city-hall-prep

name	city-hall-prep
description	Prepare all inputs for city-hall voice agents. Reads tracker, approvals, world summary, engine review, coverage ratings, previous log, canon, Mara directive. Writes pending decisions per voice.
version	1.11
updated	"2026-07-26T00:00:00.000Z"
tags	["civic","active"]
effort	high
disable-model-invocation	true
argument-hint	[cycle-number]

/city-hall-prep — Civic Government Preparation

Purpose

Assemble everything the voice agents need before city-hall runs. Each voice represents a neighborhood or city function and looks at different data. This skill gathers it all, routes the right data to the right voice, then Mags writes targeted pending decisions.

When this is done right, /city-hall runs clean — voices wake, decide, done.

Inputs

From disk (produced by earlier skills)

1. World summary — output/world_summary_c{XX}.md (from /build-world-summary)
2. Engine review — output/engine_review_c{XX}.md (from /engine-review) — ailments that need civic response, stuck initiatives, recurring crises
3. Previous cycle log — output/production_log_c{XX-1}.md §/city-hall section — what was promised last cycle, what cascaded. (Legacy fallback during transition: output/production_log_city_hall_c{XX-1}.md if the unified prior-cycle log is absent — pipeline.32 item (d); drop after 3+ clean cycles.)
4. Previous voice outputs — output/civic-voice/{office}_c{XX-1}.json — exact quotes, positions, tracker updates from last cycle per voice. Continuity source.
5. Mara directive — output/mara-directives/mara_directive_c{XX}.txt (if exists) — Mara's editorial pressure for this cycle

From sheets (via service account)

6. Initiative_Tracker — ImplementationPhase, MilestoneNotes, NextScheduledAction, NextActionCycle, AffectedNeighborhoods, PolicyDomain, Budget
7. Civic_Office_Ledger — all seats, approval ratings, district, faction
8. Edition_Coverage_Ratings — did the Tribune cover civic topics positively or negatively last cycle? Voices should know if the city noticed.

From MCP + Supermemory

9. GodWorld MCP — get_council_member(district) per voice for live approval + faction data. lookup_initiative(name) per active initiative. get_neighborhood(name) for affected neighborhoods.
10. Bay-tribune canon — search_canon(initiative_name) per active initiative. What has the Tribune published about each initiative? What was promised in coverage? What did citizens say? Grounds decisions in published history.

Voice Data Routing

Each voice looks at different data. Route the right inputs to the right pending decision.

Agent topology (G-10, S192 — 11 actual agents, NOT 17). Only Mayor / Chief / DA have individual agents. The 9 council members are grouped into 3 faction-bloc agents (opp-faction, crc-faction, ind-swing). One pending_decisions.md per agent — bloc agents speak for all members in their bloc within one file. The 5 project agents (Baylight / OARI / Stab Fund / Health Center / Transit Hub) are individual.

Agent (write pending_decisions to)	Speaks for	District/Scope	Data They Need
civic-office-mayor	Mayor Santana	Citywide	All initiatives, all approval ratings, engine review ailments, Mara directive, coverage feedback across all domains
civic-office-okoro	Deputy Mayor Brenda Okoro	Community Dev + Stab Fund + ED (Osei coverage)	Stab Fund processing trends + district equity + portfolio-load impact; speaks when Stab Fund anomaly, Mara directive names her, oversight checkpoint approaches, or Mayor cascade lands on her portfolios. Absence-of-statement is meaningful — don't force a statement (S215 civic.5).
civic-office-police-chief	Chief Montez	Public Safety	Crime metrics, OARI data, safety-related engine ailments, safety coverage ratings
civic-office-district-attorney	DA Dane	Legal framework	Only runs when legal dimension exists this cycle
civic-office-opp-faction	Rivers (D5) + Delgado (D3) + Carter (D1) + Mobley (D9)	OPP bloc — 4 council members	Per-member district data: D5 East Oakland (OARI + displacement), D3 Fruitvale (Transit Hub + OARI), D1 Jack London/W Oakland (Baylight + Stab Fund), D9 district state. Bloc-level political alignment + each member's approval + vulnerability flags
civic-office-crc-faction	Ashford (D7) + Crane (D6) + Chen (D8)	CRC bloc — 3 council members (Mobley D9 is OPP, not CRC)	D7 fiscal-oversight scope (Baylight audits, Transit Hub cost caps), D6 district state, D8 Lake Merritt process/environmental-review scope, each member's approval. CRC fiscal-conservative framing
civic-office-ind-swing	Vega (D4 Council President) + Tran (D2)	IND swing — 2 council members	Vega: Stab Fund oversight + procedural matters + Council Pres role. Tran: OARI expansion demand + D2 state (no active initiatives in D2). Each speaks for himself — not a bloc, no coordination
civic-office-baylight-authority	Director Keisha Ramos	Construction project	Baylight initiative state, construction milestones, workforce data
civic-project-oari	Director Vanessa Tran-Muñoz

