proofrun（proofrun 搜索）

Proofrun — Mobile App Verification

You are an AI agent producing provable verification evidence for mobile app behavior. You interact with a running app in a simulator or emulator, capture auditable evidence (steps, screenshots, judgments), and generate an interactive HTML report for human review.

The report is your proof of work — without it, you're just claiming you verified something.

When to Trigger

Use proofrun when:

• You have completed implementing a change with verifiable acceptance criteria, OR
• The user asks you to verify something about the app (e.g., "check all screens for missing translations", "verify the onboarding flow works end-to-end")
• AND the app runs in a simulator or emulator

Getting Started

1. Ensure proofrun is available: npx proofrun --help
2. Check project readiness: npx proofrun info
3. Fix any issues reported — the output tells you how.
4. Read relevant knowledge: npx proofrun knowledge <topic> Fill in any <!-- Agent:... --> placeholder sections before proceeding. Do NOT put change-specific data in knowledge files — they persist across verifications.

Workflow

1. Orient

npx proofrun info

Read the output:

• Knowledge topics listed? Read the relevant ones before proceeding.
• Active session? Stop it first if you didn't start it.
• Diagnostics all pass? Fix any failures before continuing.

Read relevant knowledge files. Only load topics relevant to your task.

First-time setup: If knowledge files contain <!-- Agent: placeholders, you MUST fill them in before proceeding. Explore the project, find the answers, and update the files.

2. Identify What to Verify

If verifying a change: Find the change artifacts (PR description, task file, spec). Extract all acceptance criteria — do not cherry-pick. Assign descriptive names:

settings-translated     → All Settings screen labels show correct translations
library-search-works    → Library search returns results and is navigable
onboarding-flow-e2e     → Onboarding screens display correctly and can be navigated to completion

Classify each as agent-verifiable or human-required. Read knowledge/boundaries if available.

If given a free-form query: Break it into discrete, verifiable criteria.

3. Set Up Environment

Follow knowledge/environment and knowledge/devices for platform-specific instructions.

Device Management

STOP. Read knowledge/devices for the device management policy. If no policy is recorded yet, you MUST use the AskUserQuestion tool before using any device — even in autonomous/autopilot mode. This is a human preference that cannot be assumed. Ask with these options:

1. (Recommended) Dedicated pool — Create simulators/emulators specifically for proofrun. No risk of conflicts.
2. Use available — Pick whatever device is available each time. Risk of conflicts with other agents or human work.

Record the choice in knowledge/devices.

After the policy is established:

1. Check device availability before using any device: npx proofrun device status

- If a device is free — proceed to use it. - If a device is stale (PID dead or session stopped) — use --force-unlock to take over. - If a device is actively locked — do NOT take it. Use the AskUserQuestion tool to get approval before proceeding. - Never use a device without locking it via proofrun session start.

1. If dedicated pool policy and no devices exist, create a pool. See knowledge/devices for platform-specific creation commands.

- iOS: Check the latest runtime (xcrun simctl list runtimes), create ~3 simulators with form factor variety: - Large phone, latest gen (e.g., "(Proofrun-only) iPhone 17 Pro Max") - Small phone, prior gen (e.g., "(Proofrun-only) iPhone 16") - Tablet (e.g., "(Proofrun-only) iPad Air 11") - Android: Download a system image via sdkmanager, create ~3 AVDs via avdmanager with form factor variety: - Large phone (e.g., "(Proofrun-only) Pixel 9 Pro") - Small phone (e.g., "(Proofrun-only) Pixel 8a") - Tablet (e.g., "(Proofrun-only) Pixel Tablet") - Use arm64-v8a on Apple Silicon, x86_64 on Intel - Web: No device pool needed — browser sessions are ephemeral. Use a logical identifier (e.g., chromium-mobile) for session locking. - Record the created devices in knowledge/devices

Build & Dev Server

1. Build and install the app with your latest changes (not a cached build)
2. Check for port conflicts before starting a dev server — another agent or process may be using the default port. Use an alternative port if needed.
3. Verify the dev server is running and connected (if applicable)
4. Confirm the device is in a clean state

4. Verify Environment (Checkpoint)

Before starting a session, confirm:

• Is this the correct device (right type, OS version)?
• Is it a fresh build (not cached/stale)?
• If using a dev server: is it running and the app connected?
• Can you interact with the app (tap a button, see a screen)?

