- Home
- Skills
- Agents & Orchestration
- Council Deliberation
Works with the AI Agents you already use
Council Deliberation
Council replaces that with a round table where 3-6 roles argue the question in separate contexts and cannot quietly agree.
$1.99
Council Deliberation
Example session with this skill installed
I need to decide whether to migrate our primary database from Postgres to a multi-region
distributed SQL database to fix cross-region write latency. I want the opposing side and the cost
people on the panel, not a set of yes-men.
- Read your context and instructions
- Compiled the council deliberation
R2 digest · Should we migrate the primary database from Postgres to a multi-region distributed SQL database to fix cross-region write latency? · 4/4 present
role stance change core claim
arch-migration conditional restated Run one shared phase-0 measurement gate; make region-partitioned primaries the default destination and allow distributed SQL only if phase 0 proves writes need multi-region commit.
rel-sre conditional restated Conditional: quorum floor regresses the median write, so go/no-go is error-budget arithmetic; approve only with measured baselines, pre-flip divergence gates, and named rollback triggers.
st-cost conditional restated Conditional: fund only phase 0 (~$25k) until it proves write locality and >= $8M/mo affected revenue; the full path costs ~$1.1-2.2M over 3 years vs ~$100k for region-local primaries.
st-critic conditional restated Not-yet-valid: all agree to measure first, but phase 0 cannot select between the three fixes and the two rollback claims contradict; fund only measurement.
open questions 13:
- @rel-sre Do you accept reverse-replication rollback (CDC from target back to Postgres) as satisfying the rollback gate, in place of application-level dual-write?
- @st-cost Does a phase 0 that produces the revenue-at-stake input, at zero dual-run cost, satisfy your requirement that the break-even number exist before approval?
- @st-critic If phase 0 delivers the network/commit/queue breakdown and the feature inventory, is the platform decision decidable, or is a proof-of-concept benchmark still a hard precondition?
- @arch-migration Publish the proof-of-concept's measured per-region p50/p99 write latency on the real transaction mix, next to the phase-0 Postgres baseline — the quorum-floor claim must become a number per region, not a protocol argument.
- @st-critic Do you accept a two-week production baseline (per-region p50/p95/p99 write latency split into network, queue, and commit time) as sufficient root-cause evidence, with the phase-0 gate as its single owner and deadline?
- @st-cost What sunk-cost cap would you set for reaching the phase-0 measurement gate, so that a migration cancelled at the gate is still net-positive versus never starting?
- @arch-migration What locality threshold does phase 0 use — the maximum allowed share of write transactions touching more than one region's data — and how many engineer-weeks does phase 0 itself consume?
- @rel-sre How many engineer-hours per month does the current Postgres on-call rotation actually consume, so the new store's staffing delta can be netted against freed hours instead of grossed?
- @arch-migration State the numeric locality threshold that cancels the project at phase 0, who signs it before phase 0 starts, and whether phase 0 will add the latency-component breakdown and same-base pricing of the alternatives.
- @arch-migration Is per-batch post-cutover rollback rehearsed at full production scale before any batch moves, and what reverse-replication lag ceiling makes it safe? Name the owner.
- @rel-sre Do you accept the migration plan's batched reverse-replication rollback as a genuine post-flip rollback path — yes or no?
- @rel-sre Define the degraded write mode under a region partition and confirm backup and point-in-time-recovery parity with today's Postgres guarantees.
- @st-cost Name the source for your infrastructure baseline and for the revenue-on-affected-paths figure; if none exists within this decision's scope, say so.
required items 11/12 · missing: Capacity levels
signals: theses restated 4/4 (new members 0) · stance labels changed 0 · open questions 13
(signals are inputs only; convergence is proposed by the facilitator and confirmed by the user)
Should we migrate the primary database from Postgres to a multi-region distributed SQL database to fix cross-region write latency?
2026-10-28 · mode decide · scenes architecture, reliability · panel arch-migration, rel-sre, st-cost, st-critic · 2 rounds
Answer
Not yet — and probably not to that destination. Approve one cheap measurement phase, not the migration. Fund roughly two engineer-weeks (about $25k) to produce three numbers: the share of write transactions that touch data owned by more than one region, the per-region split of write latency into network, commit and queue time, and the monthly revenue that flows through the latency-sensitive write paths. Those three numbers decide the question; nothing else in this plan does.
The reason to doubt the premise is arithmetic. A multi-region distributed database commits every write through a majority of replicas spread across regions, so each commit costs at least one inter-region round trip, typically 40–120 ms. That is the protocol, not a tuning knob. Writes near today's primary therefore get slower, and writes from far regions can only improve down to that same floor. If most of your writes are already local, you would be paying to make the median worse in exchange for a better tail.
If the measurement shows that only a small minority of writes genuinely must commit in more than one region, the answer is not distributed SQL but region-local primaries per tenant or data domain, with asynchronous replication for cross-region reads. That captures most of the latency fix for roughly $60–110k over three years, against $1.1–2.2M for the distributed path. If instead most writes are cross-region and revenue-bearing, distributed SQL becomes worth costing seriously — but even then it should be adopted second, not first.
