AEM as a Cloud Service — Code Assessment

Single skill for detecting and fixing AEM CS code-quality issues, entirely against the local workspace — no external services or network calls. Findings reach the runbook from one of two sources; everything downstream is identical.

Findings sources

Discovery runs through the deterministic analyzer (scripts/analyze.sh): it parses the workspace once and runs the enabled detectors, emitting the shared findings shape. Every ready pattern has an analyzer detector. One detector — remove-deprecated-api — loads its rules dynamically from a preflight-produced cache (remove-deprecated-api/scripts/detect.sh runs the AEM Analyser Maven Plugin and writes the cache TSV before the analyzer is invoked); the detector's shape and integration are otherwise identical. Patterns without a detector are planned only — not yet detectable and not yet built; there is no LLM-scan fallback in this version (see Scope & limitations) — the scan value on planned rows in references/patterns.md marks the intended future detection method, not an active code path.

Routing

1. User named files / coordinates → run the runbook in with_findings mode against those paths.
2. "Scan my repo" / no files named → run the runbook in discover mode (per-pattern Discovery, workspace roots only).

Then follow the runbook: references/runbook.md.

Manual Pattern Hints (classification → expert skill)

Route the request to one expert skill. Two pattern families share this skill:

Mechanical fixes (analyzer-driven detection, deterministic edits — follow the runbook flow):

Architectural migration patterns (guided remediation — full before/after, troubleshooting, modern alternatives; invoked directly or via migration for BPA/CAM-driven discovery):

Broad / correctness-review asks ("check my Sling Models are implemented correctly", "review my code", "is my AEM project healthy", "assess this project") are not a single pattern: run the runbook in discover mode with intent report — the analyzer runs every detector and the report covers all built patterns, explicitly noting aspects not yet supported. Only narrow to one pattern when the user targets a specific fix.

If nothing matches, say the issue is not yet supported and offer to file a request for a new expert skill.

Full catalog (built + planned patterns, with severity / detection / fix): references/patterns.md.

Invocation from the migration skill

migration performs BPA/CAM/MCP discovery and handles batching + one-pattern-per-session workflow. After it has identified (pattern, file) pairs from BPA findings, it hands off here for the actual transformation. When invoked with (pattern, file) from migration:

• Skip HA/analyzer discovery (caller already identified the pattern + file)
• Open the pattern's expert skill directly (per the Manual Pattern Hints table above)
• Apply the steps in the expert skill against the named file(s)
• Return the result; migration continues with the next finding in its batch

The pattern guides themselves are agnostic about who invoked them — they apply identically whether reached from migration (BPA/CAM) or from the runbook in this skill (HA / analyzer).

Runbook

All detection, planning, edits, verification, git/in-place handling, and the run log live in references/runbook.md. The runbook is the sole owner of repo-environment detection (edit_mode, git snapshot) — this control plane does not duplicate it.

One pattern per session

Report may span every pattern found; apply touches one pattern per session (atomic revert, single-story diff). Refuse "fix everything" for the apply phase. Rationale: references/shared-principles.md.

Critical rules

• Local only — no network calls or external services; operate solely on the workspace. Documented exception: remove-deprecated-api is plugin-driven and needs Maven Central (to resolve aemanalyser-maven-plugin and, transitively, the AEM SDK's api-regions data) plus optionally Adobe Experience League as a fallback source for successor guidance. If offline, that one pattern is skipped with a clear message; all other patterns remain local-only.
• Requires a local JDK (Java 11+) for detection — the analyzer compiles/runs in memory; no install beyond the JDK, no network. If absent, detection stops with a clear message.
• The analyzer is detection — never substitute external tooling. Do not run mvn versions:display-dependency-updates / mvn versions:display-property-updates, npm outdated, or Maven Central / registry lookups in place of analyzer discovery. Those answer "what is the latest on the network" — outside this skill's local-only contract. If the user explicitly wants a live registry comparison, say it needs network and offer it as a separate step after delivering the skill report. remove-deprecated-api's preflight (remove-deprecated-api/scripts/detect.sh) is the one documented exception: it invokes the AEM Analyser Maven Plugin against the project to populate its rules cache, then hands off to the shared analyzer.
• Never commit, push, or open a PR — branch (git) or in-place edits only; the developer reviews and commits.
• Surgical edits — no reformatting / re-serialization.
• Skip with a reason — record un-applicable findings as skipped with an exact reason; never silently drop.
• One pattern per session for apply.

Full rationale: references/shared-principles.md.

Scope & limitations

Local static detection and remediation only — no external services, no network, no live AEM instance. Issues that require runtime or live-repository state, telemetry, or history across runs are out of scope for this skill. Detection requires a local JDK (Java 11+); there is no remote or LLM-scan fallback in this version. A large apply (e.g. an @Inject migration across 100+ files) is processed in resumable batches: the run checkpoints each file to .autofix/last-run.json and pauses at a per-pass cap, so it survives context limits — reply apply <pattern> to continue (see references/git-workflow.md).

Adding a new pattern

Full end-to-end procedure — detector → fixtures/tests → catalog + routing → expert skill → verify: references/adding-a-pattern.md. The [wiring] test keeps the detector, catalog row, and expert-skill directory in sync.

Triggering scales without touching the description. The description above is intentionally broad (intent verbs + AEM domain), so it already fires on "check / review / fix my <AEM thing>"; a new pattern is reached by its Manual Pattern Hints + patterns.md rows, not by editing the description. Update the description only if the new pattern introduces a domain keyword it does not already cover (a new subsystem or file type). The [wiring] test keeps the detector, catalog row, and expert-skill directory in sync.

Related skills

• migration — drives BPA/CAM/MCP-based legacy-AEM migration workflow. Discovers findings, batches them, enforces one-pattern-per-session, and hands off (pattern, file) pairs to this skill for transformation. See the "Invocation from the migration skill" section above.