troubleshoot-app

Troubleshoot App

Use this skill to diagnose live app problems from the outside in: reproduce the user-visible issue, correlate it with data-plane truth, inspect logs and code, then recommend a fix. Keep it globally usable; discover the project's actual sources instead of assuming a stack.

This is quality regime B (interactive product). Wrong outputs from APIs/pipelines with no UI → diagnose-bug (A). Thrashing agents / bad LLM generations → agents analyze / Langfuse traces (C), not this skill.

Use lightweight BDD completion scenarios and existing requirements as the definition of what should be happening (handbook/concepts/13-quality-trace.md). Look in DocSlime docs/ (PRODUCT.md, experience/, REQUIREMENTS.md, engineering/TESTING.md, …), issue bodies, PR descriptions, existing .feature/tests, and acceptance criteria. Prefer refining an existing definition over inventing a parallel one. If none fits, draft a concise Given/When/Then scenario and say where it should live. A broken, confusing, or falsely promised experience is still a bug — including craft/a11y/framing debt (handbook/concepts/12-bugs-and-debt.md).

For backend-only, API, worker, data-pipeline, or algorithm failures (no meaningful browser UI), use diagnose-bug instead.

Operating contract

Shared by every DecisionNerd/dev-skills skill. Canonical copy: handbook/concepts/14-operating-contract.md.

• Define done first. Before using tools, write the finish line in one or two lines: acceptance criteria (existing issue AC, BDD scenarios, tests, or contract; otherwise propose them and where they should live) and the evidence that will prove them. Re-check it before reporting done. Never report done on unverified work.
• Requested scope sets the finish line. A question ("is it ready?", "why is it broken?", "what next?") ends with the answer and a Next: line naming the exact next invoke. An outcome request ("fix", "finish", "land", "#42") continues through the chain (diagnose → fix → test → check-readiness → merge-it) until the outcome or a real blocker. Do not end a turn with "Do you want me to…?" for in-scope, in-repo work.
• Stop only for real blockers. Stop and ask only when you cannot continue without the user, or before: deleting data or unmerged work, force-push or history rewrite, changing anything outside this repository (GitHub objects, deployments, live data, production or paid resources, external services), or leaving the requested scope, unless the user's request already named that exact action. Keep the harness's permission prompts for risky commands. Otherwise keep going and put status notes in the same message as the next action.
• Ask well, once. For a genuine question use the harness's structured question tool when it has one (Claude Code: AskUserQuestion; Codex: request_user_input when the current mode supports it; T3 Code: the question tool) with concrete options; otherwise plain text with numbered options. Treat the answer as settled; do not re-open earlier verdicts, plans, or answers unless asked.
• Fan out when work is parallel. Use subagents for independent reads (repo survey, evidence gathering, per-option research, per-area audits) and independent verification (a reviewer that did not write the change). Writes stay single-owner per path set and sequential. Brief every child with goal, done-when, constraints, must-not, and return shape; verify each result before consolidating. Use Claude Code's Workflow tool only when orchestrating many subagents truly needs it; it is expensive. T3 Code: its delegated subagents — with optional worktree isolation for independent verification — cover this without the Workflow tool.
• Pick the model tier per child; defer to routing config. If the harness or user config already routes subagents, follow it and do not pass a model. Otherwise pick a tier per child: mechanical search or inventory → fast/cheap; implementation and evidence gathering → mid; planning, review, adversarial verification → top. Tier names differ per harness (Claude Code: haiku / sonnet / opus or fable; T3 Code: inherit by default or pick from its live provider/model catalog). Set the choice with the harness's spawn model param or agent frontmatter model: (per-harness details: skills/agents/references/sub.md). Where the harness cannot choose, children inherit or the harness picks; say which in the status note.
• Keep a checklist on long runs. For more than about five steps or work that crosses skills, keep TASKS.md at the repo root and tick items as they finish. Do not commit it unless the repo already tracks one.
• Close every run with three headings. Blocked on me (the one genuine question or blocker, else "none"); Changed (files, commits, GitHub objects, deploys, else "nothing"); Found (evidence, verdict, and Next: <exact invoke>).

Core Rule

Do not jump straight from screenshot to code. Establish what the user sees, what the app believes, and what durable data says. Treat analytics as evidence, not authority.

Done when: the named BDD scenario or contract passes, a regression test is added, and no new failures appear in the narrowest suite.

A bug report is a fix request. After diagnosis, implement the smallest fix and add a regression test in the same run — report the diagnosis as a status note before coding. Stop before implementing only when:

• the user asked for diagnosis or explanation only, or said "don't change code";
• the fix requires a production config change; or
• the fix requires a live-data mutation or a change outside this repository that the user's request did not name exactly; or
• two materially different fixes are viable and the choice changes product behavior — then ask one structured question with the options (<option A> / <option B> / stop).

If data repair is needed, describe the exact mutation first and ask before mutating live data; skip the question only when the user's request named that exact mutation.

Workflow

1. Reproduce the visible state.

• Use Computer Use for Atlas or the user's named browser when they ask to see what they see.
• Capture the URL, visible state, selected account/org/workspace, error copy, disabled controls, and any loading or redirect loop.
• Avoid risky browser actions. Follow Computer Use confirmation policy for account changes, submissions, billing, permissions, uploads, destructive actions, or sensitive data transmission.