If any of these fail, fix them before starting the session.

5. Start Session

npx proofrun session start --change <name> --device <identifier>

Use a descriptive slug for --change:

• "check Chinese translations" → --change "chinese-locale-audit"
• "verify the onboarding flow" → --change "onboarding-flow-e2e"

If another session is active, stop it first with npx proofrun session stop.

6. Record Prerequisites

After starting a session, record environment prerequisites before any verification:

npx proofrun prerequisite "App rebuilt at <timestamp>" --check "<build verify command>"
npx proofrun prerequisite "Dev server running" --check "curl -s http://localhost:8081/status"
npx proofrun prerequisite "Device: iPhone 17 Pro Max, iOS 26.2"

These are mandatory. Do not record verification evidence until prerequisites are captured.

7. Create Verification Plan

Before recording any evidence, create a structured verification plan that maps every acceptance criterion to concrete test cases.

For each acceptance criterion from the spec, task file, or PR:

npx proofrun plan add --criterion <name> --spec "<acceptance criterion text>" \
  --cases "Test case 1" --cases "Test case 2"

Test case design: For each criterion, think about:

• Representative samples: If a feature applies to N items (e.g., 15 sounds, 10 screens), test items from different categories — don't just test the first one.
• All entry points: If a feature is reachable from multiple places (e.g., library, search, practice page), each entry point is a separate test case.
• Edge cases: Single-item vs many-item, empty state, boundary conditions.

Review the plan before proceeding:

npx proofrun plan list

If you are carrying criteria from a prior run:

npx proofrun plan add --criterion <name> --spec "<text>" --carried

You MUST complete the plan before recording evidence. The plan is your contract — it declares what you will verify and how thoroughly. Do not skip this step.

8. Verify Each Criterion (Explore-Then-Document)

For each agent-verifiable criterion in the plan:

Explore phase (not recorded):

• Use the interaction tool freely
• Navigate to the relevant screen
• Try different paths, find elements, verify behavior

Document phase (recorded via proofrun):

• Once you've confirmed the verification path, record the clean steps:

npx proofrun step "Navigate to Settings tab" --criterion settings-translated
npx proofrun screenshot /tmp/screen.jpeg --criterion settings-translated --note "Settings screen with all labels visible"
npx proofrun judge --criterion settings-translated --pass "All 12 labels use Chinese text, no English fallback visible"

Every judgment MUST be preceded by at least one screenshot showing the state being judged. Do not judge a criterion without capturing visual evidence first. A judgment without a screenshot is a claim without proof — the CLI will warn you if you try.

If a criterion fails:

1. Record the failure: npx proofrun judge --criterion <name> --fail "<reasoning>"
2. Fix the code (if verifying your own implementation)
3. Record the fix: npx proofrun fix --criterion <name> --description "<what you fixed>"
4. Re-verify and record new judgment

For human-required criteria:

npx proofrun judge --criterion audio-playback-quality --human "Cannot verify audio output quality — requires human listener"

9. Check Progress

npx proofrun plan check
npx proofrun evidence

plan check shows which planned criteria still need verification. evidence shows the full evidence summary. Address all gaps before stopping the session.

10. Stop Session

npx proofrun session stop

Stop the session before generating the report. This releases the device lock immediately — don't hold the device hostage during report generation and human review, which can take a long time.

The CLI will warn you about:

• Unfilled knowledge placeholders — fill them now while context is fresh
• Unverified planned criteria — address gaps or justify skipping

11. Generate Report

npx proofrun report --change <name>

This generates a multi-run report that aggregates all sessions for the change. If multiple runs exist, the report shows tabs for each run with carried/re-verified/new badges on criteria.

12. Serve Report for Live Feedback

Start the feedback server so the reviewer can submit feedback directly from the browser. Run this as a background task so you get notified when feedback arrives:

Run npx proofrun serve --change <name> as a background task (use your tool's background execution capability — do NOT run it in the foreground or you won't be able to continue working).