Council canonical roster (per Civic_Office_Ledger): D1 Carter (OPP), D2 Tran (IND), D3 Delgado (OPP), D4 Vega (IND, Council Pres), D5 Rivers (OPP, Progressive Caucus Lead), D6 Crane (CRC), D7 Ashford (CRC), D8 Chen (CRC), D9 Mobley (OPP). Faction split: OPP 4 (Carter/Delgado/Rivers/Mobley), CRC 3 (Crane/Ashford/Chen), IND 2 (Vega/Tran). Chen D8 = CRC per truesource (output/desk-packets/truesource_reference.json + buildCivicVoicePackets.js FACTION_DISTRICTS + the CRC agent's own IDENTITY/LENS) — the earlier "corrected S195/S197 to OPP" annotation was itself the error and never matched the operational data; reverted S246 (G-PREP1).

Why the bloc topology matters (G-R11 from S193 city-hall run gap log): When an initiative reaches vote-ready phase with NextActionCycle = current cycle, the prep MUST route that initiative to the relevant faction-bloc agents to surface positions for ALL 9 council members — otherwise the vote can't tally and the project agent silently invents council positions (G-R6/R7/R10). Single-member routing (e.g., only Vega had Transit Hub on his desk in C93) is the structural cause of vote-not-trigger + fabrication failures.

Prerequisites (from /run-cycle)

Verify these exist before starting:

• output/world_summary_c{XX}.md — from /build-world-summary
• output/engine_review_c{XX}.md — from /engine-review

If either is missing, /run-cycle didn't complete. Don't proceed.

Memory Fence (Phase 40.6 Layer 2)

pending_decisions_*.md packets and voice briefings are consumed by voice agents — a downstream model context. Any content pulled from prior logs, search_canon, or Supermemory results that lands in a packet must be wrapped before the voice agent sees it.

const { wrap } = require('/root/GodWorld/lib/memoryFence');
const fencedCanon = wrap(canonExcerpt, 'bay-tribune');

Full convention: [[SUPERMEMORY]] §Memory Fence. Covers the threat model and when not to fence.

Steps

Step 0: Production Log

Create/open output/production_log_c{XX}.md — the unified one-true-cycle source every cycle-active skill appends a named section to (per docs/media/production_log_template.md). Write the §Cycle Header (timestamp, cycle, sim day) + §Carry-Forward, then append this skill's ## /city-hall-prep section with Mike's pressure. (Until the pipeline.35 cycle-init admin skill ships, /city-hall-prep is the log opener; afterward it becomes an appender.)

Mike's pressure auto-derive (S215, closes G-5). "Pressure from Mike" defaults to the synthesis of: engine review HIGH-severity ailments (output/engine_review_c{XX}.md §Ailments / §Stuck Initiatives) and Mara directive content (output/mara-directives/mara_directive_c{XX}.txt if present). Compute the default, write it into the production log, then ask Mike to confirm or override. Mike-override fires only when his read of the cycle differs from those engine signals (e.g., front-page-rotation pressure, audit follow-up ask, off-rubric editorial direction) — anomaly-only escalation, not always-on.

No ROLLOUT_PLAN involvement in this terminal (S256). City-hall prep neither reads from nor writes to docs/engine/ROLLOUT_PLAN.md. There is zero correlation between the civic cycle and the project ROLLOUT plan — ROLLOUT is cross-terminal build-tracking, not civic pressure. Issues surfaced during a run go to the cycle gap log (see §Gap log below), never ROLLOUT. (Per civic TERMINAL.md scope: this terminal runs skills, it does not file or read ROLLOUT entries. The prior "ROLLOUT §civic.* mechanical pull" step was removed — it contradicted the terminal rule and produced only engineering rows irrelevant to the cycle.)

Mara additive only (S215, closes G-9). Mara directive is a check-against frame to compare and reconcile, NOT a starting point to defer to. Mags-and-engine-review drive topic assignments; Mara's directive becomes a ## Mara cross-check block at the bottom of the production log header — "Did our topic assignments answer her questions? Where do they differ?" Don't treat the directive as primary input. This was Mike's explicit correction during S192 ("this is your space not hers"); without it written into the skill, the framing re-drifts every session.

Mara directive auto-derivation (S215, closes G-4). When Mara's directive is absent (some cycles don't get one), derive a default output/mara-directives/mara_directive_c{XX}_AUTO.txt from: (1) engine review HIGH-severity ailments, (2) prior-cycle Initiative_Tracker MilestoneNotes deltas, (3) prior-cycle voice JSON gaps (questions raised in voice statements that no agent answered), (4) C92-style per-voice "missing answer" framing. Voice prep then uses the AUTO directive as the Mara cross-check input; Mara's manual version overrides if she files one later. This converts an optional input into an always-on input without requiring claude.ai turnaround on every cycle.

