Changelog

0.5.0 — 2026-10-11

• Rename the visible entry agent to concierge and the CLI invocation to copilot-concierge:concierge. There is no frontdesk agent alias; update launch commands and start a fresh session.
• Use models.concierge in launcher configuration while retaining models.frontdesk as a deprecated input alias. Reject both keys in the same file to avoid ambiguous settings.
• Add bounded Scout project notes: up to 12 sourced facts and 2 KiB in the assigned repository, checked against current evidence. Keep notes advisory, honor explicit no-write instructions, and stop note maintenance on unsafe paths or denied access.
• Add an optional, orchestrated Ship it specialist for GitHub draft PRs and GitLab draft MRs. Read-only preflight and separate confirmation of the exact publication proposal are required. No implicit commits, pushes, merges or non-draft creation. Provider access comes from existing authorized tools.
• Tighten specialist routing to discovered, explicitly pinned profiles; do not fall back to an unpinned built-in worker. Preserve the Scout note exception across delegation.
• Keep generated local project notes out of public source and release packages; check for accidentally tracked local state before release.
• Rewrite the README around installation and a first request, with an editable orchestration diagram and clearer configuration, migration and verification limits.
• Record release validation and live-trial scope in references/runtime-verification.md. Static instruction checks do not establish live provider draft creation or Windows/Linux model behavior.

0.4.0 — 2026-10-11

• Rename the maintained product to Concierge and the public repository, package, plugin and marketplace IDs to copilot-concierge.
• Breaking identifiers: start copilot-concierge:frontdesk, install copilot-concierge@copilot-concierge, use .copilot-concierge/config.json or ~/.config/copilot-concierge/config.json, and rename custom routing skills to copilot-concierge-routing (including prefixed variants).
• Rename the internal message markers to CONCIERGE-BRIEF, CONCIERGE-HANDOFF and CONCIERGE-REPORT; keep the existing models and role boundaries.
• Require explicit migration from the Zilot plugin. Users remove the old installation, install Concierge and manually copy chosen settings and update custom routing. There is no silent fallback to old configuration or routing.
• Preserve the v0.3.1 tag/release and its original Zilot IDs. The public repository remains the maintained native product; the former private development repository and history remain separate.
• Pass 39 source tests, validation, catalog synchronization and offline host catalog discovery for the new metadata and namespace. Custom routing checks now reject a plugin agent under the wrong manifest namespace; full YAML parsing remains outside the preflight checker.
• Live app/model workflows have not been replayed for 0.4.0. Historical Zilot 0.3.0 runtime observations remain under their original names and do not establish live acceptance of the new identifiers.

0.3.1 — 2026-10-11

• Accept LF and Windows CRLF Markdown frontmatter during validation and packaging; retain required-model and tool-boundary checks. Markdown files also use explicit LF Git attributes.
• Enumerate test files in Node so npm test also works with Windows shells that do not expand wildcards.
• Add the Concierge Copilot logo to the README and distributable package while retaining the native zilot plugin and marketplace IDs.
• Release the plugin, documentation and included logo under Apache-2.0, with LICENSE and NOTICE included in packaged copies.
• Prepare a plugin-only public distribution, with no prior private Git history, local sessions or legacy benchmark outputs.
• Keep the existing default agents and model policy. Known custom-preflight limitations remain: malformed YAML and a wrong plugin namespace can escape the checker; verify custom profiles in Copilot before use.
• Live runtime evidence remains the historical macOS 0.3.0 verification. CRLF regression checks do not establish Windows Copilot runtime acceptance.

0.3.0 — 2026-10-11

• Fixed the 0.2.0 scout failure. Copilot gives a nested agent only the tools its parent agent also has, so the control-only orchestrator left every specialist without file access. The orchestrator now carries the full tool set for its specialists and delegates by charter; a regression test guards this.
• The front desk runs the orchestrator as a background agent, routes follow-ups to it, asks questions and plan approval with Copilot's native question tool, and mirrors the orchestrator's task list into the root to-do list.
• Added a Sol 6.1 escalation role for explicit escalation; every role pins one model with no fallback list.
• Customers can add their own agents and skills in the places Copilot already reads and route Zilot's roles to them with a zilot-routing skill table; scripts/check-custom.mjs checks them, and examples/custom/ has templates. The launcher's extraDirectories loads extra folders with --add-dir or --plugin-dir.
• The front desk runs several orchestrators at once, recovers lost IDs with list_agents, and keeps every open objective in the root task list.
• Every report carries a short handover (request, follow-ups, decisions, done, open, sources). After 10 turns, a new request for a finished objective goes to a fresh <name>-2 orchestrator that continues from the handover, keeping orchestrator context small without hooks or token measurement.
• The front desk records a plan-approval preference only from the user's explicit words.
• Only the front desk appears in agent pickers; internal roles are user-invocable: false.
• Skills are self-contained, so workers stay inside the user's project.
• Removed the experimental extension adapter and its --native-config launch option.
• Live acceptance in Copilot CLI 1.0.93/1.0.95 and the GitHub Copilot app 1.1.28 is recorded in references/runtime-verification.md, with known limits in references/capability-matrix.md.

0.2.0 — 2026-10-08 (local preview)

• Zilot replacement plugin with new internals, planning, plan grilling, source scouting, enterprise UX, React, Node, GraphQL and independent review skills; legacy repository history is retained.
• Required-model role definitions for Luna frontdesk, Sol orchestration and Luna specialists.
• Session-local configuration with explicit model IDs, per-setting provenance and task/session/project/user approval precedence.
• Dependency-free configuration tests, package checks and reproducible local packaging.
• Experimental session-local native configuration adapter; observed computer-use and extension-management tools are excluded from the frontdesk role.
• Live ordinary CLI evidence confirms required native Luna → Sol → Luna routing and Sol todo updates. Nested source access remains blocked; actual app transcript rendering does not establish a working app workflow or root objective list.

Runtime and desktop acceptance results are recorded in the compatibility reference. A valid package is not proof of the complete three-tier runtime contract.