One thing must be settled before any cutover plan is trusted: the panel contradicted itself on rollback. The migration plan claims per-batch rollback after cutover using hot reverse replication; the SRE entry states that once the source of truth flips, rollback is effectively one-way. Both cannot be true. Until a full-scale rollback rehearsal has been run and owned, treat the cutover as a one-way door and gate it accordingly.
Why
The chain that produces that answer has four links.
First, the question names a symptom and assumes a cause. "Cross-region write latency" can come from network distance between client and primary, from application fan-out across services, or from the database commit path. Distributed SQL only addresses the third, and it removes none of the first two: if clients are simply far from the primary, the distance remains and a consensus round trip is added on top. This is why the latency-component breakdown, not the locality share alone, is the number that lets a decision be made.
Second, the cost asymmetry is not marginal. The distributed path carries $450–900k of year-one cost and $300–600k per year after that, plus a permanent staffing floor of a quarter to half an engineer, plus what st-cost called the permanent developer tax: every future schema change and feature is built against a distributed planner, a lasting 10–20% premium on backend delivery. The region-local alternative costs one to two person-months and under a 20% infrastructure bump. A gap of one to two million dollars over three years should buy evidence before it buys a platform.
Third, the reversibility asymmetry points the same way. Moving ownership batch by batch between local primaries can move back. Crossing the quorum line cannot, if the SRE reading is correct. st-critic's objection is the strongest one raised in the session — that a gate which can only cancel and cannot select between the three candidate fixes is a formality, not a decision — and it stands. The answer to it is not to skip the gate but to specify it properly: phase 0 must also produce the latency breakdown and price all three options on the same measured base, so the gate selects as well as cancels. That is a requirement on phase 0, and it is cheap.
Fourth, everything above is conditional on numbers nobody in the room had. No production per-region latency breakdown, no published Postgres feature inventory, no revenue figure for the affected write paths. A panel without a business role cannot manufacture that last one; someone who owns the revenue has to sign it.
Evidence
| # | Claim | Source |
|---|---|---|
| 1 | A quorum commit floors every write at the inter-region round trip (40–120 ms); far-region writes cannot beat the floor, near-primary writes regress to it | R1 / rel-sre; R2 / rel-sre |
| 2 | The latency's cause is unmeasured: network distance, commit path and application fan-out are not separated, so the platform choice is undecidable | R1 / st-critic |
| 3 | Three-year cost: distributed SQL ≈ $1.1–2.2M; region-local primaries ≈ $60–110k; the difference is ~$1.0–2.1M | R2 / st-cost |
| 4 | Break-even needs the affected write path to carry ~$8–15M/month of revenue at a 0.5% conversion uplift from removing 100–300 ms | R2 / st-cost |
| 5 | Region-partitioned primaries have better rollback geometry than a quorum cutover: ownership moves one batch at a time and can move back | R2 / arch-migration |
| 6 | Post-cutover rollback is claimed to be per-batch reversible by the migration plan and effectively one-way by the SRE view; unresolved | R2 / st-critic |
| 7 | Quorum converts a regional outage into a possible global write outage; a globally-scoped 99.9% target is not the same budget as today's per-region one | R2 / rel-sre |
| 8 | Phase 0 was costed at ≤ ~$25k / ~2 engineer-weeks, with the parallel run capped at 12 weeks and ~$150k cumulative before proof | R2 / st-cost |
Conditions
- Phase 0 must deliver all three numbers before any schema work: multi-region-owned write share, the per-region latency component breakdown, and a revenue figure signed by someone who owns it. Missing any one, the project stops.
- Phase 0 must price all three destinations on the same measured base (distributed SQL, move the primary region, region-local primaries), otherwise the gate cannot select and the plan is not decision-grade.
- Any quorum cutover requires a full-scale rollback rehearsal with a named on-call owner, a reverse-replication lag ceiling (60 s proposed), and per-batch circuit breakers before the first batch moves.
- The SLO scope must be declared before comparison: today's budget is per region, a quorum system's is global. Re-baseline the error budget if the scope changes.
- Do not adopt a distributed store to solve a latency whose cause has not been located.
Preserved dissent
| Who | Objects to | Answered? |
|---|---|---|
| st-critic | The quiet convergence on "start the path, gate it with phase 0" is invalid: one measurement cannot both cancel and select; the two rollback claims contradict; the cost base is self-declared and its decisive number has no source; the source-of-truth flag has hidden writers | Partly. The gate specification now includes component breakdown and same-base pricing (defect 1 addressed). The rollback contradiction is recorded as unresolved rather than closed. Defects 3 and 4 stand |
| rel-sre | Rollback after the source-of-truth flip is effectively one-way, so the phase table overstates safety | Not answered by the migration plan; recorded as a blocking contradiction |
| arch-migration | Region-partitioned primaries should be the default destination, with distributed SQL only if phase 0 proves it | Adopted into the answer |
| st-cost | Only phase 0 should be funded (~$25k); the rest waits for proof | Adopted into the answer |
Unresolved
- Rollback: is per-batch post-cutover reversal real (rehearsed, owned, lag-bounded) or is the cutover a one-way door? The two entries disagree and no rehearsal has run.