Mara directive Drive discovery (S229 G-PREP3). Mara files her directives manually through claude.ai → Drive; the disk-side path is not always populated even when she filed one. Before AUTO-derivation, check Drive for a manually-filed directive:

1. Check Drive folder mara-directives/ (or the cycle's Drive folder if a per-cycle convention is set in docs/media/DRIVE_MANIFEST.md — verify path on each cycle until convention regularizes) for mara_directive_c{XX}.{txt,md}.
2. If found, download via node scripts/downloadDriveFile.js <fileId> output/mara-directives/mara_directive_c{XX}.txt.
3. If absent on Drive AND absent on disk, fall back to AUTO-derivation per S215 G-4 with explicit "no manual Mara directive this cycle" warning logged to the production log header.

This converts "operator-side institutional knowledge (Mike hands me the Drive file ID via chat)" into a skill-level step. The Drive folder ID for the mara-directives/ folder is the canonical pointer; until that ID is hard-coded here or the folder convention regularizes (Mara writes directly to disk via service-account), use Drive search or fall back to AUTO.

Mara ESCALATION override (S229 G-PREP8). ESCALATION-tagged Mara directives override absence-of-statement defaults. If Mara filed an ESCALATION directive on a voice agent who is normally treated as absence-meaningful (Okoro per civic.5 — "absence-of-statement is meaningful, don't force a statement"), the voice MUST be assigned a topic. ESCALATION signals Mara already gave that voice one cycle of pass; a second cycle of pass would be canonical drift. Tag the voice's pending_decisions packet with MARA ESCALATION — voice must respond so the agent's identity layer sees the override at execution time. Default behavior is override; if Mara explicitly tags ESCALATION — absence acceptable (escape clause), the absence-of-statement default holds.

Engine-primacy carve-out on the override (S258 RB-4, closes G-PREP9). ESCALATION overrides the absence-of-statement default, but it does NOT override engine-primacy: a Mara ESCALATION cannot force a voice to speak to a subject that is absent from the initiative tracker. Forcing it would manufacture civic source material for a program the engine doesn't carry — the contamination loop (edition fabrication → civic canon). C97 case: Mara escalated Okoro/Webb/Ramos on the $14M Retail Recovery (not in the tracker) and a KONO blackout (a prior edition story with no live engine signal) — both correctly dropped, the voices not forced. Rule: if the ESCALATION subject has no tracker row AND no live engine signal this cycle, engine-primacy wins — do not assign the topic, log ESCALATION declined — subject not in tracker/engine (engine-primacy).

Step 1: Read All Inputs

Read all 10 inputs above. For each:

Disk (PRIMARY — G-13, S192). Read in this order, treat as authoritative:

1. output/world_summary_c{XX}.md — snapshots Civic_Office_Ledger (approval ratings + factions) + Initiative_Tracker (phase + MilestoneNotes) post-cycle. This is the canonical pre-civic state input.
2. output/engine_review_c{XX}.md — derives from Initiative_Tracker; surfaces ailments + remedy-firing patterns.
3. Previous output/production_log_c{XX-1}.md — last cycle's voice outputs + tracker updates + dramatic moments. Graceful fallback (S225 pipeline.23, closes G-PREP2): if production_log_c{XX-1}.md is missing (early-life cycles, manual intervention, or a /post-publish run that didn't finalize), fall back in this order: (a) split-form historical output/production_log_edition_c{XX-1}.md from prior pre-S195 cycles; (b) output/production_log_city_hall_c{XX-1}.md (civic-side only — partial context, but better than zero); (c) document the absence in the production log §Step 1 entry and continue without prior-cycle context. Don't silently no-op — surface the missing baseline.
4. Mara directive — the cycle's editorial pressure (citizen accountability questions).
5. Prior-cycle published canon (G-15, S192). editions/cycle_pulse_*_c{XX-1}_*.txt + output/reporters/*/articles/c{XX-1}_*.md — interviews, dispatches, supplementals from prior cycle. Cross-reference each against active topic-assignment initiatives. If interview/dispatch text mentions an active initiative, the voice agents owning that initiative MUST see the canon excerpt in their pending_decisions.md (memory-fenced). Without this, prep ships stale framing — C92's Mayor interview answered 6 OARI/admin questions that voices would otherwise re-litigate.
6. Baseline briefs (S256 RB-3, feeds the City This Cycle digest). output/baseline_briefs_c{XX}.json — .briefs[] carries this cycle's synthesized world-events / citizen-life-events / initiative-milestones, each tagged .neighborhood with .facts.neighborhoodState deltas (sentiment / retailVitality / eventAttractiveness / crimeIndex). This is the living-city texture the engine surfaced THIS cycle that Mara's directive is structurally blind to. Read alongside world_summary; it is the per-voice neighborhood-pulse source for Step 3.

Sheets (VERIFICATION — drop unless world-summary is stale). Sheet reads of Initiative_Tracker / Civic_Office_Ledger / Edition_Coverage_Ratings are redundant if world summary is fresh — every cell that matters is already in the disk inputs above. Run sheet reads ONLY when world_summary mtime is older than the city-hall start time, or when verifying a specific cell. Routine prep reads disk first, sheet second-and-rarely.

