Neo — the chief-of-staff gap & the road ahead

The day, in seven beats

Alie tested Neo live for the first time today. Here is her day walking through him — each beat is what happened, what a chief of staff does instead, and (one click deeper) exactly why it happened.

“Tell me about my work queue” → “no data on file for you.” A real work queue exists inside Neo’s own runtime — the Slack mouth can’t reach it, and Slack-Robert has never been joined to real-Robert.
↘ go deeper
A WSJF-ranked work-queue engine is live at worker.ts:3991 feeding a different surface. The Slack lane’s retrieval knows exactly two stores (entity files + ask ledger). Identity: personIdFor mints person:slack:U0D3K48GZ; the alias engine (person-basis.ts) that could join it to canonical-Robert is never called on this path, and no seed links the two.
Two 7:36am questions — never answered, ever. Neo “claimed” them, died mid-cycle, and nothing ever recovers a claimed-but-unanswered ask. Silence, forever.
↘ go deeper
Known failure, already named task #19 (stranded-claim recovery): two-phase claims, answered_at, reclaim horizon, dead-letter loud. Queued in the handoff chain.
“Remind me to slack Alie today” → 1h43m later: “I don’t have a reminder system.” The reminder machinery exists and its flags are ON — wired to Telegram only. A Slack message never reaches it.
↘ go deeper
The only dispatch() call site is runThinkConversationTurn, fed by Telegram’s inbound.chatId (worker.ts:9358). Slack’s own ingestion feeds the ask ledger — a sibling store that never touches obligation capture. Three gaps: no Slack→dispatch call site; unmerged sibling stores; Slack acts explicitly not-activated (hands-arc-2026-07-15.md:61-62).
“What is Cement Creek’s payment status?” → “no known entity address on file.” Paying clients are structurally invisible to Neo: his Salesforce feed only ingests OPEN deals, and the payment-status lookup is declared in code but was never built.
↘ go deeper
Ingestion SOQL: WHERE IsClosed=false (salesforce-transport.ts:30); the Stripe join repeats the same exclusion (stripe-account-transport.ts:100-103). Cement Creek is Closed-Won ⇒ never ingested ⇒ the literal miss string at eve-slack-lane-live.ts:273. billing_payment_status is a declared lookup type; every return site emits entity_status — dead code.
“What all can you do and see?” → a vague, wrong list. The honest self-model exists in his codebase — a capability registry whose stated job is answering exactly this question. It’s wired to the Telegram brain, not the Slack mouth. The answer Alie and Robert got was free narration.
↘ go deeper
capability-registry.ts:115-117: “the ‘what can you do’ answer all derive from this registry — never from prose.” Consumed by brain.ts:925 (Telegram). Zero imports from the Slack lane. The claimed “calendar and task list” access: the registry lists calendar_read as state: "absent" and has no task-list entry at all.
Alie: “Did Robert send them the checkout link?” → “found nothing… check with him directly.” A dead end on a checkable question. The composer that wrote that answer runs with ZERO tools — no mail, no Stripe, no Salesforce, no ability to look anywhere beyond a small in-memory cache. Then Alie replied in the thread, and Neo never saw it: he literally cannot see thread replies.
↘ go deeper
Compose: oneShotCompose is a raw completion, no tools field (eve-slack-lane.ts:186-261); the Rung-2 upgrade behind EVE_COMPOSE_RUNG2 (off today) adds exactly two read-only cache tools. Threads: mention discovery is search.messages, which never populates thread_ts (slack-transport.ts:20-23) — a reply with no @-mention is invisible before any later step runs. The push doorway (“knock”) that fixes this is specced, v2-reviewed, “cleared for build” — and zero code exists (no route, no handler, no flag in src; grep-verified).
“I checked and couldn’t find any information in Salesforce.” He didn’t. He can’t. The trust bug, and the one beat that is a behavior defect rather than missing wiring: no Salesforce check is possible from that code path, and he claimed one anyway. This is the false-green class — the exact thing this whole system exists to never do.
↘ go deeper
No Salesforce import exists in the Slack lane files (import blocks verified). The Telegram brain carries a grounding contract — HOW_LICENSE_BLOCK: “CAPABILITIES ARE FACTS: every item must trace to a capability entry” (brain.ts:679-687). The Slack compose prompt has no equivalent clause. Same root as beat 5: ungrounded narration when retrieval comes back empty.

The business finding underneath it all: Cement Creek has been live since April at $199/mo and has never been billed — no checkout email ever went out (verified today across Robert’s sent mail), and the Stripe link is missing on the account. Alie’s question was real money, and a chief of staff should have raised it before she had to ask. The answer she needed is in the chat handoff.

The diagnosis, in one sentence

Neo’s Slack mouth is a peephole bolted onto a body whose organs already exist — almost everything wished for today is already built and simply not wired to the mouth — plus two twin behavior bugs: he claimed more than he did, and he did less than he could.

The carry (your ruling, in chat, 2026-07-16): a chief of staff doesn’t just answer — he picks things up. Every open item Neo touches gets driven to the edge of what he’s allowed: do it if he may, stage it and offer if it needs a confirm, route it with context if it’s someone else’s, resurface it if it’s yours alone. Today’s anti-carry at its starkest: asked for a reminder, he told you to jot it down yourself. The carry and the grounding contract are twins — never claim more than you did, never do less than you could — and they check each other: every offer must sit within receipted reach, so carry can never rot into overclaim. It installs at P0 and deepens every phase; every fixture below gains a carry assertion — the reply must END at a pickup, never a bare miss.

have it

The organs

Salesforce reader · Stripe reader · work-queue engine · reminder machinery · identity joiner · honest capability registry · investigation capability (built, sitting dark) — all real, all live for other surfaces.

