Debug App

📋 Shared instructions: shared-instructions.md — read first.

Debug App — Monitor & Fix

Monitor the running app by reading the Metro dev-server terminal output, detect runtime and bundle errors, and fix them autonomously by editing the affected files (or routing to the right skill when the fix belongs in a domain like Dataverse schema or auth registration). For silent failures, inject temporary console.log statements at data-path boundaries, read the Metro terminal for output, then clean them up after the root cause is fixed. Modeled on the upstream app-debugger.agent.md pattern — foreground loop, 5-second cadence, exit on 3 consecutive clean polls.

Dev-client limitation: the standalone dev client outputs app/runtime logs, React errors, and Metro bundler output to the terminal running npm run dev. This includes host runtime diagnostics that use strings such as [AuthProvider] MSAL init failed:, [bridge] fetch THREW for, [bridge] HTTP <status> for, [addAadAppToConnectionAcl] failed HTTP <status> for connection, [useConnectionRefs] could not verify connection ACLs; treating existing connections as setup-required, and [PAHost][ErrorBoundary] Unhandled JS error:. There is no separate device log stream. All diagnosis happens by reading that terminal and, where needed, injecting strategic trace statements into source files.

Subcommands (parsed from $ARGUMENTS)

Dispatch rule: if $ARGUMENTS is non-empty and is not one of the reserved subcommand tokens (status, stop, help, --help, -h, version, --version), treat the entire string (everything after the command name; outer quotes optional) as the symptom and use symptom-driven mode. For help / --help / -h, print the subcommands table above and exit.

Tip — "play around then debug": in primary mode, BashOutput($METRO_TERMINAL_ID) returns Metro output accumulated since the last read. So if something weird just happened, keep using the app the way you would normally — then run /debug-app (or /debug-app "<what you saw>") and the very first cycle will see the entire history of your session, not just what arrives after the skill starts. No need to reproduce the bug under the agent's eye.

Core Principles

• Foreground autonomous loop — Once started, this skill owns the conversation until 3 consecutive clean polls confirm the app is healthy, the user types stop, or the escalation rule trips. Do not run other skills concurrently — they'll queue behind the loop.
• Run AFTER the app is loaded — npm run dev must be running and the simulator/device must have the app open. Phase 0 verifies this; the skill stops cleanly if no app is detected.
• Native-only runtime target — The app must be loaded in a native dev client on a device or simulator; Metro terminal output is the log source for that native session.
• No web or direct Metro probes — Do not use React Native Web, browser automation, curl, fetch, WebFetch, or any direct request to a Metro/localhost endpoint for runtime diagnosis. Read only the Metro terminal and source files.
• No screen-by-screen verification — Do not crawl routes or validate every screen. In symptom mode, focus only on the user-reported workflow and the terminal/source evidence needed to diagnose it.
• One fix at a time — Fully resolve one issue (context → fix → type-check → reload → re-poll) before starting the next. No batching.
• Working-dir state — All session state lives in .claude/debug-app/ (gitignored): fixes.md for audit log, unresolved.md for escalations, injected-logs.md for tracking injected console.log statements. Survives across runs.
• Reference resolution order — For unfamiliar errors: in-repo references first (skills/add-dataverse/references/dataverse-reference.md, etc.), then mcp__microsoft-learn__microsoft_docs_search, then general web search.

Workflow — Task List First

Before entering the monitor loop, write a task list and keep it up to date:

- [ ] Verify dev server is running (BashOutput on Metro terminal — expect Metro banner)
- [ ] Capture baseline terminal state (read BashOutput, note most recent activity)
- [ ] (Symptom mode only) Phase 0.5: parse symptom → ask user to navigate → inject console.logs → read terminal → walk data path → clean up logs
- [ ] Monitoring cycle 1: collect → classify → fix if needed
- [ ] Monitoring cycle 2: collect → classify → fix if needed
- [ ] Monitoring cycle 3: collect → classify → fix if needed
      (add cycles as needed; stop after 3 consecutive clean cycles AND symptom resolved/flagged)
- [ ] Fix: <error summary> → <inline edit | skill route>  (one task per error found)

Mark each cycle complete (clean OR fixed) before starting the next.

Phase 0 — Startup Check

Before entering the loop:

0.0 Resolve the Metro terminal

The Metro terminal is the only log source. The dev-player routes all JS output there.