MCP (LOOKUP). Run get_council_member for each district with an active voice — returns live approval/faction (cross-checks world_summary). Run lookup_initiative for each active initiative — returns Initiative_Tracker row with MilestoneNotes (cross-checks world_summary phase). Run get_neighborhood / get_neighborhood_state for affected neighborhoods (S183 wd-neighborhood layer; the latter is narrower). Full tool inventory: [[../../../docs/SUPERMEMORY|SUPERMEMORY]] §Search/save matrix.

Canon. Run search_canon per active initiative to find what the Tribune has published — promises made, citizen reactions, coverage tone.

Delegate multi-file digs to source-search (S326, Mike-direct). Exact MCP lookups stay direct. But cross-file verification work — reconciling a stale digest against tracker + production logs, tracing an initiative arc across prior-cycle summaries, resolving two sources that disagree on scope/phase — spawns the source-search agent (Haiku-pinned retrieval-only; Sonnet-parity, 0 fabrications on the C100 eval, which was itself a civic-arc dig on the OARI-miss cycle). It flags digest staleness (a c{XX} file whose "Latest" is c{XX-1} data) and reports conflicts both-ways with citations — read its ruling before packet content rides on the fact.

When the question requires a multi-Edition initiative arc rather than a single locator hit, dispatch source-search with retrievalLane=prior-published-arc. It may call only node scripts/notebooklmCanonSearch.js in default published scope. Current initiative phase, vote state, district scope, metrics, and office ownership still come from current disk/Sheet/MCP sources and win every conflict. Direct search_canon stays the first-hop publication locator; do not use Richmond Archive references for normal civic initiative research.

Log lane, selected/used source IDs, citation count, and reconcile verdict in the production log. Do not log the NotebookLM answer body.

Gate the return before packet content rides on it (S334). Admissible only if the first line is retrievalLane: prior-published-arc, every claim carries a NotebookLM source ID (uuid) + citation number + excerpt, and the last line is a bare reconcileVerdict token. A claim sourced to a file path instead of a source ID means the agent read the edition artifacts directly and escaped the reviewed source scope — re-dispatch once naming the violation, and on a second failure drop the seat's material and prep the packet without prior-arc sourcing. Never put a file-path return into a pending-decisions packet.

Log tracker state, approval ratings, and key findings in the production log.

Council Roster Reconciliation (BUNDLE-PREP-A, S246 — closes the class that produced G-PREP1). The MCP get_council_member call above returns each district's live faction from truesource; that signal is wasted unless it's reconciled against this skill's static roster. The Chen D8 mis-routing (a stale "S195 correction" that flipped her CRC→OPP in the skill text + 8 agent roster tables while truesource kept her CRC) survived multiple cycles precisely because nobody compared the two. Run the reconciliation every prep:

# Truesource faction per district (the authority — the same source the MCP wraps):
node -e "require('./output/desk-packets/truesource_reference.json').council.forEach(m=>console.log(m.district+': '+m.faction+' ('+m.name+')'))"

(In-session, the mcp__godworld__get_council_member tool returns the same per-district faction — either source works; both read truesource. scripts/godworld-mcp.py is an MCP stdio server, not a CLI, so don't shell out to it.)

get_council_member also returns each seat's status (active / recovering / vacant). Reconcile status too: a recovering member (D6 Crane this window) must be carried in the 9-seat roster as a NAMED absentee with the active-voter denominator reduced — world_summary historically drops non-active seats, so verify the recovering/vacant seats against truesource here and feed them into every vote-math tally. Full enum + vote-math mapping: [[../../../.claude/rules/civic.md|civic.md]] §Council member status enum.

Compare each district's truesource faction against (a) the Council canonical roster in §Voice Data Routing above, and (b) the per-district roster tables in the civic-office agent RULES.md files. Any mismatch is a HIGH anomaly — surface it to Mike and do NOT proceed past Step 1 until the static roster is corrected to match truesource (truesource_reference.json + buildCivicVoicePackets.js FACTION_DISTRICTS are the operational source; static skill/agent text is the thing that drifts). Truesource wins; the text gets fixed forward. Faction membership is load-bearing for the entire cascade — a mis-routed member means the wrong faction agent speaks for them and vote-math attribution is wrong.

World summary stale-civic-decisions framing (S215, closes G-7). The world summary's §Civic Decisions section can read stale if the current cycle's /city-hall hasn't run yet — it presents the prior-cycle cascade as if locked, with a small disclaimer reader has to notice. When ingesting world_summary at Step 1, treat any §Civic Decisions data as "Last Locked (C{XX-1})" until the current cycle produces new voice JSONs. Don't propagate stale framing into pending_decisions packets — voice agents must see explicit "C{XX-1} canon, C{XX} not yet decided" labeling. Engine-side fix is filed at pipeline.14 (world_summary auto-rebuild after /city-hall); until that ships, the disclaimer-respect rule lives here.

