vraj-ai Council
Council is a protocol run either by the selectable council primary for
standalone work or by the goals primary inside a goal run. What it needs from
the harness: seven members, each a non-spawning subagent with read-only
filesystem, skill, web, and MCP access, on models that differ from each other
and from the parent. Any provider that supplies those satisfies the protocol;
the slugs below are the shipped defaults, not a requirement, and the runner
overrides them per member through COUNCIL_<MEMBER>_MODEL:
council-grok:opencode-go/grok-4.5council-kimi:openrouter/moonshotai/kimi-k3council-qwen:openrouter/qwen/qwen3.8-maxcouncil-sol:openai/gpt-5.6-solcouncil-gemini:openrouter/google/gemini-3.6-flashcouncil-deepseek:openrouter/deepseek/deepseek-v4-flash-0731council-glm:opencode-go/glm-5.2
The active primary performs its own independent pass alongside all seven, making eight perspectives for contested research. Never pre-solve the question and ask members to ratify an answer.
When the configured roster is unreachable (no binary, no auth, an exhausted workspace balance), the parent session runs members through whatever headless CLIs the machine does have, instead of skipping review. Two are enough to keep maker separate from checker. For example:
grok -p "<prompt>" --output-format plain # council-grok
codex exec --model gpt-5.6-sol --sandbox read-only - # council-sol
Two things generalise from that example: a TUI-first CLI usually needs a plain or non-interactive output flag, or it never returns in a headless shell; and a long diff has to arrive on stdin rather than as an argument. Record the substitute in the review verdict so the run's provenance stays honest, and never let the fallback silently become no review at all.
Research rounds
Round 1: independent work
In one tool message, spawn all seven members concurrently with the full task
verbatim. Each independently reads real files, cites file:line, uses primary
web sources where relevant, cites URLs, and returns a complete evidence-backed
report. Goals performs its own pass before reading member reports. Do not merge
yet.
Round 2+: debate
Build a neutral conflict index containing only material claims on which reports disagree. In one tool message, invoke all members again. Give each member the other reports and conflict index, but not its own report. Require point-by-point agreement, rebuttal, or concession with evidence.
Rebuild the conflict index after each round. A point remains open only when members still disagree after seeing the same evidence. Stop when no new material disagreement remains or after three total rounds.
Vote and deliverable
If dissent remains at the cap, each member votes on each open point. Majority wins; goals casts only the evidence-based tie-breaking vote. Produce one clean converged report with conclusions, actions, and verdicts. Keep member framing, conflict indexes, and vote tallies in working notes, not the deliverable.
Sanctioned Phase B exception
For backlog sanity only, invoke exactly council-grok and council-kimi in
one parallel message for one independent round. No debate and no vote. Require
the fixed checklist:
COVERAGE / DUPLICATES / SCOPING / DEPENDENCIES / FEASIBILITY
Goals applies findings mechanically; additive evidence wins. At most one rerun. Never generalize this shortcut to contested research.
Review roles
T0 item review
Exactly two eligible non-maker members independently review one item diff. Review items separately, never as a batch. The CLI path runs the locked verification command in each disposable review worktree. In the in-session fallback, goals runs it mechanically and supplies the evidence to reviewers, whose shell and editing permissions remain denied. Each emits:
VERDICT: PASS | FAIL
FINDINGS: [{"severity":"P0|P1|P2|P3","file":"path:line","trigger":"...","wrong_behavior":"..."}]
FOLLOWUPS: []
Any evidenced P0/P1 fails closed. The reviewer never fixes the item. With the fixed GLM worker, default T0 reviewers are Grok and Kimi; GLM may research but must not judge GLM-authored code.
T1 milestone review
Use one rotating eligible member after mechanical checks pass. Scope is
integration only: whether merged items compose, cross-item coverage gaps, and
drift from the plan. Do not re-review item code quality. Emit PASS,
PASS-WITH-FOLLOWUPS, or FAIL.
Rotation is council-grok -> council-kimi -> council-qwen -> council-sol,
never the same member twice in a row and never the maker model.
Context ownership
Council is read-only with respect to durable context. It never edits
AGENTS.md, CONTEXT/architecture.md, CONTEXT/progress.md, a goal
handoff.md, backlog.jsonl, or any reviews/*.json. Review verdicts and
research reports hand back to goals — the sole writer of goal state — which
records them. When a member's findings imply a documentation change (a locked
decision shifts, a non-goal boundary moves, architecture.md should grow),
emit it as a finding with file:line evidence and the documentation impact;
never edit the artifacts yourself, never rewrite context broadly, and never
mutate the backlog. The public vskills repository's own machine-local notes
(CONTEXT.md, docs/, CONTEXT/, CLAUDE.md) are local-only under that
repository's own .gitignore and are not council artifacts; they are not a
universal rule a consumer project applies to its own tracked CONTEXT/, whose
architecture.md, progress.md, handoff, and review verdicts council reads
but never edits.
Anti-over-engineering guard
Pass this in every invocation and round:
Default to the lowest-complexity option meeting the strict requirements. Prefer existing stack components. Avoid resume-driven architecture, microservices, speculative extensibility, and heavy frameworks when the standard library suffices. If a heavy option is necessary, state exactly why simpler options fail and justify its operational tax.