v0.2.0

Wake can now send what it collects somewhere else — and still says no by default.

Remote delivery

Records can be delivered to any OTLP/HTTP JSON collector, Langfuse included. The capability ships in every binary and stays off until you configure it:

wake remote set     # at a terminal: prompts for the URL, the public key, the secret key
wake remote on
wake remote flush

• wake remote status is the only command allowed to name the endpoint, and it prints the bare host. Neither set nor status ever echoes the secret key or the full URL; doctor reports presence only.
• What travels is one span per record: a trace per session, named after the repository, with each invocation parented onto its session. Spans carry structure, timing, model, tokens and outcome — never prompt or completion text. Built-in tool calls are counted locally and not sent as spans of their own.
• Nothing OTLP-shaped is persisted. The payload is computed from the local event store at flush time, held for one POST, then discarded. Every id derives from event_id, so a retried flush deduplicates at the receiver instead of double-counting.
• A wake ingest scan triggers a detached flush once delivery is on, throttled by remote.min_interval (15 minutes by default). wake remote flush --dry-run prints the exact payload the next flush would send, without sending it.
• WAKE_REMOTE_AUTHORIZATION overrides the stored credential, for CI or anyone who prefers no secret on disk.

The repository path never leaves the machine. The salted hash does, and so does the repository's readable label beside it, once you turn delivery on.

Consent a whole tree at once

wake init -g ~/Developer consents everything under a directory; wake init -g on its own means your home directory. Each repository under the boundary registers under its own identity the first time a scan sees a session in it — including repositories you create later — so report and serve keep their per-project breakdown rather than folding the tree into one. Collection starts forward-only from registration, the same default a plain wake init gives; wake init -g --full imports the existing history under the boundary in the same call. wake doctor says whether a boundary is set and counts the consented repositories it encloses.

Report

• A REPO column names the repository each primitive was used in — the readable label from your local project map, never the path. The unused list has no such column: a repository is a property of an invocation, and an unused primitive has none.
• An ERRORS column gives the failure count and its rate. Unknown outcomes are excluded from the denominator rather than counted as successes.

Correctness

• One session_end record derived per finished session (#47)
• Claude Code sessions resolve across every source that carries them, not just the first (#48)
• A subagent's health record derives from its own transcript (#49)
• A typed invocation is counted per occurrence (#50)
• parent_event_id gives delivered spans real parentage (#51)
• A throttled flush is reported as throttled instead of silently doing nothing (#40)

Install

curl -fsSL https://raw.githubusercontent.com/SupermodularAI/agents-wake/main/install.sh | sh

Already on v0.1.0: wake update. Archives ship with an SBOM, and the binaries carry GitHub build provenance — gh attestation verify ~/.local/bin/wake --repo SupermodularAI/agents-wake.

Full changelog: v0.1.0...v0.2.0

v0.1.0

Changelog

• f216c58 feat: MVP (#25)
• 9e6e3eb Initial commit