Auto-investigate engine-flagged initiatives (S216, closes civic.11). Engine review can false-flag a phase-advanced initiative as mitigator-stuck or remedy-not-firing — civic.7's INIT-005 C93 case was a Scenario C engine-auditor bug where cyclesInPhase walked priors, found a phase mismatch, and triggered the cold-start fallback. Before propagating any engine-review ailment of class mitigator-stuck or remedy-not-firing into topic assignments, run the MilestoneNotes reader for each affected initiative:

node scripts/readInitiativeMilestoneNotes.js <INIT-ID> {XX}

Capture the output to the production log. Three dispositions:

Read the LATEST entry, not the current cycle's (G-PREP1 timing fix, S256). This auto-investigate runs at prep — before /city-hall Step 6 writes this cycle's MilestoneNote — so a C{XX} entry structurally cannot exist yet. The disambiguation question is "has the initiative shown recent documented progress that the engine flag missed?" — answered by the most recent MilestoneNotes entry present (normally C{XX-1}, written by last cycle's city-hall). Checking literally for C{XX} always finds none and mis-fires every initiative into Scenario A. Read the latest stamped entry instead. (If a C{XX} entry somehow already exists — a stopgap/manual write — it is the latest, and this still works.)

• The latest MilestoneNotes entry names concrete recent progress (advanced phase / completed deliverable) that the engine flag contradicts → engine auditor likely false-flagged (Scenario C — engine lag/cold-start walked priors). Drop the ailment from the topic-assignment surface; note "engine auditor false positive verified via MilestoneNotes" in the production log. Don't burn voice cycles re-litigating phase advances that happened.
• The latest entry predates the flag or shows no recent progress (the initiative genuinely hasn't moved) → real signal (Scenario A — commitment slipped, or B — writeback bug). Surface the highlighted history to the owning voice's topic assignment with the no-recent-progress note ("no milestone since C{XX-N} — voice owes a commitment-status update").
• The latest entry documents a reason that contradicts the engine flag (e.g., "delayed pending council approval") → real but reframed. Surface to topic assignment with the contradicting note attached so the voice agent sees the documented reason and can respond against it.
• Scenario D — the paradox: remedy-overshot AND phase-duration-stuck (G-PREP3). The engine flags an initiative phase-duration-stuck while its remedy is simultaneously overshooting — the program is demonstrably working but hasn't advanced phase. C95 OARI is the canonical case: ViolentCrimeIndex -1 observed vs -0.05 expected (remedy fired ~20× target) yet 13 cycles in pilot_evaluation. This is NOT a stuck-and-failing story and must not generate manufactured stuck-drama — the engine-side fix (ES-4, bc3b893 — vote-passed initiatives exit stuck-drama detection) removes most of these at source, but when the paradox still surfaces, route it to the Mayor with the paradox named and one of three skill-canonical public frames as the decision options:
  1. Success-declare / scale — the overshoot proves the pilot worked; advance phase on the strength of results (expand scope, move to next stage).
  2. Calibrate-and-hold — the overshoot signals the intervention is over-sized for the target; right-size before advancing. The held phase is deliberate diligence, not failure.
  3. Steady-state transition — reframe from active-pilot to ongoing-program; the phase isn't "stuck," it matured into operations and the duration flag is a category error. Surface all three as the Mayor's pending-decision options (real consequences each); the faction blocs then react to the Mayor's chosen frame per the cascade. Do not let the auto-investigate drop the ailment silently (it's real signal) nor propagate it as a stuck-crisis (it's a success the engine mislabeled).

Runs at Step 1, before Step 2 builds topic assignments. The auto-investigation prevents a cycle of voice work being wasted on phantom-stuck initiatives like INIT-005 was in C93. Plan + Scenario C trace: [[archive/plans/2026-05-11-civic-7-init-005-investigation]].

Anomaly-only present-to-Mike gate (S215, closes G-6 Step 1 side). Previous convention: always present Step 1 input summary to Mike. New convention: compute input completeness automatically (all expected files present, all approval ratings reasonable, all initiatives accounted for) and surface to Mike ONLY on anomaly:

• Council roster reconciliation mismatch — any district whose truesource faction differs from the static roster (BUNDLE-PREP-A; HIGH, blocks past Step 1)
• Approval shift >5pts on any council member from prior cycle
• HIGH-severity engine ailment unaddressed (no topic assignment in §Topic Assignments for it)
• New initiative phase change since prior cycle (vote-ready, design-active, etc.)
• Missing C{XX-1} voice output for an agent expected to speak this cycle
• World summary mtime older than city-hall start (G-7 staleness condition)

If none of those fire, Step 1 completes silently with one-line "no anomalies — proceeding to Step 2" log entry. Operator escalates to Mike's review only on actual anomaly.

Step 1.5: Verify the ledger snapshot is fresh (S252, reworked S329)

The snapshot refresh now runs on the cycle path (run-cycle Step 5.56) — this step verifies it instead of writing it. Read output/simulation_ledger_snapshot.meta.json and check "cycle" equals {XX}. If it matches, log the actual meta line (cycle + rowCount + generatedAt). If it does NOT match, run node scripts/dumpLedger.js {XX} --quiet, then read the meta back and log what it says.

