Android Memory Evidence

Build a trustworthy evidence boundary before interpreting memory values. Accept partial inputs, explain what they support now, and give exact collection guidance for what remains unknown.

Workflow

1. Preserve the supplied files. Never rewrite, rename, decompress, upload, or delete raw artifacts without explicit authorization.
2. Accept the normal RD intake as one QA handoff directory plus the issue title or symptom. Do not ask the user to classify, rename, or enumerate its files first. Record any supplied target package/process, device/API, scenario, phase, and whether the symptom appears Java, native, graphics, system-pressure, or regression oriented.
3. On every run, rescan the complete handoff directory. Never restrict a follow-up pass to the files the user says are new: unchanged, changed, missing, renamed, unclassified, and newly added artifacts all affect the evidence boundary.
4. Run this Skill's scripts/build_context.py against the directory and question. It recursively inventories the tree, validates supported artifacts by content, preserves multiple logs/screenshots/dumps, reports unclassified or skipped files, and augments vague question routing with evidence signals. Check request.intent_source and every evaluated_intent. The bundled versioned runtime and knowledge catalog require no separate repository checkout. Prefer JSON for another AI and Markdown for a human.
5. If the request mentions a prior analysis, dissatisfaction, continued analysis, or new evidence, or if evidence.analysis_history.has_prior_analysis is true, read references/analysis-iteration.md completely. For every mode except explicit initial, inspect both the current conversation/task history and every listed prior-analysis artifact. Do not assume the folder contains the previous assistant response; in explicit initial mode, retain old material as inventory without applying it as a baseline.
6. Check request.analysis_mode, request.analysis_mode_source, evidence.analysis_history.baseline_applied, and evidence.analysis_history.evidence_delta. A clarification-required mode permits inventory and context construction, but requires the user to choose full reanalysis or supplementation before a final diagnosis is presented.
7. Use --repo only when intentionally testing a source checkout. Do not make a checkout or environment variable a prerequisite for normal context generation. If an optional collection command needs the full analyzer repository, identify that dependency separately from the bundled context runtime.
8. Read the resulting primary coverage, every entry in intent_coverage, qa_observations, conflicts, invalid artifacts, limitations, and next_evidence before drawing any conclusion. A mixed question must satisfy every evaluated claim contract.
9. When QA logs or screenshots exist, read references/qa-artifacts.md completely. Use log signals only to navigate the original authorized lines and inspect each relevant screenshot; zero matches do not make a log irrelevant. The runtime deliberately performs no OCR and embeds no raw log lines or pixels.
10. Classify current statements as observed, derived, hypothesis, or recommendation. Bind observations to artifact_id; bind logs to line/hash and screenshots to a visible region.
11. Continue with a bounded explanation even when evidence is incomplete. Do not turn missing evidence into a guessed fact.
12. Give missing-evidence commands in priority order. Preserve prerequisites, permission boundaries, Android-version gates, and perturbation level.
13. Hand a valid context to $android-memory-diagnose for theory-grounded interpretation.