vercel-optimize

Vercel Optimize

Run an observability-first Vercel optimization audit. Do not inspect source files until signals.json exists and a deterministic gate points to a route, file, or project setting.

Core doctrine: read references/doctrine.md if any rule is unclear.

• Metrics first. Recommendations start from Vercel production signals, not repo-wide grep.
• Deterministic gates. scripts/gate-investigations.mjs decides what deserves investigation.
• Candidate-bound scope. Read only files named by a candidate or a route-local import chain.
• Version-aware citations. Use only references/docs-library.json; invalid or version-mismatched citations are stripped.
• Customer copy. Read references/voice.md before writing report text or chat output.

Prerequisites

• Vercel CLI v53+ with vercel metrics, vercel usage, vercel contract, and vercel api.
• Authenticated CLI session: vercel login.
• Linked app directory: vercel link. VERCEL_PROJECT_ID can help resolve project config, but vercel metrics still requires directory linkage. The link or environment must include the intended project org/team/user scope so the collector can resolve a CLI-safe --scope and keep vercel metrics, vercel usage, and vercel contract on the same account.
• Node.js 20+.
• Observability Plus for route-level metric-backed recommendations.

Never put auth tokens in shell commands. Do not type VERCEL_TOKEN=..., --token ..., or Authorization: Bearer ... into commands that may be echoed in chat.

Framework Support

The preflight reads package.json and sets expectations before metric fan-out.

If unsupported, stop and ask before scanning or gating:

This project uses <framework>. Vercel Optimize supports metric-backed code recommendations for Next.js, SvelteKit, and Nuxt. Astro support is limited. For <framework>, I can still run a limited platform/scanner audit, but route-level Vercel metrics may not map back to source files.

Do you want me to continue with the limited audit, or stop here?

If the user continues, rerun collection with --continue-unsupported-framework.

Run Directory

Use a fresh run directory for every audit. Do not reuse briefs, sub-agent outputs, or reports across runs.

RUN_DIR="$(mktemp -d -t vercel-optimize-XXXXXX)"

Pipeline

1. Collect, scan, and merge signals

Run from the linked app directory or pass --cwd where a script supports it. Keep stdout JSON separate from stderr logs. Do not combine streams.

node scripts/collect-signals.mjs [projectId] >

Collection details, schemas, metric IDs, and degradation behavior live in references/data-collection.md. The metric registry is lib/queries.mjs; keep all queries on the shared 14-day window.

collect-signals.mjs resolves the linked project owner to commandScope.cliScope and verifies that the resolved account can read the resolved project before it checks Observability Plus. Downstream scripts reuse that scope for every Vercel CLI command that accepts --scope. Do not run vercel usage, vercel metrics, or vercel contract manually without the same scope; unscoped usage can report the user's personal organization while route metrics come from the team project.

If project or scope resolution is ambiguous, stop and ask the user which Vercel project and team/personal scope they want audited. Do not infer the intended scope from the current vercel whoami team, and do not proceed with metrics, usage, or contract collection until the link, an exact project match in .vercel/repo.json, or VERCEL_PROJECT_ID + VERCEL_ORG_ID identifies the intended account.

Use this prompt for PROJECT_SCOPE_UNRESOLVED, SCOPE_UNRESOLVED, or PROJECT_SCOPE_MISMATCH:

1.1 Stop on blockers

Check blockers before gating:

Required actions:

• frameworkSupportBlocker === "unsupported_framework": use the unsupported-framework prompt above.
• PROJECT_SCOPE_UNRESOLVED, SCOPE_UNRESOLVED, or PROJECT_SCOPE_MISMATCH: stop and ask which Vercel project and team/personal scope the user wants audited. For team projects, rerun after vercel link --yes --project <project-name-or-id> --team <team-slug>; for personal projects, rerun after linking under the intended user account or after setting both VERCEL_PROJECT_ID and VERCEL_ORG_ID.
• observabilityPlusBlocker === null: continue.
• no_traffic: tell the user route metrics are sparse; continue only if they accept limited output.
• payment_required or no_oplus_probe: render references/observability-plus.md verbatim and ask.
• project_disabled: tell the user to enable Observability Plus for the project or accept a limited audit.
• daily_quota_exceeded: stop and tell the user the Observability query quota is exhausted; retry after the next UTC midnight reset, or ask whether to continue with a limited code-only audit.
• not_linked: link the app directory, then rerun Step 1. If app path and project are known:

• forbidden or project_not_found: fix auth/team scope. Do not pitch Observability Plus.
• all_failed_other: show the raw error code and ask whether to continue in limited code-only mode.

Do not silently fall back to code-only mode. If the user accepts a limited audit, rerun collection with:

Then scan and merge again.

2. Gate candidates

Output shape:

• toLaunch: code-scope candidates to investigate.
• platform: project/account-scope recommendations.
• gated: skipped, covered, or disqualified candidates that must still appear in the report.
• budget: candidate budget and selection mode.

Default budget is 6 code-scope candidates with a diversity guardrail. To expand:

Generated candidate docs: references/candidates.md.

2.1 Ask about audit scope when needed

Before deep-dive, run:

If shouldAsk is false, continue.

If shouldAsk is true:

1. Print exactChatMessage.body exactly as returned. Do not summarize, truncate, reorder, or rewrite it.
2. Then ask questionText using questionPayload when the host supports structured questions.
3. If the user chooses a different number, rerun the gate with --max-candidates <choice>.

Never put the long preview inside the question field. The preview and the question are separate surfaces.

