v0.3.0 — self-typed, actionable detector feedback

The feedback-detector now synthesizes lessons whose type is evident from the wording alone and that are actionable enough for a maintainer to fix the skill:

• technical / code — out-of-scope edits, wrong class/definition, missing typing, wrong API or structure.
• behavioral / flow — over-confirming, a skipped promised step, an ignored standing convention.

No schema or wire-contract change — the type is carried by the lesson prose, not a new field. The consent gate format is unchanged.

Changed

• Detector synthesis (Step 5) rewritten with per-type guidance + good/bad examples (SKILL.md, reference.md); /harness-feedback mirrors it.

Internal

• Eval suite scaled to 30 model-driven scenarios across four dimensions (precision, recall, redaction, and a new synthesis dimension graded by an LLM rubric judge), with streaming progress / Ctrl-C handling and a --save-outputs flag.

Packages

• @guidobuilds/loopback-setup → 0.3.0 (published to npm).
• The loopback service is unchanged and stays at 0.2.0 (no redeploy needed).

Install: npx @guidobuilds/loopback-setup

v0.2.0 — server-hosted MCP

The MCP moved server-side. The loopback MCP is no longer a client-side stdio bundle installed onto each machine — it is now hosted by the loopback service as a remote MCP, so MCP behaviour updates ship with the service and no client re-install is needed.

npx @guidobuilds/loopback-setup [agent] --service-url <your-service> --token <token>

Added

• Server-hosted MCP (service/app/mcp_server.py): FastAPI mounts a Streamable-HTTP MCP at /mcp with static Authorization: Bearer auth against the existing per-user tokens table. POST /feedback and the MCP submit_feedback tool share one ingest core (service/app/ingest.py).

Changed

• Installer registers a remote endpoint instead of a bundle — {type:"http"|"remote", url:"<service>/mcp", headers:{Authorization:"Bearer …"}} per harness; no server.bundle.js is downloaded.
• Redaction is done by the detector skill in-context; the service re-checks and quarantines leaks, returning the patterns for a one-shot re-redact + retry.
• Detector skill self-detects at every turn boundary with no hook/tripwire dependency.

Removed

• The client-side loopback/mcp + loopback/core packages and the shipped bundle. loopback is now just the feedback-detector skill + /harness-feedback command.
• Tools redact_preview, is_muted, mute_artifact, get_session_state, record_signal (the surface is a single submit_feedback), and per-machine muting.

Breaking

• Existing 0.1.x (stdio-bundle) installs must re-run the installer. The self-hosted service must run 0.2.0+ (it hosts the MCP endpoint).

Full changelog: CHANGELOG.md

v0.1.0 — @guidobuilds/loopback-setup

First public release of loopback on npm.

Install: npx @guidobuilds/loopback-setup [agent] (agents: claude-code | opencode | codex)

Highlights

• Ephemeral installer published as @guidobuilds/loopback-setup — wires the loopback MCP server, the feedback-detector skill, and the /harness-feedback command into your agent. No persistent CLI is left behind.
• Replaces the old persistent CLI; no hooks, no background turn-state.
• Wire contract: the client no longer generates a record id; the service assigns the canonical fb_<uuid> on ingest and echoes it back.

Notes

• Requires Node.js ≥ 18.
• Deploy the matching service/ (it assigns ids server-side) before the new client submits feedback.

Full notes in CHANGELOG.md.