Release v5.2.1

[5.2.1] - 2026-11-06

Fixed — internal subagents launch in the Copilot harness and Copilot CLI

• Cause, read in the Copilot engine bundled with VS Code Insiders (@github/copilot
  1.0.73): the subagent tool builds its registry from custom agents filtered by
  disableModelInvocation !== true, so an agent with disable-model-invocation: true
  is not just unlisted but unlaunchable, even by exact name ("Unknown agent type").
  VS Code's local harness honours the caller's agents: list instead.
• AL Knowledge Consultant, AL Implementation, Planning and Code Review Subagents drop
  disable-model-invocation and keep user-invocable: false. As in the Claude Code
  plugin, where these agents are always listed, each contract now carries a caller
  guard (NOT_FOR_THIS_CALLER) and its description names its only callers.
• Architect and Spec still launch the consultant by its exact name, whether or not a
  list shows it. Copilot CLI, Foundation and Codex regenerated; Claude Code unchanged.
• Copilot CLI agents now pin claude-sonnet-5.5. claude-sonnet-4.6 left the CLI catalog,
  so every CLI delegation failed before the subagent ran ("Model 'claude-sonnet-4.6' is
  not available").
• Each guard checks a fixed first line, Caller: <agent name> (requirement <name>),
  which Conductor, Architect and Spec now write into every brief. Without it, the same
  bare request was accepted by one model and rejected by another. The line is a
  declaration, not authentication.
• The Copilot CLI Conductor entry repeats the caller line. In a CLI run the Conductor
  delegated from that short entry without reading its bundled contract, where the line
  lived, and the planning subagent correctly rejected the brief.
• The Copilot CLI README explains that plugin agents are namespaced
  (copilot --agent aldc-cli:al-architect; a bare al-architect is not found) and that
  VS Code's Copilot harness also loads installed CLI plugins.
• Observed on 5 Nov 2026 (clean Directions copy, before the caller line): in the Copilot
  harness the Architect launched AL Knowledge Consultant and its selection record passed
  every check; the Conductor launched AL Planning Subagent. In Copilot CLI 1.0.105 the
  Conductor launched aldc-cli:al-planning-subagent and the Architect launched
  aldc-cli:al-knowledge-consultant (claude-sonnet-5.5); the record passed every check.
  A default agent asking the consultant directly got NOT_FOR_THIS_CALLER.
• Observed on 6 Nov 2026 with the caller line (clean Directions copies): in the Copilot
  harness and in Copilot CLI 1.0.105, the Conductor and the Architect both wrote the line,
  their subagents ran, and each selection record passed every check. A default agent
  asking directly got NOT_FOR_THIS_CALLER on both hosts. An earlier CLI run outside the
  clean copy omitted question on one record entry; the clean runs did not repeat that.
• Not yet observed on a host: Spec launching the consultant, the Conductor launching its
  implementation and review subagents, and the local harness with these contracts.

Fixed — the knowledge consultant is launched by name, not by list

• VS Code host runs on 5 Nov 2026: in the local harness a disable-model-invocation
  subagent is not shown in the agent list but launches when called by its exact name
  (the Conductor launched AL Planning Subagent and got OK); in the Copilot harness the
  same call fails with "Unknown agent_type". Architect had consulted inline because the
  consultant was missing from the list.
• Architect, Spec and the design guidance now launch the consultant by its exact name
  (AL Knowledge Consultant, or al-knowledge-consultant where agents are addressed by
  id) and consult inline only after a failed launch attempt, recorded as the reason.
• Launching by name was not enough on the Copilot engine, which cannot launch a
  disable-model-invocation agent. Dropping the field, described above, fixes that.

Fixed — using the knowledge consultant is required when it can be launched

• Host runs on 2 Nov 2026 (Architect, one domain, clean Directions copy): Claude Code
  and Codex delegated; Copilot Chat delegated once and consulted inline once; the
  Copilot CLI saw the consultant but consulted inline "because it seemed simpler";
  in the Copilot harness of VS Code the consultant was not listed as launchable.
• Design guidance, Architect and Spec make delegation required whenever the
  consultant can be launched; consulting inline needs a stated reason.
• The consultant stays user-invocable: false, so it is not in the agent picker, and
  its caller guard keeps a knowledge-response from reaching an agent not prepared to
  validate it.

