Prax Shift

Prax Shift for Claude

Use Shift as one Prax operating mode. Prax owns durable state, verification, approval boundaries, scheduling, and evidence; Claude is a replaceable worker.

Map the user's intent

• For start or handoff, require a bounded goal, verifier, workspace, and budget.
• For test-only handoff, use the one-shot verification path below; planning or integration work alone does not authorize test execution.
• For status, use prax shift status --run-id <id> to inspect saved evidence in .prax/harness-runs/<id>/, without starting work or claiming fresh verification.
• For continuing coding work, resume the frozen contract with the same run ID. One-shot verification cannot resume.
• For pause or stop, persist a pause request and preserve evidence. Remove a schedule only when explicitly requested.
• Create a recurring schedule only when explicitly requested.

Operating contract

• Keep intent, irreversible decisions, production actions, and scope changes HITL.
• Execute AFK only inside the requested workspace and budget.
• Require an independent verifier for coding work. Without one, report the missing acceptance criterion and do not start a coding worker.
• Keep Claude at dontAsk: operations that still require approval are denied. This is not an OS sandbox. Do not bypass permissions to keep AFK work going; missing narrowly preauthorized capabilities require a human decision.
• Never auto-push, auto-merge, deploy, spend money, or widen permissions.
• Treat HARNESS-ERROR as an environment failure, not a code defect.

Coding: prepare and run

Prepare the files for the user after inspecting existing configuration. Do not overwrite another task or unrelated work. In .prax/harness.json, use schemaVersion: 1, worker: "claude", and relevant existing policy_files (or []); optional model is frozen too. Preserve an existing worker choice unless the user asks to change it.

Create .prax/task.json with schemaVersion: 1, a unique id, the agreed goal, and kind: "code-check". Declare repositories with workspace-relative path, the actual current branch, and narrow relative scope arrays. Declare verifier with trusted argv, workspace-relative cwd, and timeout_seconds; use explicit pytest test paths. Optional control_files binds additional existing scripts/fixtures using workspace-relative paths. Keep acceptance outside editable scope. Set budget with max_iterations, wall_seconds, stuck_after; default conservatively to 3, 600, 2 when unspecified and disclose these limits.

Use prax or the active environment's python -m prax. Plan validates without starting tests or a worker:

prax shift plan --task .prax/task.json
prax shift run --task .prax/task.json

Continue it:

prax shift status --run-id <id>
prax shift pause --run-id <id>
prax shift resume --run-id <id>

Pause requests supervisor shutdown; inspect status to confirm it stopped. Resume checks frozen controls and source identity and retains the total budget. Do not delete evidence, change acceptance, or rotate task IDs to bypass a pause, identity mismatch, or exhausted budget. Escalate for a new task if needed.

Schedule a recurring Shift:

prax cron add \
  --name "<job-name>" \
  --schedule "<five-field cron>" \
  --run-mode shift \
  --task .prax/task.json
prax cron install

The scheduler replays prax shift run --task .prax/task.json --resume-existing; it cannot unpause or renew a budget. Do not create a schedule unless explicitly requested. Follow host scheduling instructions and avoid duplicate schedulers. Legacy goal/feature commands do not provide this guarded task contract.

One-shot verification

For existing tests without code repair, use kind: "verification", empty repository scope arrays, a wall_seconds budget, and verifier.report_schema: "prax-verification-v1". Require a reviewed project adapter and read its documentation for argv, authorization and evidence. Do not invent another test framework or bypass the existing public E2E entry point.

Validate with prax shift plan --task <task.json>. Only after current explicit authorization run prax shift run --task <task.json> --authorize-verification. Obtain the project's exact platform, case or explicit case set, environment, artifacts and side-effect authorization before E2E; never reuse old permissions.

No coding worker is constructed. Failure or interruption does not permit a retry, resume, --resume-existing, recurring schedule, or automatic task-ID rotation. Another attempt needs fresh authorization. Use structured results and sealed evidence to distinguish passed, failed, blocked and not-executed tests; a process exit code alone proves nothing, and missing coverage mapping means unknown.

Always hand back the run ID, stop reason, verifier verdict, changed files, waiting-on-human decisions, and an exact continuation command when supported. For verification, hand back status/evidence and any new authorization needed, never a resume command.

For an integration-only check, do not modify files. Return PRAX_CLAUDE_SHIFT_OK with the resolved Prax executable path.