Levyra context-efficiency workflow

Purpose

Spend model context on code and decisive evidence. This skill reduces repeated instructions, broad reads, noisy command output, and stale history. It never replaces domain skills, source inspection, testing, review, or exact diagnostics.

Token savings must come from context and output, never from shallower reasoning. If omitted evidence can change correctness, read or rerun it.

Automatic routing

Route this skill only when the task is likely to create substantial repository or command-output volume. Tiny edits, ordinary explanations, and already-local code changes should not load it just because the prompt says "implement", "modify", "analyze", or "inspect".

Before broad reading:

1. identify the architecture owner and the exact question the next read answers;
2. search symbol/path/call site first;
3. read the smallest useful source/test range;
4. expand only when a concrete unanswered question remains;
5. do not reread unchanged evidence already in context;
6. load only the domain/companion skills that materially affect correctness.

RTK

For shell-capable noisy work, prefer the repository RTK layer after checking it. Use scripts/ensure-rtk.ps1 -Quiet on Windows or ./scripts/ensure-rtk.sh --quiet elsewhere. Manual repair remains available through scripts/setup-ai.ps1 -InstallRtk or scripts/setup-ai.sh --install-rtk. If RTK is unavailable, continue raw rather than weakening validation.

Useful compact routes include:

rtk gradlew <tasks>
rtk git diff
rtk git status
rtk gh pr view <number>
rtk test <command>
rtk err <command>
rtk grep <pattern> <path>
rtk log <file>
rtk adb logcat -d -t 400
rtk summary adb shell dumpsys <service>

Compact output is not proof of success. Check the command exit status and the authoritative success/failure marker. If compression hides the deciding cause, rerun the exact command raw.

Keep decisive evidence raw

Do not compress away evidence needed for:

• compiler/test/lint failures whose exact diagnostic matters;
• security, redirects, MIME, permissions, secrets, signing, checksums, or trust boundaries;
• Perfetto/thread/frame timing, SQL/query failures, concurrency, or memory root cause;
• R8/Proguard missing-class, mapping, metadata, or release-only failures;
• exact protocol, quoting, encoding, stdout/stderr, or regression reproduction.

For ADB, bound noisy textual output first. Keep tiny control queries raw. Never wrap binary/payload commands such as adb exec-out screencap -p in a text compression path.

Cross-session context

Use claude-mem only when earlier-session context materially affects the task and the runtime exposes it. Retrieve progressively: search -> timeline when chronology matters -> get_observations for only relevant IDs -> verify against the current repository. Current code, tests, CI, runtime evidence, and owner decisions always outrank memory.

Fail open: if memory tooling is unavailable, unhealthy, or unsupported, continue ordinary engineering without it. Do not enable cloud sync or semantic injection implicitly. For ChatGPT, Repository configuration alone cannot make ChatGPT reach a local claude-mem worker.

If a shell-capable runtime genuinely needs the optional integration and it is missing, one bounded setup attempt may use scripts/setup-ai.ps1 or scripts/setup-ai.sh. Manual forcing remains scripts/setup-ai.ps1 -ClaudeMem or ./scripts/setup-ai.sh --claude-mem. Failure must not block ordinary work.

Never store or retrieve secrets, tokens, cookies, keystores, private URLs, .env, or local.properties through project memory.

Long-task checkpoints

Carry forward only the verified goal, root cause/decision, affected files or symbols, preserved behavior, current edit state, validation results, real blockers, and one next action. Drop superseded logs and disproved hypotheses. A compact handoff never replaces source-of-truth evidence.

Safety

• Do not trade correctness, review depth, or testing for a smaller context window.
• Do not enable danger-full-access, approval bypasses, or unrestricted sandboxing.
• Do not install unrelated plugins or broad system upgrades.
• Do not let RTK or memory hide security/signing/runtime evidence.
• Never infer permission to commit, push, open/merge a PR, tag, release, or deploy.

Validation

After changing this workflow or its routing, run:

python3 scripts/validate_claude_mem.py
python3 scripts/validate_agent_config.py
python3 scripts/validate_ai_efficiency.py
python3 scripts/evaluate_skill_routing.py

On Windows use py for the same scripts when appropriate.