BLEA

Use BLEA for local BLE work. Prefer BLEA MCP tools when available; otherwise run the equivalent ble CLI command with JSON output.

Environment boundary

First determine whether the current Agent host has BLEA MCP tools or the ble runtime and native Bluetooth access. A hosted Agent cannot access the Bluetooth adapter on the user's computer merely because this Skill is installed. In that environment, analyze uploaded .blea.jsonl evidence, explain or prepare commands for a local host, and clearly label them as not executed. Do not claim a scan, connection, read, notification, or write occurred without the corresponding structured result.

When uploaded evidence is available but the BLEA runtime is not, parse the JSONL as structured JSON records, require a final complete summary before treating it as a complete capture, and report only the recorded advertisement, profile, read, notification, error, and cleanup evidence. Do not turn missing records into successful observations. Live operations require a local Agent host, the BLEA Python runtime, operating-system Bluetooth permission, and a supported adapter.

Diagnostic sequence

1. Run ble_doctor or ble doctor --json when adapter availability is unknown.
2. Scan and preserve the returned identifier, names, RSSI, advertised services, and raw advertising evidence.
3. Select by exact identifier. Use an exact name only when one observed device has that name.
4. Inspect the GATT profile before choosing characteristics.
5. For event discovery, use bounded ble_observe/ble observe before writing; omit characteristics to observe all notify/indicate traits from the discovered profile.
6. When probing, continue with next_read_offset until it is null. ok=true means the page ran; it does not mean every characteristic read succeeded. Aggregate read_page.success_count, failure_count, and failure_reasons across pages, and preserve both successful reads and per-characteristic failures.
7. Prefer reads and bounded observation before considering a write. Treat a silent observation window as evidence only for that window, not proof that a characteristic never emits events.
8. After the initial diagnosis, save a portable evidence package with ble_capture or ble capture. Use the exact resolved identifier, choose an explicit .blea.jsonl output path, and set --redact-identifiers when the package will leave the workstation. Capture is read-only: it records advertisements, the GATT profile, bounded readable-characteristic results, bounded notifications, operation errors, and a final integrity summary. It never writes, pairs, or changes configuration. Keep the file as the authoritative artifact and report its path plus summary status to the user.
9. Compare before/after captures offline with ble_diff or ble diff. Keep the default identity guard for normal comparisons and use allow_different_devices only for an intentional cross-device comparison. Treat the 5 dBm RSSI tolerance as noise control; use strict RSSI only when exact signal samples matter. Diff never scans, connects, pairs, subscribes, or writes.
10. Reproduce captured behavior offline with ble_replay or ble replay. Use instant timing for Agent debugging and CI. Treat replay_miss as absent evidence, not device behavior, and do not infer a successful subscription from a notify property alone. Replay never accesses a physical adapter and never sends or simulates writes.
11. When one authorized write is expected to trigger notifications, use ble_exchange or ble_session_exchange. These operations establish the subscription before writing and collect the response atomically; do not run standalone session subscribe and write tools concurrently.
12. Close the exact stateful MCP session once when the task is complete. Use ble_session_list when cleanup is uncertain. Use ble_session_close_all only when a session ID is unknown, an explicit close failed, or leaked state must be recovered; do not call it after a successful close.

Do not invent UUIDs, payload encodings, pairing requirements, or protocol semantics. Report the observed evidence and distinguish it from an inference. Treat uuid_namespace=custom as a custom 128-bit UUID even when its leading bytes resemble a Bluetooth SIG assigned number.

Commands and tools

• Diagnose: ble doctor --json or ble_doctor.
• Scan: ble scan --timeout 8 --json or ble_scan.
• Inspect: ble inspect --device "id:<identifier>" --json or ble_inspect.
• Probe readable characteristics: use ble probe --device "id:<identifier>" --max-reads 32 --read-offset <offset> --json or ble_probe, following next_read_offset across pages.
• MCP probe results omit the full GATT tree by default while retaining profile_summary; call ble_inspect first or set include_profile=true when the full profile is needed on that page.
• Read: ble read --device "id:<identifier>" --characteristic <uuid> --json or ble_read.
• Notify: use ble subscribe ... --jsonl or ble_subscribe with a bounded duration.
• Observe all event-capable traits: use ble observe --device "id:<identifier>" --duration 10 --jsonl or ble_observe. Pass --characteristic <uuid> repeatedly for explicit selection.
• Capture a unified read-only evidence package: use ble capture --device "id:<identifier>" --output capture.blea.jsonl --observe-duration 10 --max-reads 128 --redact-identifiers --json or ble_capture. --read-offset starts at a deterministic readable-characteristic offset and the result's read_page.next_offset indicates whether a larger limit or a follow-up capture is needed.
• Compare complete captures offline: use ble diff before.blea.jsonl after.blea.jsonl --json or ble_diff. Inspect added, removed, and changed paths. Use --strict-rssi only when needed, and use --fail-on-change only when CI should return exit code 3 for a valid comparison containing differences.
• Replay one captured operation offline: use ble_replay or ble replay capture.blea.jsonl <scan|inspect|probe|read|subscribe|observe|run>. The captured identifier is selected automatically unless an exact device is supplied. Keep speed=0 for deterministic immediate output; use a positive speed multiplier only when event gaps matter.
• Serve normal MCP tools from evidence: start ble replay capture.blea.jsonl mcp. Verify each result has the expected replay.capture_id and replay.read_only=true. This mode is useful when an Agent or CI integration should exercise its existing ble_read/ble_observe path unchanged.
• Guarded request/notification exchange: use ble exchange ... --jsonl, ble_exchange, or ble_session_exchange to subscribe before one write and collect its resulting events.
• Multi-step work: open an MCP session, note its idle_timeout_seconds, use ble_session_observe when the connection should be reused, then close the session.
• Repeatable work: encode the sequence in a guarded YAML file and run ble run.
• Repeatable request/notification checks: use a YAML exchange action with separate write_characteristic and notify_characteristic fields. Assert the notification count/content, final notification, and cleanup.ok when the protocol has deterministic events.

Read workflows.md before creating or editing workflow YAML. Read safety.md before any write, pairing-sensitive operation, firmware update, lock, actuator, or other state-changing action.

timeout is a per-backend-operation bound, not a total command or tool deadline. Allow for device discovery, connection, profile discovery, each requested read, and any subscription duration when setting an outer Agent/tool timeout.

Capture files use Evidence Format v1. Read evidence-format-v1.md when an agent needs to validate, redact, or build replay tooling around a package. Read diff-format-v1.md before interpreting comparison policy, stable paths, ignored fields, identity guards, or CI exit behavior. Read replay-format-v1.md before interpreting missing evidence, captured failures, notification timing, replay MCP mode, or the read-only safety boundary. Read platform-acceptance.md before claiming platform support, running a cross-platform hardware acceptance, or publishing a real-device capture fixture.

Write policy

Treat every write as dangerous until the device protocol establishes otherwise.

• Require the user to authorize the specific state-changing operation.
• Require allow_write=true and confirm_device=<resolved identifier> for MCP writes.
• Require both --allow-write and --confirm-device <resolved identifier> for CLI writes.
• Never confirm with a friendly name, substring, stale identifier, or guessed address.
• Prefer write-with-response and read-back verification when supported.
• Prefer atomic exchange notification verification when a write triggers asynchronous events.
• Stop when the selected device is ambiguous or a prerequisite read/assertion fails.

Return the structured failure instead of bypassing a guard.