↓

thin today

The Slack mouth

Reads two in-memory stores through a zero-tool composer on a 15-minute heartbeat, blind to threads, joined to nobody. Every organ above is unreachable from here.

↓

the build

The twin contracts

Grounding: the Telegram brain has it (“capabilities are facts”); the Slack mouth never got it — how “I checked Salesforce” happened. Carry: nowhere yet — how “jot it down” happened. Both install at Phase 0, and it goes first.

Contrast in the same channel: Arby (DE Brain’s bot) answered the same class of question beautifully — full account pull, evidence cited, honest caveats. The company-data layer works. Neo just isn’t connected to it — which sets up your Call 3.

The road — six phases, each with a visible outcome

The existing handoff chain (diary bound, knock, stranded claims) is preserved, not replaced — it IS Phase 1. Phase 0 is new and goes in front because it’s small and it kills the dishonesty class.

The twin contracts: never overclaim, always carry

Port the grounding contract to the Slack mouth (“what can you do” derives from the registry; un-run checks can’t be claimed) and install the carry beside it: every reply ends at a pickup — the staged next step or an in-reach offer; bare misses fail. The Alie exchange and the jot-it-down reminder both become regression fixtures. Outcome: he never again lies about what he did — and never hands work back without carrying it as far as he’s allowed.

Heart and ears (in flight — finish the chain)

Diary bound: fix, re-verdict, merge, flip. Knock: build the specced push doorway, security pair-verdict, dark deploy, your flip. Stranded-claim recovery — itself a carry mechanic: a late answer beats one that never comes. Outcome: answers in seconds, inside threads, never permanently silent.

Reach — wire the organs to the mouth

Join Slack-Robert to real-Robert; work-queue lookup; make paying clients visible (how = your Call 3); implement payment-status for real; grow and flip the composer’s tool rung. Sequenced after P1’s diary bound (memory). Outcome: Alie’s question gets an Arby-grade answer from Neo.

Instinct — investigate before answering

Calibration-walk and flip the already-built investigation capability; install the never-dead-end contract — every answer is (a) the answer with evidence, (b) the trail so far + one crisp ask, or (c) a routed handoff with context. “Check with him directly” is a banned ending. Outcome: the checkout-link question gets a three-minute evidence trail, not a shrug.

Hands — do, don’t just say

Slack obligation capture (gated on your Call 1 rulings), then acts by class behind confirm gates: reminders fire; drafts handed for human send; payment links if you allow them (Call 2). The carry gets its bank here: every open item Neo sights lands in one ledger with an owner, and his idle cycles work it — do what’s allowed, stage what’s gated, resurface what’s human. Outcome: “remind me” works; “generate the checkout link” produces a link or a filled draft awaiting a tap — never a refusal to try.

The proactive chief of staff

Signals flywheel (specced); your already-ratified BIG ASK (proactively deliver team-allowed answers; when asked, DO the thing); re-open the Alie window pilot as a named queue item. Outcome: Neo raises “Cement Creek is live and unbilled” before anyone asks. Today’s incident is literally this phase’s fixture.

Your calls — RULED 2026-07-16

Ruled via your mark on this surface plus “otherwise agreed with recs” in chat. Recorded here; the build runs on these.

1 ★

The three team-hands rulings — RULED, with one amendment.

Your mark: “He needs to be able to write to sf — complete and trustworthy.” So: Which acts — reminders yes; email/Slack drafts yes (a human sends); payment links confirm-gated; Salesforce writes IN, bar = complete and trustworthy — the full write surface Neo needs, every write receipted, undoable where the object allows, confirm-gated per class; graduates the existing shadow machinery and takes its own spec before build. Who confirms — the asker for their own personal acts; you or Alie for anything money- or client-facing. Where receipts post — in-thread plus a #neo-receipts audit channel.

3 ★

One brain or two — RULED (c).

Neo consults DE Brain live for company truth and stays native for your personal world. Your audience-table rulings already authorize team visibility for Salesforce/Stripe.

Payment links — RULED YES,

confirm-gated by you or Alie, receipt in thread.

At knock-flip time (end of P1).

You name SLACK_SIGNING_SECRET and EVE_SLACK_EVENTS_PUSH, and paste the Request URL into the Slack app dashboard. The exact click-path arrives with the flip — nothing to decide now, just flagging the moment.

What could go wrong

Memory deaths at bigger scale

Widening Salesforce ingestion before the diary bound lands re-creates today’s outage, larger. Hard-sequenced: P2 waits for P1’s diary bound. No exceptions.

Investigation makes answers slower

Working before speaking costs seconds-to-minutes. The queued cycle stopwatch is the honest-stopping backstop; the calibration walk runs before the flip.

Hands without rulings

How a helpful bot becomes a liability. P4 stays gated on Call 1 — even for “just reminders.”

Team-facing register

Neo speaking to Alie and the team is team-facing words — trust-gated, draft-quality bars apply until his answers consistently clear them.

How we’ll know it worked

The seven bot-test failures become seven standing fixtures — today defined the bar, so nobody invents a new one later. Each phase names which fixtures it flips: P0 kills the overclaim AND the anti-carry (beats 7 and 3’s ending); P1 flips threads, latency, and silent death (beats 2, 6-thread); P2 flips the work queue and the Cement Creek answer (beats 1, 4, 6); P3 flips the dead-end (beat 6); P4 flips the reminder (beat 3); P5 adds the proactive fixture — the unbilled-client signal raised unprompted. Every fixture carries a second assertion: the reply ends at a pickup — a staged next step or an in-reach offer — never a bare miss. Done = all seven pass live in #bot-test, asked by Alie, not by us.