1. Read memory-bank.md for the Metro terminal id: line (written by /create-mobile-app Step 12).
2. If found, call BashOutput against that id once. If it returns any Metro output (even just the banner), set $METRO_TERMINAL_ID and continue.
3. If memory-bank.md has no terminal id, or BashOutput returns "shell not found" / "no such background shell": ask the user:

   "Which terminal is running npm run dev? I need its terminal ID to read Metro logs. If you started it in VS Code, look for the active terminal tab name." Wait for the user to provide the ID, then retry BashOutput against the provided id. Set $METRO_TERMINAL_ID and continue.

Record the resolved id in fixes.md:

[<HH:MM:SS>] Log source — Metro terminal $METRO_TERMINAL_ID

0.1 Ensure state directory

mkdir -p .claude/debug-app
touch .claude/debug-app/fixes.md
touch .claude/debug-app/unresolved.md
touch .claude/debug-app/injected-logs.md
rm -f .claude/debug-app/symptom-state    # per-session — Phase 0.5 rewrites it if symptom mode is active

If fixes.md is empty, write a session header:

# Debug session — <date>

0.2 Verify Metro bundled and the app is running

Telemetry checkpoint: validate_metro_session

Branch on the source resolved in 0.0.

If $METRO_TERMINAL_ID is set (primary path):

Call BashOutput on it once and scan the captured Metro output:

• Most recent error-class line is SyntaxError, Unable to resolve module, transform failed, or error: Bundling failed → bundle is broken. Treat as a Step B "Import / Bundle" critical error and route through Step D immediately. Do NOT enter the steady-state loop until the bundle is healthy.
• Output contains Bundling complete / iOS Bundled / Android Bundled with no later error-class line → Metro is healthy. Proceed.
• Output contains a Metro banner (Metro waiting on, Logs for your project) but no native Bundled / bundling lines yet → Metro is up but no native client has connected. Tell the user:

  Metro is running but no app is connected yet. Open the app on a device or simulator, then re-run /debug-app. Stop here.

• Output is empty, OR contains no Metro banner at all → the recorded shell is alive but Metro isn't running in it (the user repurposed the terminal). Tell the user:

  Metro not detected in the recorded terminal. Restart with npm run dev and re-run /debug-app — the new terminal id will be picked up from memory-bank.md. Stop here.

If $METRO_TERMINAL_ID is NOT set:

Ask the user:

"Which terminal is running npm run dev? Please provide the terminal ID so I can read Metro output."

Wait for the user to reply. Set $METRO_TERMINAL_ID to the provided ID, call BashOutput($METRO_TERMINAL_ID) once, and continue with the checks above.

0.3 Capture baseline

Telemetry checkpoint: capture_runtime_baseline

Read the latest output from BashOutput($METRO_TERMINAL_ID). Note the most recently bundled native platform (iOS / Android) and any recent runtime log lines. Append to fixes.md:

[<HH:MM:SS>] Baseline — last Metro activity: <1-line summary of most recent lines>

0.4 Initialize cursor

BashOutput maintains an internal stream cursor against $METRO_TERMINAL_ID — each call returns only output produced since the previous call. No separate cursor file is needed. The .claude/debug-app/cursor file is no longer used and can be ignored if present from a previous session.

Phase 0.5 — Symptom-driven setup (only when $ARGUMENTS is a symptom string)

Skip this entire phase if no symptom was provided. The standard log-polling loop alone is good at visible errors but blind to silent ones: an empty list because the connector wasn't added, a blank screen because useFocusEffect wasn't wired, blank rows because column names don't match the model. Phase 0.5 closes that gap.

0.5.1 Parse the symptom

Extract three signals from the user's text:

Append to fixes.md:

[<HH:MM:SS>] Symptom — class=<class> screen=<path> entity=<service>

If no screen/entity match: keep screen=unknown / entity=unknown and proceed — Phase 0.5 still injects diagnostic logs and reads the terminal from whatever data path is most likely based on the symptom text.

0.5.2 Ask the user to navigate to the affected screen

The dev-player has no automation API for navigation. Ask the user:

"Please open the <screen> screen on the device/simulator, then reply ready."

Wait for the user to confirm before proceeding.

0.5.3 Inject diagnostic console.log statements and read terminal

Inject targeted console.log statements at the boundaries of the suspected data path so the Metro terminal reveals what's happening.