- No production latency breakdown exists, so the root cause is still unknown; the locality threshold has no pre-agreed value or signatory.
- Postgres feature inventory (extensions, isolation levels, advisory locks, triggers, procedures) is unpublished, so portability is unquantified.
- Degraded write behaviour under a region partition, and backup/PITR parity, are undefined.
- Write-path inventory: which long-tail services, batch jobs and scheduled scripts bypass the source-of-truth flag.
- Revenue on the affected paths has no named source; the panel had no business role.
Revisit triggers
- Phase 0 shows the majority of write transactions touch data owned by more than one region, and those paths carry ≥ ~$8–15M/month of revenue: reconsider distributed SQL, with the gates above.
- Phase 0 shows the database accounts for a minority of the write latency: close the question, fix network placement or application fan-out instead.
- A full-scale rollback rehearsal passes at production scale with a named owner: the one-way-door objection is withdrawn and the phase table may be trusted.
- The Postgres feature inventory comes back with major gaps: portability cost rises and the balance tilts further toward region-local primaries.
- The globally-scoped error budget is re-baselined and shows the quorum floor already consumes a large share of it: stop before phase 1.
Connects securely to your tools. The creator never sees your data.
About this skill
The problem it solves
Most consequential decisions die in a loop: you ask an AI, it gives you a measured answer that smooths over the actual disagreement, and you act on it. Council replaces that with a round table where 3-6 roles argue the question in separate contexts and cannot quietly agree.
Instead of one ask and one answer, 3-6 roles with explicit stances take positions in separate contexts, then see each other's arguments and keep going until the session converges into a decision record.
What it produces
- One table per round — who stands where, their core claim, who they challenged, and whether they moved
- Required-item coverage — every scene (architecture, security, product, business, org, legal, research, personal decisions, ...) ships a checklist of what must be examined; uncovered items are named so four roles cannot all discuss the same thing
- A decision record (
conclusion.md) — the conclusion, every claim traced to a round and a role, the conditions it holds under, preserved dissent, what stayed unresolved, and the triggers that would reopen it
Why this is not "make the AI play several characters"
- Separate contexts: each role is its own subagent, so it never sees your chat history and cannot be anchored by your opening sentence. Swapping hats inside one prompt cannot do this.
- Auditable: every claim traces to a specific round and role, and dissent cannot be "synthesized away".
- Stateful: the session lives on disk in
.council/, so you can stop and resume it in another session. - Mechanical signals: attendance, coverage and whether a thesis was restated are computed by a script, not self-reported (measured: models self-reporting "I did not change my position" were wrong 3 out of 4 times).
Built-in role library
16 scenes x 4-5 specialist roles, plus 9 cross-scene stances (progressive, conservative, diagnostician, critic, cost skeptic, affected party, systemic risk, contrarian, synthesizer) — 73 roles total. The facilitator picks the scene itself; you do not have to.
Three modes
| Mode | For |
|---|---|
diverge | opening up a technical choice or a design space |
review | reviewing something that already exists (code, doc, architecture, PR) |
decide | producing a conditional call between options |
Three tone levels and multi-language output
--tone | For | How it talks |
|---|---|---|
professional | contracts, security, launches, finance | measured, conclusion first, every term explained on first use |
neutral | ordinary tradeoffs | direct and specific, no flattery, no confrontation |
casual | personal, curious, "shower thoughts" | everyday words, no invented terminology, no interrogation |
Output follows the language you write in (--lang en/zh/ja/... pins it). Under casual, the
adversarial roles are dropped from the panel by default: sometimes you want a different angle, not a debate.
It also keeps itself readable
Terminology and length snowball round over round (measured over 4 rounds: terminology-laden words +179%, body length +65%). So the script mechanically checks: a thesis over 40 chars, a body over 1000 chars, or more than 3 invented quoted names all raise a warning and the role is asked to compress and rewrite.
Requirements
python3 >= 3.7 (standard library only, nothing to install) and an agent that can run shell commands.
When not to use it
Questions with a factual answer; when you just want an action performed; when you are not actually
undecided. This skill costs panel size x rounds model calls — it is for decisions where being wrong
costs far more than the session.
Requires python3 >= 3.7 and a host agent that can run shell commands.
A session costs panel size x rounds model calls, so it is not for questions you can answer yourself.
Role independence depends on the host's subagent support; without subagents it degrades to sequential role-play and says so up front.
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before any payment.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 20 days ago
Requires
Permissions
Terminal/Shell: runs scripts/council.py for session bookkeeping and round tables. Read Files: reads the previous round's entries and digests under .council/. Write Files: writes meeting.json, entries/, digests/ and bundle.md under .council/<date>-<slug>/. Network Access: the skill's own script makes no network calls; participants may fetch reference material when the topic is a research or external-artifact question. Environment Variables: reads XDG_CONFIG_HOME / HOME to locate the optional user-level config for language and tone.