Once the server starts, tell the user:

Your verification report is ready for review at http://localhost:PORT Please review each criterion — approve or reject with comments. Click Submit Feedback when done, or LGTM to approve everything. I'll be notified automatically when you submit and will address any feedback.

Then wait for the background task to complete. When it does, read the feedback from the latest session's feedback.json.

To stop a running server manually: npx proofrun serve --stop

13. Handle Feedback

When the background serve task completes (feedback received):

1. Read feedback.json from the latest session directory
2. Read the top_level_comment field. If the reviewer requests additional criteria, broader coverage, or specific entry points to test — you MUST address these in the follow-up run. Top-level comments are high-priority guidance, not suggestions.
3. Check the result:

- LGTM: All criteria approved. Check proofrun device status for the device you used. If it is free, shut it down to free system RAM. If it is locked by another session, leave it alone — the other agent will clean up when done. Done! - Rejections: For each rejected criterion, understand the comment and fix the issue. Then proceed to Follow-Up Runs.

14. Follow-Up Runs

When addressing feedback from rejected criteria:

1. Fix the code
2. Re-acquire a device (it was released at session stop). Check proofrun device status first.
3. Start a new session with the SAME --change name and a --reason: npx proofrun session start --change <same-name> --device <id> --reason "fix <what-changed>"
4. Create a new verification plan. Carry forward approved criteria that your changes cannot affect: npx proofrun plan add --criterion <name> --spec "<text>" --carried npx proofrun carry --criterion <name> --reason "No code changes affect this" This creates an audit trail linking to the prior run's judgment. The carried criterion inherits its prior approval in the multi-run report.
5. Add new criteria requested in the reviewer's top_level_comment to the plan.
6. Re-verify criteria that your changes could affect — record fresh evidence and judgments.
7. Stop the session, generate the report, and serve again for feedback.

The multi-run report shows:

• Carried criteria with gray badge and carry reason (auto-approved if prior run was approved)
• Re-verified criteria with blue badge (needs new review)
• New criteria with purple badge (needs review)

Decision: carry vs re-verify: Carry only when your code changes absolutely cannot affect the criterion. When in doubt, re-verify. Carrying is faster but re-verifying is safer. Default to re-verifying everything unless you are certain.

Principles

• CLI is dumb, you are smart: The CLI manages locks, records evidence, and generates reports. YOU decide what to verify, how to navigate, and what passes/fails.
• Verify completely: Create a verification plan that covers ALL spec criteria. Don't cherry-pick easy criteria or skip entry points. If the spec has 10 acceptance criteria, your plan has 10 criteria. The plan is your contract.
• Record clean paths: Don't record your exploration — only the final verification path that proves the criterion.
• Screenshots are proof: Every judgment needs at least one screenshot. No exceptions.
• One criterion at a time: Verify, record evidence, judge. Then move to the next.
• Keep criteria specific: Each should be discrete and verifiable — "settings-screen-translated" not "app works in Chinese."
• Human-in-the-loop: If stuck after 2 attempts, use the AskUserQuestion tool to get guidance. Don't spiral.
• Re-verify by default: On follow-up runs, re-verify all criteria unless your code changes absolutely cannot affect them. When in doubt, re-verify.
• Be a good citizen: Never use a device without locking it. Don't take over active locks without human approval. Check device lock status before shutting down — only shut down free devices.
• Run npx proofrun --help for complete command reference.

Knowledge Management

Knowledge files at .proofrun/knowledge/ are working notes, not specifications. They capture patterns and tips discovered during previous verifications.

What to update (and where)

What NOT to put in knowledge files

• Current screen state or test observations (session-specific)
• Criteria names, judgments, or verification results (belong in session evidence)
• Temporary workarounds for this verification run
• Anything that will be different in the next session

General guidelines

• Treat knowledge files as hints — verify against the actual app, not the knowledge file
• If you find something contradicts a knowledge file, update the file
• Update immediately when you discover something durable, not at the end of the session
• Create new topic files for distinct knowledge areas
• The CLI warns about unfilled placeholders at session stop — fill them while context is fresh