Injection sites — choose the minimum set that covers the symptom class:

Console.log injection pattern — all injected lines MUST use this exact format:

console.log('[TRACE <tag>]', <value>); // [INJECTED-TRACE]

• <tag> — short unique label for this site (e.g., items, render, service-response)
• // [INJECTED-TRACE] trailing comment on the SAME LINE — this is the cleanup grep key
• Log the smallest useful value; use JSON.stringify(value) for objects
• Never inject inside src/generated/ — inject in the hook/screen that calls into it

Record every injection in .claude/debug-app/injected-logs.md:

[<HH:MM:SS>] Injected [INJECTED-TRACE] at <file>:<line> — tag=<tag>

Then tell the user:

"I've added diagnostic console.log statements. Please reload the app (press r in the Metro terminal), navigate to <screen>, and trigger the symptom (e.g., scroll the list, tap the button). Reply done when finished."

Wait for the user to reply, then call BashOutput($METRO_TERMINAL_ID) and filter for [TRACE lines.

0.5.4 Walk the data path from terminal output

Use the [TRACE lines to walk the chain:

1. Screen TSX (app/<route>.tsx)

   • Find the useListData(...) / use*Data(...) call.
   • Check service-call options — a stray top: 0, an over-strict filter, a search: query bound to a never-cleared input, or orderBy on a missing column can each silently return zero rows.
   • Check any client-side .filter(...) after the data lands.

2. Data hook (src/hooks/useListData.ts or sibling)

   • Critical: the template hook has TWO mock-fallback paths:
     • Error path: service returns { error } → hook substitutes mock AND may call setError. Silent if the screen ignores error.
     • Empty-result path: service returns { data: [] } (no error) → hook silently substitutes mock. Always invisible without a [TRACE] log.
   • Detect: Grep for MOCK_ imports in the screen file. If present, mock data is wired in.
   • Confirm useFocusEffect is used (not useEffect) — useEffect won't re-run on back-navigate.

3. Generated service (src/generated/services/<Name>Service.ts)

   • If a TODO stub or file missing → route to /add-connector or /add-dataverse. Do NOT edit src/generated/.
   • If it exists and the [TRACE service-response] log shows an error field → read that error; 401/403 = auth issue; 404 = wrong resource name.

4. Generated model (src/generated/models/<Name>Model.ts)

   • Confirm field names match what the screen references. item.title vs cr3e9_title produces blank rows.

5. power.config.json

   • Confirm the datasources array contains the suspected entity / connector. If absent, npx power-apps add-data-source was never run for it.

6. Auth state (src/playerConfig.ts, app.config.js, auth.config.json, useAuth() hook)

   • 401 from the service wrapped as { error } — the [TRACE service-response] log surfaces the error string.
   • OAuth deeplink handoff: verify app.config.js → expo.scheme matches src/playerConfig.ts → connectorOAuthRedirectUri, AND the same redirect URI is in auth.config.json and the Entra ID registration. If the app registration is missing, route the user to the Power Apps Wrap page via /set-app-registration-native.

Classify the [TRACE output:

Record the outcome in .claude/debug-app/symptom-state (single line: resolved, flagged, or pending).

• Fix is clear and local → apply via Step D3 + D4 (type-check + reload + re-poll). After the fix, ask the user to interact with the screen again and read the terminal. If the [TRACE items] line shows N > 0, write resolved.
• Fix routes to another skill → tell the user, log to unresolved.md. Write flagged.
• No obvious cause → log a structured note to unresolved.md. Write pending and enter the monitor loop.

0.5.5 Clean up injected console.log statements

After the root cause is identified and a fix is applied (or Phase 0.5 concludes), remove ALL injected logs:

grep -rn 'INJECTED-TRACE' app/ src/hooks/ src/services/

For each matching file, edit out the console.log(...); // [INJECTED-TRACE] lines. Verify with:

grep -rn 'INJECTED-TRACE' app/ src/hooks/ src/services/  # must return zero results

Clear the tracking file:

echo '' > .claude/debug-app/injected-logs.md

Run npm run type-check once after cleanup.

Hard rule: Never leave [INJECTED-TRACE] lines in code. Clean up before marking the session done, even if the symptom is pending or flagged.

0.5.6 Re-enter the standard monitor loop

After Phase 0.5 completes, fall through to the monitor loop (Step A). The "3 consecutive clean cycles" exit condition is suspended until the symptom is either marked resolved or recorded as NEEDS ATTENTION in unresolved.md. After that, the loop exits per the standard rule.

Monitor Loop

Repeat until 3 consecutive clean cycles, OR the user types stop, OR the escalation rule trips.

Step A — Collect logs

Telemetry checkpoint: collect_runtime_logs

BashOutput(bash_id=$METRO_TERMINAL_ID)

BashOutput returns only output produced since the previous call against the same shell — its built-in stream cursor IS the cursor.

