages to inspect scope, tasks, and real progress. If the writer fails, preserve its diagnostic and repair the installed plugin before continuing.

Practice

Use the host-native sprint command on the prepared attempt

Request the complete bounded sprint. Your original request supplies its scope and needs no second spec or plan approval. The agent asks only when an actual consequential choice is missing.

Why

Your original request authorizes the sprint. The native specification and plan keep its scope and checks visible as work proceeds. This exercise authorizes a fork PR; merging or publishing requires a separate request.

Your agent · Claude Code harness · All operating systems

/ca:sprint academy-sprint: Write docs/academy-sprint-summary.md from the prepared U01 brief. Explain authorization from the original request, meaningful SMARTS decisions, and hard-gate stops. Create and maintain the native spec and plan, record material choices if any, verify the guide, and commit through the gate. Record the guide as one work item bound to the specification and plan, and leave it queued. Open a PR only in my fork against its main branch. Do not change product code, tests, dependencies, or remotes. Do not merge or publish.

Your agent · Codex harness · All operating systems

$ca-sprint academy-sprint: Write docs/academy-sprint-summary.md from the prepared U01 brief. Explain authorization from the original request, meaningful SMARTS decisions, and hard-gate stops. Create and maintain the native spec and plan, record material choices if any, verify the guide, and commit through the gate. Record the guide as one work item bound to the specification and plan, and leave it queued. Open a PR only in my fork against its main branch. Do not change product code, tests, dependencies, or remotes. Do not merge or publish.

Your agent · Pi harness · All operating systems

/ca-sprint academy-sprint: Write docs/academy-sprint-summary.md from the prepared U01 brief. Explain authorization from the original request, meaningful SMARTS decisions, and hard-gate stops. Create and maintain the native spec and plan, record material choices if any, verify the guide, and commit through the gate. Record the guide as one work item bound to the specification and plan, and leave it queued. Open a PR only in my fork against its main branch. Do not change product code, tests, dependencies, or remotes. Do not merge or publish.

Your agent · Pi harness · All operating systems

/skill:ca-sprint academy-sprint: Write docs/academy-sprint-summary.md from the prepared U01 brief. Explain authorization from the original request, meaningful SMARTS decisions, and hard-gate stops. Create and maintain the native spec and plan, record material choices if any, verify the guide, and commit through the gate. Record the guide as one work item bound to the specification and plan, and leave it queued. Open a PR only in my fork against its main branch. Do not change product code, tests, dependencies, or remotes. Do not merge or publish.

Expected result

A native specification and plan, the operator guide, observed verification, and a fork-only PR handoff. Real missing choices remain visible.

If that does not happen

If the installed writer fails, preserve its diagnostic and repair the installation. If scope or a consequential choice is unclear, answer that specific question before affected work continues.

Evidence

Preserve the saved specification, plan, operator guide, and any material decisions recorded in the sprint log. Check validates repository facts, not host invocation, approval, or pull-request creation.

From codeArbiter 3.8, a sprint records each deliverable as one work item. On this practice board that is one queued row bound to the sprint's specification and plan, and it rides the sprint commit.

Inspect the specification

Open the rendered specification and check whether it reflects your request. No further approval reply is required for in-scope work. Correct a real misunderstanding when you see it.

Why

The learner, not the autonomous process, owns the scope decision that permits later execution.

You · Claude Code harness · All operating systems

Show the saved academy-sprint specification. Explain how it serves the requested operator guide and identify any real unresolved choices. Continue authorized work; do not add another approval ceremony.

You · Codex harness · All operating systems

Show the saved academy-sprint specification. Explain how it serves the requested operator guide and identify any real unresolved choices. Continue authorized work; do not add another approval ceremony.

You · Pi harness · All operating systems

Show the saved academy-sprint specification. Explain how it serves the requested operator guide and identify any real unresolved choices. Continue authorized work; do not add another approval ceremony.

Expected result

A readable specification that reflects the original request and the actual state of the work.

If that does not happen

Correct the concrete misunderstanding. Preserve existing work and resume from its actual state.

