# City Cave Slice 0 — Topology > Drafted 2026-05-20 · scope: the thin vertical · subject: Penrith > Lens: compositional semantic grammar + parser-combinator brain (per `.planning/2026-05-19-city-cave-brain.md` §6b) --- ## 0 · Thesis **The brain is a parser interpreter.** Skills, rules, knowledge, client folders, and per-instance intel are all productions in a single compositional grammar. The agent runtime walks the grammar against a live input (an email, a question, an interaction) and emits a typed Answer. There is no separate "City Cave brain" — there is the **unBasic grammar**, **extended** by a **City Cave franchise grammar**, **extended again** by a **per-cave instance grammar**. Composition is by parser combinator. The runtime executes the composed grammar. This document maps the topology of Slice 0 (the thin vertical for Penrith / Kevin Lee) along five axes: 1. Composition (how the brains compose) 2. Hierarchy (the cascade of rules) 3. Grammar (the productions per layer) 4. Flow (how a query traverses the topology) 5. Time (the two clocks) --- ## 1 · Composition — unBasic ⊕ City Cave ⊕ Cave ### 1.1 The composition model ``` Brain :: ServiceContext → Query → Answer unbasicBrain :: Brain Agency ccBrain :: Brain (Franchise CityCave) -- extends unbasicBrain penrithBrain :: Brain (Cave Penrith) -- extends ccBrain answerKevin :: Query → Answer answerKevin = unbasicBrain <> ccBrain <> penrithBrain -- left to right: each layer narrows the context -- right wins on specificity; left wins on invariants ``` This is *not* containment ("CC lives inside unBasic"). It is *extension* — each brain is a typed extension of the one above. The current filesystem layout (CC nested inside unBasic) is an **implementation choice**, not the composition model. Other clients (DTC, non-franchise) would extend unBasic differently: ``` otherClientBrain :: Brain (DirectClient Foo) -- also extends unbasicBrain ``` ### 1.2 What each layer contributes | Layer | Contributes | Example | |-------|-------------|---------| | **unBasic** | Voice, frameworks (campaign analysis, audience segmentation, content optimisation), agent hierarchy (strategic/tactical/operational), security/audit invariants, capability skills | "No auto-send" (D5) · "Citations required on numerical claims" · `meta-ads-manager` skill · `email-triage` skill | | **City Cave (franchise)** | Service offerings (float / massage / sauna), brand-voice constraints (City Cave global guidelines), franchise SOPs, head-office relationship (Maddi · Jess Altmann · Jeremy · Tim · BDC), franchise-wide commercial rules | "Ads must run from local cave page when local audience" · "Territory expansion requires Jess sign-off" · franchise voice register | | **Penrith (cave)** | Owner identity (Kevin Lee), per-cave campaign portfolio (LSM × 4 dirs, gift cards, NDIS, custom Penrith-page), relationship intel (cares about / said no to), territory map (Richmond N / Liverpool S / Minchinbury E / west open), local data (MyTime sales, invoicing) | "Sincerely, Kevin Lee — formal sign-off" · "No geo-redirect to homepage" · "NDIS retargeting only" · ROAS 12.34 on Penrith-page custom | ### 1.3 The composition operator The plan defines `caveParser = CaveContext <$> profileParser <*> salesParser <*> invoicingParser <*> adHistoryParser <*> commsParser <*> relationshipParser`. The brain composition is the same shape, one altitude higher: ``` agentRuntime = AnswerContext <$> unbasicRules -- always-on invariants <*> activeSkill -- selected by query <*> clientContext -- City Cave <*> instanceContext -- Penrith <*> liveQuery -- the input ``` Two algebraic facts: - **Applicative for parallel layers**: when two layers are independent (e.g. `activeSkill` and `clientContext` don't depend on each other), `<*>` is the right operator. Order doesn't matter. - **Monadic when later layers depend on earlier ones**: