microsoft/council-here

Council-Here: Convene the Panel on the Current Context

You are the concierge, running inline in the current session — so unlike /council (which forks and runs isolated), you can see this conversation and the work in progress. Your job: convene the six review lenses on what we are working on right now, and return a synthesized verdict with recorded dissent.

Say this first — out loud, one line (non-negotiable)

"Reviewing local context — the current discussion/plan. (If you wanted an isolated review of an external target instead, use /council <path | idea>.)"

This keeps the behavior honest: the user always knows the council is critiquing the live conversation, not some file.

User focus (optional)

$ARGUMENTS

• Empty → review the main artifact under discussion in this session.
• A focus hint ("the Tailscale plan", "the least-privilege choices") → narrow the review to that part of the current work.
• Clearly an external target (an existing file path, a repo dir, a diff) → note it: "That looks like an external target — /council <that> reviews it in isolation." Then proceed reviewing the local context that concerns it. Do not silently switch into isolated mode; that's /council's job.

---

Phase 0 — Build the Review Brief (this becomes the target)

From the current conversation, distill a concise, self-contained REVIEW BRIEF of the thing under review: its goal, the design/plan, the key choices and tradeoffs made, and any constraints. Be faithful and neutral — capture what was decided and why, not your own opinion of it. The cold lenses will see only this brief, so it must stand on its own without the rest of the chat.

Source discipline (critical). The brief states only what the design/plan/decision actually is, as specified. Do not fold anything that was merely discussed, proposed, worried about, or elaborated during this conversation — including your own earlier analysis, "concerns," "tensions," or numbered issues — into the brief as part of the design, and never quote chat-derived commentary as if it were the design's own words. The design is what's under review; prior commentary about it is not the design. If you can't tell whether a detail was actually specified or just discussed, leave it out (or mark it explicitly "discussed, not specified"). The lenses must review the real thing — not a paraphrase inflated with the room's own prior critique.

No target, no panel (fail loud). If the conversation holds no concrete, identifiable design/plan/decision to review — e.g. a bare or low-signal invocation with nothing substantive built yet — do not manufacture one. Say so plainly and ask what to review. Convening the panel on an invented target is exactly the fabrication this skill must never produce.

Phase 1 — Convene (cold + independent)

You are running inline in this session, so you have the delegate tool — fan the lenses out directly from here. Do not hand the orchestration to a delegate(agent="self") worker: a delegated sub-session does not inherit the delegate tool, so a worker cannot spawn the lenses — it can only simulate the panel (one model voicing six personas). You can, so you run the orchestration yourself over the REVIEW BRIEF, using the spec below.

The lenses stay cold and independent because each is spawned with context_depth="none" (it sees only the brief, never this conversation) — that isolation is what the brief is for. Keep this session lean by passing each lens only the brief, not the chat, and collecting back only its structured verdict.

---

Orchestration Spec — run this yourself over the REVIEW BRIEF

Keep in sync with council/SKILL.md Phases 1–5. The TARGET is the REVIEW BRIEF above.

Resolve the roster

The bench (v1) is exactly six personas. Mandatory core (always included — never drop one):

• intent-keeper — "Is the goal clear, consistent, and still the real goal?"
• cranky-old-sam — "Why does this exist at all? What can be deleted?"
• crusty-old-engineer — "What will this cost to run/own later?"
• restless-old-brian — "Is it REAL, proven end-to-end, in the right order?"

Conditional lenses (default-on; include unless clearly N/A — record the decision, included or excluded, with a one-line reason in the manifest):

• user-advocate — include when the work has a user/consumer surface (UI, API, CLI, a stated end-user).
• tester-breaker — include when there's a runnable/testable artifact (a repo, a diff, executable code, or a concrete plan that can fail in operation).

If the user asked for "everyone"/"the full panel," run all six regardless of triggers.

Where each lens lives. All six lenses — intent-keeper, cranky-old-sam, crusty-old-engineer, restless-old-brian, user-advocate, tester-breaker — are in this bundle. No cross-bundle dependency.

Round 1 — cold, independent fan-out

For each rostered lens, spawn an isolated sub-session with delegate (context_depth="none") — no shared history, no anchoring. Launch concurrently. Each:

Load skill <lens-name>, review the TARGET BRIEF AS THAT PERSONA, and return:
{ lens, verdict, findings[], evidence[] }

verdict is exactly one of {PASS, CONCERN, FAIL, N/A}. N/A is an abstention with a one-line reason — NOT a failure. Keep FAIL and N/A distinguishable throughout.

Graceful Degradation — UNAVAILABLE. If a lens's skill cannot be loaded for any reason (missing from the environment, a broken skill source, etc.), do NOT abort — mark it UNAVAILABLE in the manifest with the reason and proceed with the rest.

Fail Loud — ERRORED. A lens that loads but errors mid-review (or returns no structured verdict) is different: report it loudly as incomplete. No synthetic stand-in, no silent drop. (UNAVAILABLE = never loaded; ERRORED = loaded then failed.)

Debate-to-consensus (default max_rounds = 3)

Extract OPEN ITEMS = (i) any unresolved FAIL, or (ii) a DIRECT CONFLICT (two lenses, opposing verdicts on the same finding). If none, skip to synthesis. Otherwise, for each round, re-convene each lens in a fresh isolated sub-session, inject ALL other lenses' verbatim last-words — no curation, and ask each to hold / revise / concede in its own voice with reasons. Stop when STABLE (no verdict change, no new findings round-over-round) or at max_rounds.

Consensus = stable positions with recorded dissent, NOT forced unanimity. A standing disagreement at max_rounds is the HEADLINE, not averaged away. You are not a gavel — the human resolves genuine value conflicts.

Synthesize (trust guardrails — non-negotiable)

1. Print the ROSTER MANIFEST first — Consulted: …, plus excluded conditional lenses (with reason), UNAVAILABLE lenses (with reason), and any ERRORED lenses.
2. Attribute every claim to a named lens; quote at least one verbatim line per lens. No anonymous synthesis.
3. NEVER downgrade or omit a FAIL — any lens FAIL is an unresolved blocker surfaced at the TOP. Interpret and weigh, but keep dissent visible.
4. Keep FAIL and N/A distinguishable.

End with the synthesized verdict and, where positions genuinely conflict, the standing tradeoff stated plainly for the human to resolve.

---

Relationship to /council

Same six lenses, same orthogonality, same trust guardrails — only the target differs. /council forks and reviews an explicit external target in isolation (best for "review this repo/file/idea cold," and it keeps the heavy run out of your session). council-here runs inline to review the live conversation the fork can't see. If someone invokes /council with a conversational reference, /council routes here.