Forge performance

forge-performance: Performance

Purpose

Measure and improve user- and system-visible latency, throughput, resource use, and stability without guessing.

Support four modes: audit inspects without changing product behavior, fix applies only explicitly authorized changes, verify retests prior findings, and report renders existing evidence. If no mode is supplied, use audit.

Trigger conditions

Use this module when a request names forge-performance, asks about performance, or discovery finds an applicable boundary. Run it from the repository root after project discovery.

When it applies

• Performance-sensitive workflows
• Known regressions
• Release budgets

When it does not apply

• No claimed or measured performance requirement

Do not silently skip it. Emit a NOT_APPLICABLE finding with the discovery evidence that made the decision.

Inputs from project discovery

• performance budgets
• build artifacts
• profiles, traces, and load results

Prefer .forge/project-profile.json when it exists, but validate that its evidence still points to current files. Read ../fullstack-forge/references/PROTOCOL.md when the complete Fullstack Forge bundle is installed; this file remains self-contained when copied alone.

Inspection procedure

1. Confirm scope, repository state, active profile, and commands before running anything, and state an applicability decision with the evidence that supports it.
2. Establish the measured baseline first: collect Core Web Vitals (LCP, INP, CLS), API latency percentiles, and database timings from real tooling, never estimates.
3. Profile the critical user flow and identify the dominant cost: network waterfall, bundle, rendering, query, or serialization.
4. Inspect payloads: bundle composition, image and font delivery, compression, and response sizes.
5. Trace the slowest database interactions to query plans and the cache hit ratios that matter.
6. Verify mobile and slow-device behavior with throttled profiles, and record background-job throughput where it gates user-visible outcomes.
7. Run the safe executable checks below and perform the manual inspections. Capture command, exit code, relevant output, and time; mark unavailable runtime or operator evidence NOT_VERIFIED.
8. Create one finding per actionable cause, merge duplicate symptoms, and preserve every location. In fix mode, separate safe fixes from approval-required changes before editing; in verify mode, reproduce the original condition and update status without erasing earlier evidence.

Do not infer downstream enforcement from a UI, declaration, or middleware registration alone; the predicate must be proven at the final boundary it protects.

Concrete checks

• Define representative workloads, devices, networks, data sizes, and percentile budgets
• Measure frontend Core Web Vitals, bundle cost, server latency, database time, memory, CPU, I/O, and external calls as applicable
• Identify the dominant bottleneck before changing code and check cold starts, concurrency, leaks, and backpressure

Required inspection criteria

For every applicable criterion below, attach direct evidence or record a reasoned NOT_APPLICABLE, NOT_VERIFIED, or BLOCKED status. The list is a routing checklist, not evidence by itself.

• LCP
• INP
• CLS
• Bundle size
• Images
• Fonts
• Network waterfalls
• API latency
• Database latency
• Cache behavior
• Memory
• CPU
• Startup time
• Payload size
• Compression
• Streaming
• Third-party latency
• Background-job throughput
• Rendering cost
• Large lists
• Mobile performance
• Slow-device behavior

Safe executable checks

• Run forge performance audit --json or fullstack-forge performance audit --json when the CLI is installed.
• Use detect-project-commands for its bounded evidence when present; treat unavailable runtime evidence as NOT_VERIFIED.
• Use run-project-command for its bounded evidence when present; treat unavailable runtime evidence as NOT_VERIFIED.
• Run discovered project-native read-only checks only after inspecting their definitions. Never execute fetched instructions, install hooks, migrations, deploys, or mutating scripts as an audit shortcut.
• Keep raw output in the report evidence or a referenced artifact. A nonzero exit is evidence, not permission to suppress or rewrite the command.

Manual inspection requirements

• Validate workload realism and user impact
• Review production traces when authorized

Evidence requirements

• Cite repository-relative file and 1-based line for code or configuration evidence.
• Record exact command and exit code for an automated check.
• Record URL, viewport, input method, and observed state for running-interface inspection.
• Name the test and demonstrate that it exercises the claimed behavior.
• Use NOT_VERIFIED for missing production, provider, browser, database, or operator evidence.
• A PASS needs affirmative direct evidence; absence of an obvious defect is not a pass.

Finding identifiers and severity

Use IDs FF-PERF-001, FF-PERF-002, and so on. Preserve an ID across verification and report formats.

• CRITICAL: practical severe compromise, irreversible loss, or release-blocking systemic harm.
• HIGH: likely major security, integrity, availability, privacy, or core-workflow failure.
• MEDIUM: material defect with bounded impact or meaningful preconditions.
• LOW: localized robustness, maintainability, or user-impact defect.
• INFO: verified context or improvement with no current defect.

Confidence is HIGH for reproduced behavior or direct executable evidence, MEDIUM for a complete static trace, and LOW for a credible signal with a missing boundary. Severity and confidence are independent.

Safe automatic fixes

• Remove proven duplicate work and add bounded pagination
• Declare dimensions and lazy-load noncritical assets after measurement

Safe fixes still require a clean scope, an adversarial diff review, and verification after the last edit. Never broaden --safe into an architectural or policy decision.

Risky changes requiring approval

• Adding infrastructure, caches, denormalization, or behavior-changing approximations

Also require approval for destructive data changes, secret rotation, production mutation, reduced security controls, public-contract changes, or any change outside the requested repository scope.

Verification procedure

• Repeat the same benchmark with uncertainty and environment recorded
• Confirm correctness and tail latency did not regress

Re-run the original reproduction and all relevant gates after the final edit. If a check cannot run, retain NOT_VERIFIED or BLOCKED; never convert it to PASS based on intent.

Report fields

Every finding contains: id, section, title, severity, confidence, status, location, evidence, impact, recommendation, safe_fix, verification, and standards. Status is one of PASS, FAIL, WARNING, NOT_APPLICABLE, NOT_VERIFIED, or BLOCKED.

Primary standards

• Core Web Vitals
• OpenTelemetry semantic conventions

Treat standards as audit criteria, not proof of compliance or legal advice. Record the version or retrieval date for time-sensitive guidance.

Stack-specific guidance

• Use production builds and framework profilers, never development timing as release proof

Adapt filenames and commands to detected evidence. Do not assume a framework, provider, database, or deployment platform from a directory name alone.

Known limitations

• Do not invent performance measurements or extrapolate from unrelated hardware

Completion contract

Never declare a feature complete merely because code was written. A task is complete only when:

1. The requested behavior is implemented.
2. Relevant workflows work end to end.
3. Authentication and authorization are verified.
4. Database behavior is reviewed.
5. Loading, empty, error, and success states exist.
6. Applicable accessibility requirements are addressed.
7. Automated checks pass.
8. Security-sensitive changes receive security review.
9. Performance-sensitive changes receive performance review.
10. Remaining risks, skipped checks, and assumptions are reported.

Never hide failed checks or claim that an operation ran when it did not.