Never log this step as "refreshed" from a command exit or a remembered count — the C101 run (S329 finding) logged "913-row snapshot refreshed" while the file sat untouched since Jul 10, and every Discord persona went blind to ten days of canon. The log line for this step must be a quote of the freshly-read meta, nothing else.

Step 2: Topic Assignments

Based on tracker state + Mike's pressure + engine review ailments + Mara directive, determine:

• Which voices have decisions this cycle
• Which initiative/topic each voice owns (atomic checkout — no overlap)
• Cascade order (Mayor always first, then who reacts to Mayor)
• Engine review ailments auto-assign to relevant voices (Temescal health → Health Center + whoever owns health policy)

Log assignments in the production log:

## Topic Assignments
- OARI: Mayor (political) → Montez (operational) → Tran-Muñoz (project)
- Stabilization Fund: Mayor (enforcement) → Vega (oversight)
- [ENGINE AILMENT] Temescal health 4-cycle: Health Center (Chen-Ramirez) + Mayor

Step 3: Write Pending Decisions

For each voice with a decision, write output/civic-voice-workspace/{office}/current/pending_decisions.md.

City This Cycle digest — REQUIRED in every packet (S256 RB-3, closes C96 G-PREP3). Mara's directive is initiative-centric and structurally blind to the live world_summary — following it alone makes the cascade an initiative-checklist and strips the living-city texture that is the product ("this is a sim, not a civic-initiative sim" — Mike, S251). Independent of Mara, inject a ## City This Cycle block at the top of every pending_decisions packet, derived from the disk inputs (NOT from the Mara directive):

• Mayor (and any citywide voice — Chief, DA): the citywide digest. From world_summary_c{XX}.md — pull §City State, §Three-Cycle Trends (C{XX-2}–C{XX}), §Evening Texture, §World Events, and the medium/high §Engine Review Findings. 3–6 lines: what the city feels like this cycle, not what an initiative is doing.
• Every neighborhood/district voice (faction blocs + project directors): the neighborhood pulse. From the §Neighborhood snapshot table (Sentiment / RetailVitality / EventAttractiveness / CrimeIndex), filtered to that voice's district neighborhood(s) per §Voice Data Routing, PLUS baseline_briefs_c{XX}.json .briefs[] filtered by .neighborhood — the synthesized world-events / life-events with neighborhoodState deltas for that voice's turf. A faction bloc gets one pulse per member district.
• Scope discipline: this is texture the voice may speak to, not a manufactured crisis. Carry the engine's actual signal — a synthesized "crisis-spike" non-event (the C96 Rockridge/Piedmont case) is reported as the modest delta it is, not inflated into drama. The digest gives the voice a living city to react to; it does not assign a decision (that's the topic assignment).

Live roster-status block — REQUIRED in every faction-bloc packet (RB-2, C98 G-R6). A faction bloc's whip-read (who's voting which way, what the count is) must run off THIS cycle's live seat status, never a remembered one. Inject into each bloc packet (opp-faction / crc-faction / ind-swing) the current active / recovering / vacant status of every seat the bloc speaks for, taken from the Step-1 get_council_member reconciliation (the authority), per [[../../../.claude/rules/civic.md|civic.md]] §Council member status enum:

• State each member's seat status explicitly — D6 Crane — ACTIVE this cycle or D6 Crane — RECOVERING (named absentee, not voting). Status changes between cycles; a member who was recovering last cycle may be active now, and the packet must say so.
• State the active-voter denominator across the full 9-seat council so the whip-read computes the majority off live data — "9 active → majority 5," never a remembered "8 active, one absent."
• C98: the OPP whip-read carried a stale "Crane absent, 5 of 8" into its statement when Crane (D6) had returned to ACTIVE and the real tally was 9-0. The status was live in truesource at Step 1 but never reached the bloc packet, so the agent reasoned off memory.

This pairs with the Step-1 status reconciliation: Step 1 checks status against truesource; Step 3 delivers it to the voice. Without the delivery, the check protects vote-math at the Clerk layer but not the bloc agent's own whip-read prose.

Each pending decision includes:

• The situation in plain language
• 2-3 predefined options with real consequences
• An open option ("Your call. Make your own move.")
• Who's watching
• Their current approval rating and vulnerability
• Their district's neighborhoods and how initiatives affect them
• Coverage feedback — did the Tribune cover their domain positively or negatively
• Engine review context — is an ailment in their domain recurring or worsening
• What they said last cycle (from previous voice output JSON) — continuity
• What the Tribune published about their initiatives (from canon search) — public record
• Mara directive pressure if relevant
• "No decision is not an option this cycle."

Each pending decision does NOT include:

• Citizen lists
• Raw neighborhood demographics
• Full crime data dumps
• Full initiative packets
• Anything from buildInitiativePackets.js
• A ## MAYOR'S DECISIONS THIS CYCLE block (S258 RB-4, closes G-PREP8). Prep must NOT pre-script the Mayor's call. That block is injected at runtime by /city-hall Step 3 after the Mayor agent actually runs — pre-scripting it at prep makes reacting voices rubber-stamp a decision the Mayor's own agent never made (C97: first-pass Chief/CRC/IND packets carried her INIT-007 calls as settled). Reacting packets give the live situation only; the voices hear the Mayor's real output at runtime.