e.g. *which* skill to load depends on what the query is about; the query is parsed first, then a skill is bound. This is `query >>= chooseSkill`. --- ## 2 · Hierarchy of rules — the cascade ### 2.1 The cascade ``` ┌──── INVARIANTS (cannot be narrowed away) ────────────────────┐ │ Layer 0 — Cosmic Anthropic guardrails │ │ Layer 1 — Constitutional CLAUDE.md + .claude/rules/ │ │ No auto-send · audit trail · client │ │ success first · transparency │ ├──── COMPOSITIONAL (narrow on the way down) ──────────────────┤ │ Layer 2 — Skill .claude/skills//SKILL.md │ │ domain-specific procedure rules │ │ Layer 3 — Knowledge knowledge/playbooks/ · research/ │ │ patterns, frameworks, prior art │ │ Layer 4 — Client knowledge/clients/city-cave/ │ │ franchise voice, SOPs, escalation │ │ Layer 5 — Instance knowledge/clients/city-cave/penrith/│ │ owner preferences, history, intel │ ├──── SITUATIONAL (per-invocation) ────────────────────────────┤ │ Layer 6 — Hooks .claude/settings.json hooks │ │ pre/post-tool validators, audit │ │ Layer 7 — Live the query + current state │ └──────────────────────────────────────────────────────────────┘ ``` ### 2.2 Cascade laws Three laws govern how layers interact: 1. **Specificity wins on narrowing.** Layer 5 ("Kevin wants Penrith page only") is more specific than Layer 4 ("City Cave brand voice"). When narrowing, lower wins. 2. **Invariants are not narrow-able.** Layer 1's "no auto-send" cannot be overridden by Layer 5 even if Kevin said "please auto-send." Invariants form a downward closure: every lower layer must satisfy every higher invariant. 3. **Promotion is deliberate.** A pattern observed at Layer 5 (multiple cave owners hate weekend-only ads) only becomes a Layer 4 rule via human-reviewed promotion. The capture loop writes to Layer 5 by default; promotion to Layer 4 is a separate decision. ### 2.3 The cascade walk For any query, the runtime walks the cascade in two passes: **Pass A — accumulate constraints (top down):** - Load invariants from L0/L1 (always in context). - Load active skill from L2 (matched on query). - Load relevant L3/L4/L5 knowledge (matched on entities mentioned). - Activate hooks at L6. **Pass B — generate answer (constrained):** - Generate a candidate answer that satisfies the accumulated constraint set. - Hook validators reject violations. - Output goes to user. This is the same shape as CSS cascade + computed style, applied to brain rules. Specificity + !important. --- ## 3 · Compositional semantic grammar — the productions Each layer contributes grammar productions. A query is parsed by composing them. ### 3.1 unBasic layer (productions) ``` ::= | | | | ::= "what is" "for" ::= "draft" "to" ::= "reply" | "follow-up" | "summary" | "report" ::= direct + no-fluff + sentence-short + ai-as-collaborator ::= must-cite-sources · must-not-auto-send · must-log ``` ### 3.2 City Cave (franchise) layer (productions) ``` ::= | -- adds to unBasic ::= "City Cave" | "the network" ::= | "float volume" | "membership" | "gift card sales" | "NDIS leads" ::= | "territory request" | "campaign brief" ::= "→" "Jess Altmann" "→" {Jeremy, Tim, BDC} ::= + brand-words-only · no-superlatives · wellness-register ``` ### 3.3 Penrith (cave) layer (productions) ``` ::= "Penrith" | "Kevin's cave" ::= "Kevin" | "Kevin Lee" ::= "Sincerely," NEWLINE "Kevin Lee" NEWLINE -- for reply drafts the agent should match this register ::= + numbered-when-multi-point + paper-trail-required + timely-or-he-chases ::= ads ::= local? "Penrith page" : "City Cave main page" + NEVER geo-redirect-to-main-homepage ::= ndis-leads ::= retargeting-only + NEVER email-outreach ::= {Richmond, Liverpool, Ed-Park, Minchinbury, west-open} + if-purchased(t) → STOP ads-in(t) ::= LSM × 4 dirs ⊕ gift-card ⊕ NDIS ⊕ Mothers-Day ⊕ Penrith-page-custom ⊕ national-group-buy ``` ### 3.4 What "compositional" buys you A query like *"Draft Kevin