In the new output, surface as classifiable signal:

• Runtime ERROR / WARN / LOG prefixes
• Host diagnostic lines with prefixes like [PAHost], [bridge], [AuthProvider], [AuthContext], [useConnectionRefs], [useConnectionSetup], [addAadAppToConnectionAcl], [PAHost][ErrorBoundary]
• Stack frames (at <fn> (<file>:<line>:<col>))
• Bundle-class errors (Unable to resolve module, SyntaxError, transform failed) — re-classify as Step B "Import / Bundle" Critical and route through Step D
• HTTP method + status from Metro's request log (e.g., "GET /index.bundle?platform=ios&dev=true ..." 500 -) — non-200 on .bundle is a bundle/transform failure; non-2xx on connector / Dataverse hosts feeds Step B "Network / API"
• Lines containing Bundling complete / iOS Bundled / Android Bundled are informational — log to fixes.md at debug volume but do NOT classify as an issue
• [TRACE prefixed lines from injected trace statements — classify under the symptom walk (Phase 0.5.4), not as errors

Interpretation rule for host diagnostic lines:

• Treat as classifiable signal only when lines match emitted host strings such as:
  • [AuthProvider] MSAL init failed:
  • [AuthProvider] Intune enrollment failed:
  • [AuthContext] acquireTokenSilent failed for scopes:
  • [AuthProvider] Intune unenroll failed:
  • [bridge] unhandled plugin call
  • [bridge] fetch THREW for
  • [bridge] HTTP <status> for
  • [addAadAppToConnectionAcl] failed HTTP <status> for connection
  • [addAadAppToConnectionAcl] error:
  • [useConnectionRefs] could not verify connection ACLs; treating existing connections as setup-required
  • [useConnectionRefs] Failed to load connections:
  • [useConnectionSetup] could not grant connection ACL: missing Power Apps token or user OID
  • [PAHost] getConnectorToken: acquireToken threw for apiId="...":
  • [PAHost] getConnectorToken: acquireToken returned null for apiId="..."
  • [PAHost] getDataverseToken: acquireToken threw for orgUrl="...":
  • [PAHost] getDataverseToken: acquireToken returned null for orgUrl="..."
  • [PAHost][ErrorBoundary] Unhandled JS error:
  • [PAHost][ErrorBoundary] Error stack:
  • [PAHost][ErrorBoundary] Component stack:
• Treat as informational when lines are lifecycle/status-only, such as bridge setup/ready, token acquisition start/success, bridge registration, and connection-setup screen visibility.

Step B — Classify each new log entry

Telemetry checkpoint: classify_runtime_failures

Apply the 8-category table. Treat each unique stack trace / error message as one issue.

Parsing the multi-line bundle/transform block (primary mode): Metro prints these as a banner (e.g., error: Bundling failed, iOS Bundling failed) followed by an indented block. Unlike runtime stacks, the file:line is on the first non-banner line of the block, formatted as <absolute or relative path>:<line>:<col>. There are usually no at <fn> stack frames. Example to recognize:

iOS Bundling failed 412ms
SyntaxError: /Users/.../app/(tabs)/todos.tsx: Unexpected token (47:12)
  45 |   return (
  46 |     <YStack>
> 47 |       <Text>{title</Text>
     |             ^
  48 |     </YStack>
  49 |   );

Take app/(tabs)/todos.tsx:47:12 as the fix site. The recipes in D3.1 below operate on this format.

Ignore known-noisy lines:

• Require cycle: warnings from Metro
• VirtualizedList: You have a large list… without an associated crash
• Expo SDK informational banners (Starting Metro…, Connecting to…)
• React Navigation development warnings about non-serializable params (unless tied to a crash)
• USER_CANCELLED from pen input when the screen leaves state unchanged and does not show an error
• Host lifecycle/info lines with no failure indicator, for example [bridge] setupNativeHost: bridge ready, [PAHost] bridge registered, [PAHost] render: waiting for connection resolution (spinner)

Step C — If NO issues found

Telemetry checkpoint: confirm_runtime_health

Increment the consecutive-clean-cycle counter.

Before exiting at 3 clean cycles, check the symptom guard. Read .claude/debug-app/symptom-state and pick the matching exit path below. (If no symptom-driven mode was used this session, Phase 0.1 cleared the file at startup, so the "file missing" branch fires.)

resolved OR file missing (no symptom mode this session):

✓ App is running cleanly — no errors detected across 3 consecutive log checks.
  Symptom verification: <PASS | n/a — no symptom provided>.
  Session summary written to .claude/debug-app/fixes.md.