Evidence

A progress label is not a test result or independent acceptance.

Inspect the plan

Open the rendered plan and check whether it reflects your request. No further approval reply is required for in-scope work. Correct a real misunderstanding when you see it.

Why

Reviewing the plan helps you catch scope misunderstandings and missing checks. Work already covered by your request can continue without another approval reply.

You · Claude Code harness · All operating systems

Show the saved academy-sprint plan. Explain how it serves the requested operator guide and identify any real unresolved choices. Continue authorized work; do not add another approval ceremony.

You · Codex harness · All operating systems

Show the saved academy-sprint plan. Explain how it serves the requested operator guide and identify any real unresolved choices. Continue authorized work; do not add another approval ceremony.

You · Pi harness · All operating systems

Show the saved academy-sprint plan. Explain how it serves the requested operator guide and identify any real unresolved choices. Continue authorized work; do not add another approval ceremony.

Expected result

A readable plan that reflects the original request and the actual state of the work.

If that does not happen

Correct the concrete misunderstanding. Preserve existing work and resume from its actual state.

Evidence

A progress label is not a test result or independent acceptance.

The inspection cards help you follow the work. They do not require approval replies. Request a pause explicitly if you want one; otherwise, ordinary work proceeds.

Read the result and verification

Read the finished guide, inspect the diff, and compare reported verification with the actual output.

Why

The verifier-owned evidence boundary is explicit before the lesson becomes publicly runnable.

You · Claude Code harness · All operating systems

Show the operator guide and final diff. List the checks actually run and any failures or unrun checks. Explain any material decision recorded during the sprint. Give me the fork PR link if creation succeeded; otherwise report the remaining handoff.

You · Codex harness · All operating systems

Show the operator guide and final diff. List the checks actually run and any failures or unrun checks. Explain any material decision recorded during the sprint. Give me the fork PR link if creation succeeded; otherwise report the remaining handoff.

You · Pi harness · All operating systems

Show the operator guide and final diff. List the checks actually run and any failures or unrun checks. Explain any material decision recorded during the sprint. Give me the fork PR link if creation succeeded; otherwise report the remaining handoff.

Expected result

The guide explains scope, decisions, and genuine stops. The saved plan reflects work completed, with verification reported separately.

If that does not happen

If any artifact is absent or the worktree is not the prepared attempt, stop. Do not create substitute files to imitate evidence.

Evidence

Inspect the actual PR in GitHub; a local Check cannot prove that a PR exists.

Read the guide as if you were about to use it. Remove instructions to approve artifact codes, certify hashes, or reapprove each routine plan update.

Recognize success

One learner commit contains the maintained native spec, plan, and guide, plus a sprint-log append if a material choice needed recording. The prior log prefix is unchanged and the worktree is clean. The sprint opens a pull request in your fork after its commit gate.

Explain material choices. SMARTS helps compare methods; it grants no permission. A reported implementation status does not certify tests, review, or publication.

Check

Check the bounded local evidence

Ask the agent to run the installed external Academy Check with the host-loaded plugin installation for native artifact inspection.

Why

The installed release accepts only the reviewed sprint-evidence contract and fails closed for altered scope or unsafe remotes.

Your agent · Claude Code harness · All operating systems

Run the installed external Academy verifier for U01-autonomous-sprint with --repository set to this clone and --plugin-root set to your loaded codeArbiter installation. Show the exact command and result. Use the installed verifier outside this checkout. Preserve any failure and report what remains.

Your agent · Codex harness · All operating systems

Run the installed external Academy verifier for U01-autonomous-sprint with --repository set to this clone and --plugin-root set to your loaded codeArbiter installation. Show the exact command and result. Use the installed verifier outside this checkout. Preserve any failure and report what remains.

Your agent · Pi harness · All operating systems

Run the installed external Academy verifier for U01-autonomous-sprint with --repository set to this clone and --plugin-root set to your loaded codeArbiter installation. Show the exact command and result. Use the installed verifier outside this checkout. Preserve any failure and report what remains.