1. Identify the intended workflow.

• Determine the user journey, entry URL, auth state, selected tenant/org/account, and expected landing state.
• Search for the intended behavior in existing docs, requirements, acceptance criteria, BDD/Gherkin files, test files, issue text, PR text, and relevant code comments.
• Link the expected behavior back to a concrete source when one exists, such as DocSlime REQUIREMENTS.md / experience/, a .feature file (only if the repo already uses one), an e2e/unit/integration test, or a GitHub issue.
• If an existing definition is close but stale or incomplete, say how it should be refined rather than creating a competing definition.
• If no definition is suitable, draft a proposed BDD completion scenario using Given/When/Then and recommend where it should live.
• Name the invariant that appears violated, such as "active browser org exists but app has no private workspace," "checkout created a session but subscription is missing," or "UI says saved but database lacks row."

1. Discover data-plane sources.

• Read project docs, .env* variable names, scripts, package files, and deployment metadata to identify systems in use.
• Look for likely sources: application database, auth provider, billing provider, analytics, feature flags, queues/jobs, object storage, logs, and deployment platform.
• Use only available credentials/tools. Redact secrets and personal data in summaries.
• If a live data source is unavailable, say exactly which source is unavailable and why.
• Fan out evidence gathering to parallel read-only subagents (fast tier) when more than one source is available — browser/session evidence, data-plane queries, and log correlation can run concurrently; verify each result before diagnosing.

1. Query current data.

• Use read-only queries first.
• Correlate by stable IDs from the UI, URL, session, metadata, database rows, and analytics properties.
• Compare at least two layers when possible: browser/session, application database, provider system, logs/analytics.
• Check for stale/missing mirror rows, mismatched IDs, old immutable deployments, incomplete jobs, feature flag differences, and entitlement/billing denials.

1. Inspect code only after evidence.

• Locate the resolver, route, component, middleware, job, or webhook that should bridge the mismatched layers.
• Locate tests or BDD files that should already cover the intended behavior. If they do not exist, identify the missing test or BDD definition as part of the recommended fix.
• Look for fail-closed checks, idempotency gaps, retry holes, race conditions, missing backfills, stale deployment URLs, and unsafe fallback assumptions.

1. Diagnose.

• State the observed symptom.
• State the intended behavior definition used: existing requirement, existing BDD/test, refined definition, or newly proposed BDD scenario.
• State the data-plane facts.
• State the likely root cause and confidence level.
• Distinguish current facts from inference.

1. Recommend.

• Recommend the smallest fix that restores the intended workflow while preserving authorization, billing, privacy, and tenant boundaries.
• Recommend whether to use an existing requirement/BDD/test definition unchanged, refine it, or add a new definition before or alongside the implementation.
• Include any data repair, backfill, migration, logging, telemetry, and test coverage needed.
• Make the verification target explicit: the fix should prove the linked or proposed requirement/BDD scenario now passes.
• Identify immediate workaround only if it is safe.

1. Fix and prove.

• If the fix requires a live-data mutation or a change outside this repository that the user's request did not name exactly, state the proposed action and ask one structured question: <do it> / revise / stop. If two materially different fixes exist and the choice changes product behavior, ask the same structured question with the options named.
• Implement the smallest fix identified in step 7.
• Add a regression test that would have caught this bug.
• Run the narrowest test suite that covers the fix; confirm it is green.
• Re-check the scenario from step 2: the named BDD scenario or invariant must now pass.

Atlas And Computer Use

Use Computer Use when the user references Atlas, a current browser tab, or "what it looks like to me."

• Start with get_app_state before interacting.
• Prefer opening a new tab for separate URLs unless the user asks to continue in the current tab.
• Read the accessibility tree and screenshot. Record the page title, host, route, selected identity/workspace, and visible error text.
• Use navigation and reloads for diagnosis. Do not submit forms, create accounts, change billing, accept permissions, or modify account/org settings without the required confirmation.
• If the live page is an immutable preview/deployment URL, compare it with the current deployment for the relevant branch.

Data-Plane Evidence

Use references/data-plane-sources.md for source-discovery patterns and read-only query examples.

Best practices:

• Prefer read-only provider APIs, CLIs, dashboards, logs, and database queries.
• Correlate the same entity across systems before concluding.
• Check timestamps and deployment hashes; stale previews often explain "still broken."
• Check provider metadata and app database rows separately.
• Treat feature flags and analytics as rollout/behavior evidence, not access or billing truth.
• Do not paste secrets, tokens, raw private records, OAuth tokens, financial records, or private payloads into the final answer.

Response Shape

Keep the diagnosis concise and evidence-led. Report diagnosis as a status note before coding:

I reproduced the issue: ...

Expected behavior definition:
- Source: <docs/test/BDD/issue/PR link or "No suitable existing definition found">
- Scenario: <existing/refined/proposed Given/When/Then summary>

Data-plane checks:
- Browser/session: ...
- App database: ...
- Provider/logs/analytics: ...

Diagnosis: ...

Recommended fix: ...

## Blocked on me

<the one genuine question or blocker, else "none">

## Changed

<files changed by the fix, else "nothing">

## Found

Diagnosis: <root cause>. Fix: <what changed>. Verification: <test name(s) and result>.

Next: <check-readiness after a verified fix; otherwise the step that unblocks, e.g. fix-it or the named blocker>

Grounding

This skill's TTPs are grounded in current engineering baselines (DORA, GitHub Docs, Fowler/Beck, Google SRE & SWE book, OpenTelemetry, OWASP LLM / NIST AI RMF, Diátaxis — see handbook sources.md).

Live incidents: reproduce → isolate → evidence (SRE). Regime B (handbook/concepts/11-quality-regimes.md): journeys, a11y, Web Vitals, data-plane mismatch — not unit-test theater and not LLM-eval theater. Prefer the quality trace before inventing a private oracle (13-quality-trace.md).

Handbook card: handbook/practices/troubleshoot-app.md.