  To resume monitoring, run /debug-app again.

flagged (Phase 0.5 found a real problem that needs another skill):

⚠ Logs are clean BUT the symptom isn't fixed — it requires another skill.
  Symptom: <class> on <screen>
  Next step: <skill route recorded in unresolved.md by Phase 0.5> (e.g., run /add-connector)
  Details: .claude/debug-app/unresolved.md

  Re-run /debug-app "<symptom>" after taking that step to verify the fix.

Do NOT print the green ✓ — the app is technically log-clean but the user-visible problem persists, and the user needs to act before re-running.

pending (still active after 3 clean log cycles):

Append to unresolved.md with the Phase 0.5 chain findings, then print:

"⚠ Symptom <class> on <screen> still active after 3 clean log cycles. The runtime is quiet but the user-visible problem persists — likely a swallowed data-path error. See .claude/debug-app/unresolved.md for the chain walk. Suggested next step: ."

Exit the loop. Do NOT auto-resume.

Iteration cap (applies in both modes): independent of the clean-cycle counter, the loop exits after 50 total cycles (≈5 min wall clock on a clean loop, longer when fixes are running). Track cycle: <N> at the top of fixes.md and increment per cycle. On cap-hit, exit with:

"⚠ Loop reached the iteration cap (50 cycles). Symptom may be intermittent OR a fix is regressing on every reload. See .claude/debug-app/fixes.md for the per-cycle log. Suggested next step: review the last 3 fixes for circular regressions, or re-run with a more specific symptom."

If counter is < 3 AND the cap hasn't tripped, pause 5 seconds (sleep 5 via Bash; on Windows without bash, Start-Sleep -Seconds 5 via pwsh -NoProfile -Command), then return to Step A.

Step D — If issues ARE found

Telemetry checkpoint: repair_and_verify_runtime_issue

Reset the consecutive-clean counter to 0. For each issue, work through the sequence below one at a time before moving to the next.

D1. Gather context

Read the most recent output from BashOutput($METRO_TERMINAL_ID). Note whether the log shows:

• A crash with a full stack trace (app is crashing)
• A React error boundary message (component threw)
• A network error or HTTP status (API/connector failure)
• Empty response with no error (silent failure — consider injecting a [INJECTED-TRACE] console.log; see Phase 0.5.3 for the pattern)

For JS Runtime / React errors, the stack trace in the terminal IS the context. Read the topmost user-code frame to locate the file.

For crashes or blank screens with no terminal output: ask the user:

"Do you see anything on screen — error boundary, blank white, or loading spinner? Please copy any visible error text."

D2. Root-cause analysis

Read the relevant source file(s). Identify:

• The exact file and line causing the error. For runtime errors (JS Runtime, React, Network/API), pick the topmost user-code frame from the stack trace (skip node_modules/ and src/generated/). For bundle / transform errors (Import / Bundle category, primary mode), use the file:line on the first non-banner line of the Metro error block — see Step B's parsing note.
• Whether the fix touches the data layer, UI layer, routing, or schema.
• Whether the fix could regress other screens.

D3. Apply the fix

For Import / Bundle category errors, jump to D3.1 first — those have specific recipes that pre-empt the generic routing table below. For everything else (JS Runtime, Network/API, React, etc.), use the routing table:

PDF/pen/geolocation-specific routing:

• INVALID_URL for PDF viewer input is an inline screen/wrapper fix. Allow https:// and non-empty file:// inputs with viewer 0.2.9+. Never add support for content://, blob:, or http:// in the native viewer path.
• A generated PDF local file:// URI may be opened by native PDF viewer 0.2.9+.
• NATIVE_MODULE_MISSING for PDF viewer or pen input means the native extension is not in the running build. Do not install packages or edit native config from debug; route to /add-native pdf-viewer or /add-native pen-input to verify wrapper/package state, then tell the user a native rebuild/template update is needed if the package is absent from the app build.
• For geolocation, debug the actual failure dimension: can tracking start (startTracking, permissions, native module), are rows reaching Dataverse (default msdyn_locationrecords exists, native upload/auth errors, no JS upload path), and does behavior match the user expectation (background, restart persistence, breadcrumb/route continuity). Fix visible screen handling inline; if the native module/table is missing, block use and route to the relevant geolocation setup path, not /add-dataverse.
• USER_CANCELLED from pen input is not a bug unless the screen renders it as an error. Inline fix screens that show cancellation as failure.
• Dataverse artifact writes are local app fixes only when the schema/service already exists. If File/Image columns are missing, route to /add-dataverse.