Expected result

Academy Check reports pass only when the prepared brief, exact artifact scope, safe remote topology, one learner commit, and clean worktree all match; otherwise it reports a nonzero failure.

If that does not happen

Do not edit progress or release data to change the result. Inspect the reported repository boundary, preserve the attempt, then Reset for a numbered retry if needed.

Evidence

Check validates the native pair, guide, preserved audit history, commit boundary, and live fork configuration. It does not authenticate approval, test execution, or a remote PR.

Check uses the installed native engine to validate the committed spec and plan, then checks the guide, commit boundary, and preserved log history. It does not authenticate approval. It does not prove that a pull request was created or that a person understood the guide. Inspect the real PR separately.

Recover or continue

Hint 1. State the outcome and limits once. Correct scope drift without inventing new artifact approval ceremonies.

Hint 2. Keep the guide, spec, plan, and append-only decision record consistent.

Hint 3. Preserve a failed attempt. A non-destructive numbered retry keeps the earlier work available.

Return to main only after the attempt is clean

On a clean checked attempt, return to main from a native terminal. This is a direct shell command, so it has no !.

Why

Leaving evidence intact keeps an incomplete training run inspectable without treating it as public completion.

You · Native terminal · Windows

$status = git status --short
if ($status) {
  Write-Error "The U01 attempt is dirty. Inspect it before switching branches."
  exit 1
}
git switch main

You · Native terminal · macOS

status="$(git status --short)"
if [ -n "$status" ]; then
  printf '%s\n' "The U01 attempt is dirty. Inspect it before switching branches." >&2
  exit 1
fi
git switch main

You · Native terminal · Linux

status="$(git status --short)"
if [ -n "$status" ]; then
  printf '%s\n' "The U01 attempt is dirty. Inspect it before switching branches." >&2
  exit 1
fi
git switch main

Expected result

git status --short prints nothing before main becomes the active branch.

If that does not happen

If status shows changes, stop and inspect them. Do not discard, amend, rebase, force-update, or delete the attempt branch to make it look clean.

Evidence

The numbered attempt remains reachable for later review.

Reset safely to a numbered retry

After returning to main with the failed attempt preserved, run one installed Academy Reset command. Reset archives no history destructively and creates the next numbered U01 attempt from the reviewed base.

Why

A retry must preserve the prior evidence while creating an independent, inspectable scenario branch.

Academy · Native terminal · Windows

$academy = "$env:LOCALAPPDATA\ArbiterAcademy\preview-0.33\Scripts\arbiter-academy.exe"
& $academy --repository (Get-Location).Path reset U01-autonomous-sprint

Academy · Native terminal · macOS

academy="${XDG_DATA_HOME:-$HOME/.local/share}/arbiter-academy/preview-0.33/bin/arbiter-academy"
"$academy" --repository "$PWD" reset U01-autonomous-sprint

Academy · Native terminal · Linux

academy="${XDG_DATA_HOME:-$HOME/.local/share}/arbiter-academy/preview-0.33/bin/arbiter-academy"
"$academy" --repository "$PWD" reset U01-autonomous-sprint

Expected result

Academy reports academy/U01-autonomous-sprint/N+1 and a new preparation commit; the earlier attempt branch remains reachable.

If that does not happen

Do not simulate Reset with destructive Git commands. If Reset refuses because the current attempt is dirty or still active, inspect and preserve it before retrying.

Evidence

Both numbered attempt refs are reachable and git status --short is empty on the new attempt.

Continue to U02 after a passing Check.

Understand the mechanism

A sprint maintains a working agreement while carrying work forward. The agent checks facts, handles ordinary implementation problems, and reports real uncertainty. Run an Autonomous Sprint covers boundaries and pauses.

Use this in your own project. Give a small outcome, observable acceptance criteria, constraints, external-service limits, and any human checkpoint you want. Request a report of changes, actual checks, unresolved risks, and remaining delivery steps.

Next step

Continue with U02-override-audit-metrics.

Recovery guidance