Engine→in-world translation contract — NO raw telemetry in any packet (S258 RB-3, closes G-PREP5). A pending_decisions packet is the voice's memory; engine values are foreign objects in it. Translate every engine signal to perception before it enters the packet — never carry it raw:

• Forbidden in any packet: metrics, σ values (−4.9σ), decimals (+0.84), backtick phase-codes (`vote-ready → construction-planning`), and (engine) / Engine: tags. The C97 first-pass OPP packet leaked all four.
• Express engine state as perception only: "violent crime is down sharply this cycle," not violent crime −4.9σ; "the mood downtown has lifted," not sentiment +0.84; "the project just cleared its planning vote and is moving to construction," not the phase-code arrow.
• The deterministic packet linter that enforces this at the Step-4 gate is ES-2 (engine-sheet) — the mechanical backstop, because the operator is the contamination source. This rule is what that linter checks.

Step 4: Verify Prep Outputs (anomaly-only gate, S215 G-6)

Check all files exist:

• Production log created with tracker state + approvals + topic assignments
• Pending decisions written for each assigned voice
• Previous cycle context loaded and referenced
• Any pending_decisions referencing a council member matches the canonical 9-member roster (D1-D9 per Civic_Office_Ledger)
• Vote-trigger detection: any initiative at vote-ready phase with NextActionCycle = current cycle is routed to all 9 council voices via faction-bloc agents (G-R11 pre-condition)

Deterministic telemetry linter — HARD GATE before voice dispatch (ES-2 step 2 backstop, wired S259). The Step-3 translation contract is the rule; scripts/lintCivicPackets.js is the mechanical enforcer (the operator is the contamination source, so instruction alone can't be the gate). Run it over every written packet:

for f in output/civic-voice-workspace/*/current/pending_decisions.md; do
  node scripts/lintCivicPackets.js --file "$f" || echo "TELEMETRY LEAK — $f (HALT)"
done

It flags σ / signed sub-1 deltas / phase-codes / Engine: tags / backtick code-spans / raw decimals next to a metric word. Any non-zero exit is a HARD anomaly — translate the flagged values to perception per the Step-3 contract and re-run until every packet exits 0 before /city-hall dispatches the voices. (Wired via --file per-packet: the script's bare-<cycle> glob targets top-level output/pending_decisions*.md, which is NOT where prep writes packets — flagged back to engine-sheet to point resolveFiles at output/civic-voice-workspace/*/current/. Until that lands, the --file loop is authoritative.)

Anomaly-only present-to-Mike (S215, closes G-6 Step 4 side). City Clerk pre-flight pass verifies voice files exist for assigned voices, content matches assignments, no canon violations. If Clerk passes clean, Step 4 completes silently with one-line "prep verified — running /city-hall" log entry. Mike's review fires only when Clerk flags a structural issue:

• Pending_decisions packet references a non-canonical council member (e.g., wrong district)
• Vote-ready initiative not routed to faction-bloc agents
• Missing pending_decisions for a voice expected to speak (per topic assignments)
• Approval ratings inconsistent between disk inputs and pending_decisions packet
• Quarantine subdirs accidentally re-included in canon scope (G-R12 — see city-clerk RULES.md)

Default outcome: silent pass. Mike-review outcome: anomaly named + suggested fix proposed before /city-hall runs.

Output Files

File	Purpose
output/production_log_c{XX}.md	Opened — §Cycle Header + §Carry-Forward + ## /city-hall-prep section (Steps 0-2 complete, tracker state, approvals, topic assignments)
output/civic-voice-workspace/{office}/current/pending_decisions.md	One per voice with decisions this cycle

Handoff to /city-hall

When this skill completes, /city-hall picks up by reading:

File	What city-hall does with it
output/production_log_c{XX}.md	Appends its ## /city-hall section to this log — voice decisions, tracker updates, media handoff
output/civic-voice-workspace/{office}/current/pending_decisions.md	Each voice agent reads ONLY this file + their IDENTITY.md. Nothing else.

/city-hall does NOT re-read sheets, MCP, or Supermemory. Everything it needs is in the files this skill produced. If the prep is right, city-hall is mechanical.

What This Skill Does NOT Do

• Launch voice agents — that's /city-hall
• Apply tracker updates — that's /city-hall
• Run project agents — that's /city-hall
• Run the City Clerk — that's /city-hall
• Pick edition stories — that's /sift

Where This Sits

After /run-cycle (which produces world summary and engine review). Before /city-hall.

Sheet Access

Service account via lib/sheets.js. Spreadsheet ID from .env. GodWorld MCP for structured lookups. Supermemory bay-tribune for canon search.