a reply about the ROAS on the Penrith-page custom campaign"* is parsed by traversing all three layers: ``` parse "draft reply about ROAS Penrith-page custom to Kevin" = unBasic : = "draft" + ("reply") + "to" ("Kevin") + City Cave : ("ROAS") on ("Penrith-page-custom") + Penrith : = Kevin Lee · use ("Sincerely, Kevin Lee") · apply (numbered if multi-point, paper-trail) · apply (already Penrith page — confirm in reply) · the campaign exists in Kevin's portfolio at ROAS 12.34 ``` The Answer is constrained by ALL productions matched, in their narrowing order. The skill that *generates* the reply only needs to satisfy the composed grammar — it doesn't need to know which layer each constraint came from. --- ## 4 · Static topology (filesystem + composition) ```mermaid graph TB subgraph L1[Layer 1 · Constitutional · unBasic] CLAUDE[CLAUDE.md
org constitution] RULES[.claude/rules/
00-overview · 01-workflow · 02-commands · 03-feature-schema · 04-progress-log] HOOKS[.claude/settings.json
pre/post/Stop hooks] end subgraph L2[Layer 2 · Skills · unBasic] S_GBRAIN[gbrain-retrieval
pgvector + BM25] S_EMAIL[email-triage
gws-based] S_REPORT[monthly-reporting
Agency Analytics] S_FRAN[franchise-navigator
POC · escalation · voice] S_META[meta-ads-manager] S_ANSWER[answer-cave-query
NEW for slice 0] end subgraph L3[Layer 3 · Knowledge · agency-wide] PB[knowledge/playbooks/] RS[knowledge/research/] end subgraph L4[Layer 4 · Client · City Cave franchise] GLOBAL[_global/
brand · voice · service offerings] FRANCHISE[_franchise/
HQ relationship · Jess · Jeremy · Tim · BDC] end subgraph L5[Layer 5 · Instance · Penrith cave] PROFILE[profile.md
frontmatter + §3 relationship block] EMAILS[emails/
synced comms thread] NOTES[notes.md
human-written context] ADSEAM[ads-context.md
EMPTY SEAM] SALES[sales-history.md
OUT in slice 0] INV[invoicing.md
OUT in slice 0] ADH[ad-history.md
OUT in slice 0] CHLOG[changelog.md] end subgraph L7[Layer 7 · Live] Q[query / email / call] A[Answer / draft] end CLAUDE --> RULES --> HOOKS HOOKS -.invariants.-> L2 & L3 & L4 & L5 L2 -.reads.-> L3 & L4 & L5 GLOBAL -.narrows.-> PROFILE FRANCHISE -.narrows.-> PROFILE PROFILE -.most-specific.-> Q Q --> S_ANSWER S_ANSWER --> A style HOOKS fill:#f97316,color:#fff style PROFILE fill:#f97316,color:#fff style S_ANSWER fill:#f97316,color:#fff style ADSEAM stroke-dasharray: 5 5 style SALES stroke-dasharray: 5 5 style INV stroke-dasharray: 5 5 style ADH stroke-dasharray: 5 5 ``` **Legend:** orange = slice-0 critical path · dashed = exists as seam/stub only · solid arrows = data/rule flow · dotted = invariant enforcement. --- ## 5 · Dynamic topology — flow of a single query ```mermaid sequenceDiagram autonumber participant K as Kevin Lee participant GM as Gmail (unbasic.com.au) participant ET as email-triage skill participant AR as Agent Runtime
(Claude Code session) participant L1 as Layer 1 invariants participant L5 as Penrith profile + history participant GB as gbrain-retrieval participant FN as franchise-navigator participant SA as answer-cave-query
(Slice 0 skill) participant H as Human (Max/Joe/Brooke) participant CL as Capture loop
(D5) participant GIT as Git K->>GM: emails a question GM->>ET: gws gmail +watch → new msg event ET->>AR: classify · is this owner-query? AR->>L1: load invariants (no auto-send · cite) AR->>L5: load Penrith folder + Kevin profile L5-->>AR: relationship intel + history + active campaigns AR->>GB: retrieve relevant chunks (vector + BM25) GB-->>AR: ranked passages with citations AR->>FN: voice + escalation route FN-->>AR: register (formal · numbered · timely · sign-off) AR->>SA: compose draft SA-->>AR: candidate reply with citations AR->>H: present draft for approval (NEVER auto-send) H-->>AR: edit / approve / reject alt human approved with edits AR->>CL: propose diff to Penrith profile + changelog CL->>H: show diff H-->>CL: approve CL->>GIT: commit end AR->>GM: send approved reply (gws gmail +reply) ``` **The single dramatic point** of this diagram: step 14 (human approval). It is the constitutional invariant from Layer 1 (no auto-send · D5 from Exoskeleton 02), enforced by the hooks at Layer 6, before any send-side tool is allowed to fire. Removing this step is the failure mode that ends the project early. --- ## 6 · The capture loop — the Applicative→Monad boundary ```mermaid graph LR Q[query] -->|parsed| ANS[Answer + voice] ANS -->|human edits| EDIT[edited reply] EDIT -->|"diff vs