For inline edits, keep the change minimal and surgical. Do not refactor surrounding code, rename symbols, or change component contracts.

Append to .claude/debug-app/fixes.md:

[<HH:MM:SS>] <category> — <file>:<line> — <one-line description of fix>

D3.1 Bundle / transform error fix recipes (Import / Bundle category)

These recipes apply to errors classified as "Import / Bundle" in Step B. They are only visible in the Metro terminal (BashOutput($METRO_TERMINAL_ID)). Each recipe is opinionated: take the action listed if its precondition matches, otherwise fall through to the next.

After applying any inline edit (rows 1, 3, 4 above), Metro auto-detects the file save and re-bundles. Skip directly to D4 — do NOT manually trigger a reload. The verify step picks up Metro's Bundling complete (or the next error block) automatically.

Append to .claude/debug-app/fixes.md:

[<HH:MM:SS>] Import/Bundle — <file>:<line> — <recipe applied>

D3.2 Best-effort fix recipe for uncategorized errors

When an error falls through every row of Step B's classification table AND every row of D3.1's bundle recipes, the skill still attempts a fix instead of stopping. The discipline below keeps best-effort from degrading into wild guessing.

Step 1 — Locate the cite. Try in order; stop at the first that yields a file:line in editable user code:

1. Stack trace top user-frame — walk the stack from the top, skip frames in node_modules/, src/generated/, and React/Hermes internals (react-native/, hermes-engine/, metro/). First remaining frame is the cite.
2. Multi-line bundle block — see Step B's parsing note (file:line on the first non-banner line).
3. Verbatim grep across the repo — Grep for the exact error message text (or its most distinctive 4–6 word phrase, with regex special chars escaped) across app/, src/components/, src/hooks/, src/services/. A match at a throw new Error('...') site IS the cite.
4. Module + symbol grep — if the error mentions a function or component name (e.g., useFoo is not a function), Grep for the symbol; the unique declaration site is the cite.

If no cite can be located by step 4: log a structured note to .claude/debug-app/unresolved.md (verbatim error + which lookup attempts ran), surface to the user, advance to next issue. Do not guess at a file. Best-effort still requires a target.

Step 2 — Enrich understanding (do not skip).

• If the error contains a Microsoft-stack token (AADSTS\d+, Dataverse, Power Platform, MSAL, Entra, Graph API): query mcp__microsoft-learn__microsoft_docs_search with the exact code or token. A matching doc usually pins the fix exactly.
• Read the cited file ±15 lines for surrounding context. Note recent imports, the function signature, and any nearby try/catch or useEffect deps.
• If the error mentions a third-party module (anything in node_modules/ from the stack), one targeted WebFetch against the module's npm page or GitHub README is acceptable; do NOT do open-ended web searches in the loop.

Step 3 — Form ONE named hypothesis. Write it to .claude/debug-app/fixes.md BEFORE editing, in this format:

[<HH:MM:SS>] Hypothesis (best-effort) — <file>:<line> — <one-sentence theory>
  Evidence: <what in the error message + cited code led you here>
  Planned change: <what you'll edit, in 1 line>

Examples of acceptable hypotheses:

• "<UserAvatar> reads user.profile.image but user can be undefined during the first render — add a null guard"
• "The useEffect at line 42 captures a stale userId because userId isn't in its deps array"
• "AsyncStorage.getItem returns null for missing keys, but the caller assumes JSON-parseable string"

NOT acceptable (refuse to apply, escalate instead):

• "Something is wrong with state management" (vague — no specific change implied)
• "Try wrapping in try/catch" (mask, not fix — silently swallows the real bug)
• "Maybe also update X, Y, and Z" (multi-armed — violates one-fix-at-a-time)

Step 4 — Apply a single minimal edit. One Edit call, smallest possible diff that implements the planned change. Do NOT change unrelated code, rename symbols, or refactor surrounding structure. Re-confirm the file path is in editable user code (NOT under src/generated/, node_modules/, or any path in the Constraints section's protected list).

Step 5 — Defer to D4 verify. The existing verify cycle (type-check + reload + re-poll) decides whether the hypothesis was right. Do NOT preemptively try a second hypothesis "just in case."

Step 6 — On verify failure, ONE alternative is allowed. If D4 shows the same error

…(truncated)