2.2 Deep-dive and reconcile

--cwd must be the linked project directory so deep-dive.mjs can verify the same project link and reuse signals.json.commandScope.cliScope for any follow-up vercel metrics calls.

Reconciliation deterministically converts disproven candidates into observations before any source investigation:

• metric_mismatch
• error_storm
• deployment_regression
• scanner_only_no_metric

2.3 Generate briefs and investigate

List the work:

Generate one brief for every entry in briefs-manifest.json.briefs. The group can be toLaunch or platform; do not generate only toLaunch briefs.

Use briefs-manifest.json.briefs[].label for visible worker names, for example Low cache-hit route on /docs/llm-digest/[...slug], not toLaunch-7.

Fan-out rule:

• 1-2 briefs: investigate inline.
• 3+ briefs: spawn one sub-agent per brief when the host supports it.
• Hosts without sub-agents: run inline serially.

Sub-agent contract:

• The brief is the whole prompt.
• Read only files listed in the brief, plus route-local imports when needed.
• Emit one JSON recommendation or one JSON no-change finding using references/recommendations.md.
• Do not cite URLs outside the provided citation subset.
• Do not recommend framework features unavailable in the detected version.

If a sub-agent reaches for repo-wide grep, the candidate is malformed; drop or abstain rather than widening scope.

2.4 Collect outputs

Save each raw investigation result in $RUN_DIR/sub-agent-outputs/, then collect:

The collector extracts JSON, prepends pre-resolved records, enforces manifest order, and fails on missing, duplicate, unknown, or mismatched candidateRef values.

3. Verify recommendations

This script extracts claims, verifies files/citations/version fit, grades quality, applies sanitizers, emits verifiedRecommendations, withheldRecommendations, renderableRecommendations, and creates regenPlan for failed or unsafe recommendations.

Recommendation schema, writing rules, sanitizer order, and grading rules: references/recommendations.md. Verification rules: references/verification.md.

For each regenPlan entry, rerun the same brief with a Previous attempt failed these checks section listing topFailures. Keep the regenerated output only if verification improves without gutting citations.

4. Render report and final message

Use --debug-out "$RUN_DIR/debug.json" only when developing the skill. Customer Markdown and chat output must not expose passRate, quality, sanitizer trails, raw sub-agent names, or other implementation fields.

After rendering, print final-message.json.body verbatim and stop. Do not add highlights, debug notes, raw counts, sub-agent summaries, or extra explanation. Render-time dedupe, platform caps, and hard-safety drops can change the customer-visible count, so never summarize from raw verify.json.

Report structure and impact framing: references/scoring.md.

Recommendation Rules

Every recommendation must:

• Trace to a launched candidate, platform candidate, pre-resolved observation, or verified traffic-independent scanner finding.
• Include observed metric evidence from signals.json or evidence.deepDive.
• Cite verified files with line numbers when code is involved.
• Include at least one allowed citation that applies to the detected framework/version.
• Use precise observed performance numbers.
• Use cost magnitude phrases only; never customer-facing $N savings.
• Do not recommend duration reductions for Vercel Workflow runtime endpoints (/.well-known/workflow/v1/*). These are generated orchestration routes for durable step/flow execution and should be hard-gated before investigation.
• Workflow recommendations must name the boundary being changed. Valid examples: enqueue durable work and return a run ID instead of awaiting completion, fix stream replay/closure/locks, or reduce verified excess Workflow Steps/Storage. Do not infer cost savings from Workflow endpoint wall-clock duration.
• For streaming, SSE, resumable chat, or other intentionally long-lived routes, do not frame wall-clock function duration as a problem by itself. Require evidence of avoidable pre-first-byte work, high active CPU, duplicate invocations, or post-response work that can move out of the user-visible path.
• Name a specific cache policy when recommending caching.
• Keep unsafe responses dynamic unless evidence proves they are safe to cache: auth-sensitive paths, errors, fallback responses, missing content, invalid requests, geolocation/device-varying output, and unversioned dynamic URLs.

Never recommend "verify X is on" for facts already present in signals.project, including Fluid compute status, memory tier, regions, in-function concurrency, and timeout.

Scanner Rules

Scanner findings are supplementary. Drop findings annotated COLD-PATH or NO-ROUTE-MAPPING unless the scanner declares metadata.trafficIndependent === true.

Traffic-independent examples: middleware matcher, source maps, React Compiler config, build settings. Route-local cache or data-fetch patterns need route-level traffic evidence.

Scanner docs: references/scanner-patterns.md.

Final Customer Terms

Use:

• recommendations ready
• observations from investigation
• investigated, no change recommended
• not investigated in this run

Avoid:

• sub-agent
• abstention
• passRate
• quality score
• gate
• LLM

Failure Copy

Use these messages without adding sales copy or process detail.

No traffic in the last 14 days:

This project has no meaningful traffic in the last 14 days, so route-level metrics are sparse. I can still check traffic-independent scanner findings and project settings, but I cannot rank route fixes until traffic accumulates.

Route-level metrics unavailable:

Use the verbatim choice template in references/observability-plus.md. Do not silently fall back to code-only mode; present the two-path choice: enable Observability Plus and rerun the metric-backed audit, or accept a limited code-only run.

Project is not linked:

This worktree is not linked to a Vercel project. Run vercel link --yes --project <project-name-or-id> --cwd <app-dir> and rerun the audit. If the team is known, add --team <team-id-or-slug>.

Most route-to-file mappings failed:

The route inventory matched fewer than half of the routes we saw in observability. This is common in monorepos with custom routing. I've surfaced what I can match; the rest appear in the "Not investigated in this run" section.