relationship-intel"| DIFF[proposed diff] DIFF -->|human approve| COMMIT[git commit] COMMIT -->|writes to| L5[Penrith profile + changelog] L5 -->|next query| Q style DIFF fill:#f97316,color:#fff style L5 fill:#141428,stroke:#f97316,color:#fff ``` This is the loop that **distinguishes brain from archive**. Without it the Penrith folder is a static snapshot of 20 May 2026 and decays from that date forward. With it, every interaction is a training signal that compounds. **The promotion rule** (when does a Layer 5 observation graduate to Layer 4?): ``` if observed_at(layer_5) for ≥3 distinct caves AND consistent_signal: propose_promotion(layer_5 → layer_4) human_review(promotion) ``` Promotion is a deliberate, reviewed event. Default = stay at the layer where the pattern lives. --- ## 7 · Time — the two clocks ```mermaid graph TB subgraph Heartbeat[Heartbeat · Applicative product · monthly · 3rd of month] MT[MyTime sales] -->|parser| SALES[sales-history.md] INVS[invoicing sheets] -->|parser| INV[invoicing.md] META[Meta + Agency Analytics] -->|parser| ADH[ad-history.md] REP[reports branch] -->|parser| REPORTS[reports/] end subgraph Nerve[Continuous Nerve · Monadic fold · per event] INBOX[unbasic.com.au inbox] -->|gws +watch| ETN[email-triage] ETN -->|append| EMAILS[emails/] ETN -->|trigger| QUERY[query→answer cycle] QUERY -->|capture loop| INTEL[relationship intel + changelog] end Heartbeat -.->|derived ⟷ source
distinction critical| Nerve style INTEL fill:#f97316,color:#fff style Heartbeat fill:#141428,stroke:#3b82f6,color:#fff style Nerve fill:#141428,stroke:#22c55e,color:#fff ``` **Slice 0 runs only the Nerve.** Heartbeat sources come on in Slices 1–3. **Write discipline (algebraic, not stylistic):** - Heartbeat writes to **derived** files only. Must be 100 % re-derivable from source. Never edited by humans. - Nerve writes to **source** files (profile relationship block, notes, changelog). Edited by humans + capture loop. - The line between source and derived files must be clear in the folder layout. Mixing them = corrupting the parser. --- ## 8 · What's IN Slice 0 vs OUT (explicit) | | IN Slice 0 | OUT (later slice) | |---|---|---| | Cave folders | `_global/voice-and-tone.md` (1 page), `_franchise/profile.md`, `penrith/profile.md` (relationship block filled from Kevin's brief), `penrith/emails/` (from real history), `penrith/notes.md`, empty `penrith/ads-context.md` seam | per-cave fan-out (Slice 6+) | | Skills | `answer-cave-query` (new), `gbrain-retrieval` (existing), `franchise-navigator` (existing) | full `email-triage` continuous nerve (Slice 1+), `monthly-reporting` cron (Slice 4) | | Data sources | unbasic.com.au mailbox (read-only, **today's workshop unblocks this**), past Kevin emails | MyTime (Slice 2), invoicing (Slice 1), Meta + Agency Analytics (Slice 3) | | Two clocks | Nerve only, query-on-demand | Heartbeat (Slice 4) | | Capture loop | proposed but not auto-firing — manual edit | live & gated (Slice 4) | | Approval UI | git directly (technical users only) | thin intake surface for non-git users (Slice 4) | | Ads automation | empty seam at `ads-context.md`, never wired | post-Max-chat, post-Slice-3 | | Graph DB / Convex | none | Slice 5+ if cross-cave query volume warrants | | Brain promotion (L5→L4) | none | as patterns emerge | **Slice 0's "done" criterion (from the plan):** one real past Kevin query answered by `answer-cave-query`; human-assembly time vs agent time recorded. *That delta is the business case.