The standing view of where Fede's work stands: the three lanes, what needs a word from him, what is live at real properties, and what is parked. Each lane's detail lives on its own page; this page only points.
Rebuilt Sept 7, 2026 from every session since the last pass (117 transcripts, Aug 30 – Sept 7), the current docs pages, the decisions ledger, and memory · Replaces the Aug 29 version in full · Frame: the Constitution (Sept 4): one lane per person, nothing reaches a real property the same day it is built, turning on at a customer is a team call made once.
Since Aug 30 the company changed shape. Camellia went from "Clara as coworker" (holds, reminders, promise alarms) to autonomous on the night of Sept 6–7 on Fede's go. Two new customers arrived: Situs Group (proposal sent Sept 4, phase one targets the week of Sept 22) and its sister company Western Slope (prototype phone line built Sept 5–6, not yet handed over). Renewals lost its second engine (Sept 3) and got a lab program instead of more fixes. And the way sessions work was rewritten into a constitution after the Sept 4 retro. Almost everything in the old "portfolio" table below has been superseded by those four moves.
| Session | Lane and customer | Open PRs | Waiting on Fede |
|---|---|---|---|
| 003 | Fairhaven dry run for Western Slope; company-level tour length | 1 tour-length PR (8653), 2 reverts of the mailbox switch (8671 + one more) | Asset-type first-reply design pick; tour length in Settings next week or leave at 30 |
| 004 | Western Slope shared leasing mailbox: connect, read, reply; real-mail parser gaps | 9, all on Fede's hold | Hold-all or list them; turning the poller and Clara email on; Google login; CASA purchase |
| 000 | Shared mailbox connection and inbound email architecture; Fairhaven end-to-end harness | Poller split into 4, connect flow, harness | Test shared mailbox in our tenant; a cron slot; Gera for a Google project role |
| 002 | Onboarding knowledge review, investigation only | 1 held draft | Go on a Willows test harness |
| 008 | Camellia: owner report, listing auto-publish | none | Yes/hold on a one-time catch-up post for six Camellia units |
| 011 | Timezone correctness, all dark | none | Activation at real properties; two follow-ups |
| 001, 010 | Idle | none | Next instruction |
| 006 | No longer running |
Conflicts
Facts other sessions should know
Ordered by cost of waiting. Each is one pick. Where a page already holds the options, the link goes there.
| Lane | Owner · property | Where it stands (Sept 7) | Definition of done | Page |
|---|---|---|---|---|
| 1. Camellia back to autonomous | Fede · Camellia | ON since Sept 7, 02:24 UTC Hand-offs follow the five-line rule: answer what she knows, one honest deferral, one plain team note per person per day to the real Camellia inbox, a failed transfer stays open, the weekly scorecard shows whether the team acted. The Aug 24 phone-script gate and the hand-run phone upload are gone; phone changes deploy on merge. The overnight flip caused a 13-minute phone-routing outage at 3:48am Denver (no calls landed in it) and a robot test call that left a real voicemail on the office line (left as is, on Fede's order). Confidence self-rated 6 of 10 until a full day of real traffic and Monday's first scorecard run. | Scorecard clean two weekly runs in a row with the removable gates off; missed-call email on; owner address fixed; nothing else reaches the property. | release hub |
| 2. Maintenance intake, dialed in | Gera · Willows bench, then Situs Group | Bench Sessions on Fede's side touched it only where it crossed a bug: the "resident said thanks so no work order was filed" defect (system-wide, every channel) was root-caused and fixed Sept 6; Sean's toilet-leak call was the case. Four Willows demo bugs from Sept 4 are filed, not built (work order promised to an unknown caller never created; status question opens a team ask; descriptions are raw transcripts; recognized tenant greeted generically). | Every inbound maintenance scenario handled end to end at the bench, two clean sweeps, then partner go-live with a named owner. | Gera's page |
| 3. Situs Group onboarding | Fede · Situs Group (+ Western Slope) | Pre-go-live Read-only AppFolio pull done Sept 2 across both Situs databases; findings report, onboarding playbook and "What Clara Knows" page published Sept 3; proposal sent to Eileen and Hugo Sept 4 (opened several times). Plan: replace the maintenance call center first, then nights-and-weekends leasing, then renewals. Setup Sept 8–12, joint QA Sept 15–19, phase one live earliest Sept 22–26. Clara-sends-through-AppFolio proven at Willows for text and email, contained there; Situs activation explicitly withheld. Western Slope (sister company, ~250 scattered doors) got a prototype phone line Sept 5–6: leasing triage, Spanish, maintenance transfer, message-taking, business-hours bridge to the leasing desk. No self-booking, no texting. Died nine hours overnight Sept 5–6 from a same-night booking change; reverted. Weekday human transfer still unverified (needs one real Tuesday call). Invite blocked by item 1 above. Gera-owned inside this lane: first contact on a PMS-delivered line writes a guest card — an unknown number that texts or calls becomes a prospect and gets an AppFolio guest card so the PMS can answer (the Willows find, Sept 7; Fede confirmed the logic was never built; #7274 is the teammate stopgap). Tracked on the onboarding plan and Gera's page, lane 3. | First workflow chosen with them and its scorecard dialed in; named owner who has agreed to reply; their conversations in the corpus in week one. | Situs hub · onboarding plan · timeline |
The activation log the Constitution asks for. One line per change, who said go.