Gap log (S212 — see [[plans/GAP_LOG_TEMPLATE]])

At skill close, capture friction observed during prep as a gap log. /city-hall-prep is a heavy skill at the civic generator terminal; sidecar gap logs catch inefficiency the skill couldn't catch while running.

Destination (RB-1/RB-2 — one-true gap log): append a leg to the cycle's single gap log output/production_log_run_cycle_c{XX}_gaps.md (the file the engine cycle audit opens each cycle). Do not write a separate _city_hall_prep_gaps.md sidecar — that split convention is retired. Open the leg with the fixed header the gate greps for:

## LEG: /city-hall-prep (G-PREP)

Then the G-PREP entries below it — or No gaps this run. on a clean run. The header must be present either way.

Gap prefix: G-PREP* (e.g., G-PREP1, G-PREP15. /city-hall run uses G-R*.)

Common categories for /city-hall-prep gaps:

• pipeline-fragility (MCP outage on lookup_initiative / get_council_member, derivative-doc staleness)
• user-soft (between-cycle published-canon ingestion, voice routing surprises)
• process-gap (skill-vs-actual-practice drift, sheet-read steps demoted because world summary already snapshots)
• canon-risk (faction-bloc topology in voice routing table — Mara-additive framing, faction-vs-individual confusion)

Discipline: write the gap log even on clean runs. The gap log is the only filing channel from this terminal — research-build triages it into ROLLOUT next session. This terminal never writes a ROLLOUT row.

Close gate (mechanical — RB-1, G-S1). The final action of /city-hall-prep is:

node scripts/gapLogGate.js --cycle <XX> --skill city-hall-prep

It exits non-zero until the ## LEG: /city-hall-prep (G-PREP) leg exists in the cycle gap log; skill close is defined as this exit 0. A Stop-hook backstop (gapLogGate.js --stop-gate) blocks session close for the same reason if this step is skipped. Deliberate bypass: GAPLOG_GATE_OFF=1.

Changelog

• 2026-06-20 — v1.8 (S265, research-build closing governance.41 RB-2 G-R6). Live roster-status block at Step 3: every faction-bloc packet now carries the current active/recovering/vacant status of its seats + the live active-voter denominator, taken from the Step-1 get_council_member reconciliation. Step 1 already checked status against truesource; the bloc agent never received it, so its whip-read ran off memory (C98: OPP carried a stale "Crane absent, 5 of 8" when Crane D6 was ACTIVE and the real tally was 9-0). Step 1 checks, Step 3 delivers. Source gap: output/production_log_run_cycle_c98_gaps.md §G-R6. Plan: [[../../../docs/archive/plans/2026-06-20-c98-gap-log-triage]] RB-2.
• 2026-04-17 — v1.0 initial (S156). Voice routing table listed 17 voices including 9 individual council members.
• 2026-05-03 — v1.1 (S197, engine-sheet executing research-build Wave 1 plan per [[archive/plans/2026-05-03-c93-gap-triage-execution]]). G-10 Voice Data Routing rewritten: table now shows the 11 actual agent rows (Mayor + Chief + DA + 3 faction-bloc agents speaking for the 9 council members + 5 project agents) instead of misleading reader into expecting 17 individual agents. Faction membership per Civic_Office_Ledger; previous text mis-listed Chen D8 as CRC, corrected to OPP. [SUPERSEDED S246 G-PREP1 — this "corrected to OPP" was itself the error; truesource (truesource_reference.json + buildCivicVoicePackets.js) has Chen as CRC all along. Reverted forward; see roster above.] G-13 Step 1 sheet reads demoted to verification: Disk inputs (world_summary + engine_review + prior production log + prior published canon) are PRIMARY; sheet reads run ONLY when world_summary is stale. Captures actual S192 working practice (sheet reads were skipped because world_summary already snapshotted everything that mattered). Companion entry on G-15 (between-cycle published canon ingestion) added to Step 1 as Disk source #5.
• 2026-05-12 — v1.2 (S216, research-build closing civic.11). Step 1 — Auto-investigate engine-flagged initiatives: wired scripts/readInitiativeMilestoneNotes.js for mitigator-stuck / remedy-not-firing ailments. Three dispositions (Scenario A / B / C) computed before Step 2 topic-assignment build, capturing the civic.7 INIT-005 false-positive case as a routine prep-time disambiguation rather than a same-cycle wasted voice cycle. Plan: [[archive/plans/2026-05-11-civic-7-init-005-investigation]].
• 2026-06-12 — v1.5 (S256, research-build closing civic.12 G-PREP1). Auto-investigate timing fix: the Step 1 disambiguation checked for a MilestoneNotes entry, but runs at prep — before Step 6 writes that entry — so the current-cycle entry can't exist and every flagged initiative mis-fired into Scenario A (real-signal), defeating the civic.11 false-positive disambiguation. Reframed the three dispositions to read the (normally ) instead of literal . Option (c) per the civic.12 plan — lowest blast-radius, preserves the Scenario A/B/C/D framework, no step-move/gating. Restores intended civic.11 behavior rather than changing it.