* --- ## 9 · Ultrathink — non-obvious design pressures Ten things a less-thorough read would miss: 1. **The agent runtime IS a parser interpreter.** Improving the brain = improving the grammar. Adding a capability = adding a production. Each Skill is a typed grammar fragment. 2. **Composition is left-to-right narrowing, not override.** Lower layers narrow scope; they don't contradict invariants. Hard constraints (no auto-send) are downward-closed. 3. **Higher layers cache; lower layers re-read.** Constitutional rules load once. Per-cave intel reads per query. This is the same shape as Skills' progressive disclosure, one altitude up. 4. **Carpaccio test = composition test.** Slice 5's "cave #2 with zero new combinator code" is exactly the test the composition model must pass. If composition is right, the cave is just data. 5. **Failed parse = next slice's roadmap.** Log expected-set on every failure. Weighted by query frequency, that log is auto-prioritised ingest backlog. Without it the next slice is hand-picked (worse). 6. **The Applicative→Monad boundary is the archive→brain boundary.** The Monad (capture loop) is what makes the system *learn*. Without it: sophisticated search. With it: an intelligence that compounds. 7. **CC brain could be its own repo.** Composition by typed extension implies brain-as-library. `unbasic-brain` could be a dep that `cc-brain` imports. Other client brains compose the same way. This is the right factoring once the pattern hardens at Slice 5. 8. **Promotion rule is required.** Without "when does a Layer 5 observation graduate to Layer 4," everything ends up at Layer 5 and Layer 4 never accumulates. Pattern: ≥3 caves consistent + human-reviewed promotion. 9. **Write discipline = source vs derived separation.** Two clocks only works if Heartbeat-files are re-derivable and Nerve-files are not. Mixing = corrupting the parser. The folder layout must encode the discipline (e.g. derived files under `/derived/` or marked by frontmatter). 10. **The approval UI is the unsolved prerequisite.** Capture loop assumes a human approves a diff. Brooke won't open a PR. Until the approval surface lands (Slice 4 latest), the capture loop is half-built. Best Slice-0 stopgap: git directly, technical users only. --- ## 10 · Open questions (for the Max chat / next decision) 1. **Brain factoring.** Keep CC nested in unBasic? Or extract to its own repo when patterns harden? Affects Slice 5+ shape. 2. **Promotion mechanism.** Who decides Layer 5 → Layer 4? Same approver as capture loop? Different cadence? 3. **Approval UI.** Git direct vs thin intake. Brooke/Max/Maddi access pattern? Sequenced into which slice? 4. **Empty-seam contract for `ads-context.md`.** What fields does Max's ads-doll expect to read? Defining the schema *now* (even though wiring is later) lets ads-doll author against a stable interface. 5. **Mailbox identity for the email nerve.** Once Joe + Max are gws-authed, whose mailbox does the nerve actually read? Shared `info@unbasic.com.au` mailbox? Or per-person + filter for cave-owner senders? --- ## 11 · The one-paragraph version The brain is a parser. Layers compose by typed extension (unBasic ⊕ CC ⊕ Penrith), each contributing productions to a single semantic grammar that the agent runtime walks against any live query. Specificity wins on narrowing; invariants are not overridable; promotion to higher layers is deliberate. Slice 0 runs the continuous nerve only (inbox-driven query→draft→human-approval→capture-loop), with the four heartbeat data sources stubbed and the ads seam empty. Success = one real past Kevin Lee query answered, agent-vs-human time delta recorded. That delta is the business case for everything that follows.