Fixed — knowledge consultation executor records what actually ran

• A Codex run of Architect launched the consultant for three of six domains and
  consulted the other three inline, yet recorded every response as
  executor: "subagent:al-knowledge-consultant".
  The design guidance, Architect, Spec and the consultant now state that the
  subagent executor applies only to a consultant actually launched in that run
  through the host's delegation tool; following its instructions yourself is
  main-session. A contract test pins the rule on Copilot and Codex surfaces.
• Claude Code surface 5.2.1 carries the same rule in its maintained guide, Architect,
  Spec and consultant. A single-domain Claude Code run (plugin 5.2.0) already
  launched the consultant and recorded it correctly; the rule makes it explicit.

Added — knowledge consultant on Copilot Chat, Copilot CLI and Codex

• New canonical read-only role al-knowledge-consultant (AL Knowledge Consultant):
  runs one BCQuality al-knowledge consultation for a verbatim question and returns
  the knowledge-response with how it ran. No AL tools on any surface; Codex profile
  sandbox_mode = "read-only".
• Architect and Spec consult each knowledge domain through it when the host can
  launch it, copy its response unchanged and validate it as the consumer
  (validation.by: consumer). Copilot Chat: agent tool restricted by
  agents: ['AL Knowledge Consultant']. Copilot CLI: Architect delegates through
  task; Spec has no task grant and consults inline. Codex: native delegation when
  available, otherwise inline (executor: main-session, unverified).
• The canonical design guidance records the question before invoking, the
  invocation and validation of each response, and labels candidate lists received
  in a brief as route: hook-candidates (manual).
• Spec review criteria carry applicability, unknown and source on every surface.
• Process evidence comes from Claude Code evals; host runs on Chat, CLI and Codex
  remain to be observed.

Changed — Claude Code surface 5.2.0

• BCQuality is the official microsoft/BCQuality plugin (1.9.0 or later), with
  al-code-review, al-knowledge and al-plan-guidance. The session warns when
  the installed BCQuality has no al-knowledge.
• New read-only subagent al-knowledge-consultant: runs exactly one
  bcquality:al-knowledge consultation for a verbatim question and returns the
  knowledge-response with how it ran. Architect and Spec delegate each knowledge
  domain to it from the main session. Evals on a clean lab base: 8 + 3 real
  consultations, question verbatim and response persisted in every case, no
  nonexistent reference.
• The design guidance records the question before invoking, the invocation
  (route, executor, plugin version) and who validated (consumer/self), and
  states that the Skill tool only loads instructions.
• agent-brief labels its article list "BCQuality candidates (ALDC keyword match;
  not a BCQuality consultation)", with provenance; used articles are
  route: hook-candidates (manual). The planning subagent no longer receives it.
• Review: selection-completed means ALDC selection dispositions completed with
  the BCQuality provider not executed; it is not provider coverage.
• Spec review criteria carry applicability, unknown and source.
• The session asks Claude not to end its turn while a needed subagent runs and, in
  pre-approved mode, recommends CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1
  (Claude Code 2.1.302 moved subagents to the background after about two minutes).

What's Changed

• Docs: novedades 5.0.3/5.0.4 y evolución propia de cada superficie by @javiarmesto in #131
• Port review contract to Copilot Chat, Copilot CLI and Codex (security, BCQuality articles, agent SDK) by @javiarmesto in #133
• Port design and planning contract to Copilot Chat, Copilot CLI and Codex (Standard Flow, BCQuality references) by @javiarmesto in #134
• Port Conductor contract to Copilot Chat, Copilot CLI and Codex (delegation, Mini-design, honest report, test runner) by @javiarmesto in #135
• Port Conductor autonomy modes to Copilot Chat, Copilot CLI and Codex (human by default) by @javiarmesto in #136
• Prepare ALDC 5.2.0 toolkit and Chat configuration defaults by @javiarmesto in #137
• docs: explain Project Manager configuration in four steps by @javiarmesto in #138
• Claude Code surface 5.2.0: real BCQuality knowledge consultations, honest labels by @javiarmesto in #140
• feat: knowledge consultant on Copilot Chat, Copilot CLI and Codex by @javiarmesto in #141
• fix: knowledge executor records the consultant only when actually launched by @javiarmesto in #142
• fix(claude): knowledge executor rule on the Claude Code surface (5.2.1) by @javiarmesto in #143
• fix(knowledge): using the knowledge consultant is required when it can be launched by @javiarmesto in #144
• fix(knowledge): launch the consultant by name, not by list by @javiarmesto in #145
• fix(agents): keep internal subagents launchable in the Copilot engine by @javiarmesto in #146
• docs(readme): correct the release and Claude Code versions by @javiarmesto in #148

Full Changelog: v5.1.1...v5.2.1