| Date | Property | What turned on | Go |
|---|---|---|---|
| Aug 30 | Camellia | Promise-texting switch flipped on by script, citing a remark of Fede's that was never verified. Scan showed zero real Camellia callers had received a promise text (all 32 went to the Willows line). Now moot: the promise machinery is off at Camellia under autonomous mode. | Unverified quote |
| Aug 30 | All properties | Missed-transfer callback text shipped ungated on Friday night and texted one real Camellia caller. Walked back to a per-property switch the same weekend. This is the incident behind the "ship dark, test at Willows, turning on is a separate step" rule being restated Aug 31. | Nobody |
| Aug 31 | All properties | Two-word transfer announcement (caller's name only, no robot speech) on the phone. | Fede |
| Aug 31 | Camellia | Hold-for-review gate: Clara's team emails wait in Slack for approval, copy to hello@. Turned off again Sept 3 (nine held emails in five days, one real, and that one a false alarm), then narrowed to a temporary Camellia-only exception. | Fede |
| Aug 31 | All properties | The "application sent" email to the office deleted outright. Fede's ruling: it is a work step, not an outcome. | Fede |
| Sept 3 | All properties | One email per caller message, exactly once (an over-eager duplicate filter was swallowing real messages), plus wording fixes. Merged by a stale auto-merge arm before Fede's wording read; Fede ruled "keep." The hold-for-review label is now the only thing the merge guard honors. | Fede, after the fact |
| Sept 3 | All properties | Renewals: the parallel workflow starter deleted; the daily scanner is the only renewal starter. Holdover handling moved inside the renewal workflow, verdict-gated, no standalone sweep. | Fede (Option B) |
| Sept 4, 3:09pm | Camellia | Missed-transfer note to the office: who called, what they wanted, callback number, one fact from the record. Goes to the real Camellia inbox. No text to the caller. | Fede ("go with on when you are done") |
| Sept 4 | Camellia + Yale phone | Five phone changes pushed to the shared fleet: after-hours message desk, voucher answer, fair-housing list, payment readback, balance guardrails. | Fede ("push all 5") |
| Sept 5 | Camellia | Renewal fix shipped ungated; converted three leases to month-to-month the next morning (item 4). Origin of the "silence is not approval" rule. | Nobody (unanswered run/hold) |
| Sept 5 | Camellia | AppFolio applicants with no phone or email are now imported instead of silently dropped; a later sync fills in the contact. No outreach fires for them. | Fede ("finish this work") |
| Sept 7, 02:24 UTC | Camellia | Autonomous hand-off mode on. Promise ledger, reminder ladders, held-for-review queues off for Camellia. | Fede ("turned on tonight") |
| Sept 7 | All properties | Phone-script deploys automatic on merge again; the hand-run upload path is blocked at commit, build and push. Every phone change since Aug 24 had silently failed to reach the live agents. | Constitution IX-A |
Each is (or will be) a corpus case. Outcome in the last clause.
Per the Constitution these do not change product behavior until they become a lane. State as of Sept 7, with the page that holds it.
| Initiative | State | Open for Fede | Page |
|---|---|---|---|
| Portfolio architecture (one model for single buildings, centralized portfolios, hybrids; "never migrate again") | Study published Sept 6 with diagrams and a 16-step ladder; reviewed with Gera the evening of Sept 6 with no hard decisions. Gera owns the "how" and is running his own code-level investigation. A fake multi-company test harness merged and immediately found a real cross-company caller-matching leak on the phone side (fixed there; the login-side leak is item 1). | Decisions X10–X17 on the page (what "company" means, how a shared line binds a conversation, groups, taught answers, view grants, scattered-site homes, Situs on Google Calendar). Plus four module-switch picks from the wizard walkthrough. | the HOW |
| Voice agent architecture | Decided Sept 5 one shared set of agents for every client; when a number points at a company, Clara asks which building. Two defects must be fixed before any shared line serves a portfolio: phone renewal tools have no ownership check, and the appliance lookup can read the whole fleet by bare unit number. | Five open probes on the page (for example, can scattered-site callers name their building). | decision |
| Yale 25 Station / ConAm | Not live to the public. Real email address works (Google recipient map, Sept 2). After-hours desk and phone parity pushed Sept 4. Pricing runs in "website" mode by Fede's call. A new "not ready for showing yet" tour flow proven at Willows and Yale. Remaining gaps are provisioning on ConAm IT's side (calendar consent, ILM user, rent-roll feed, mailbox read) plus the application link they owe us. An unregistered Denver caller booked an unconfirmed tour on the unmarketed line Sept 3; unclear if real. | The reset plan (item 8). Cutover style wording conflict on the tracker (Fede picked "all sources at once" Aug 27; the ConAm access doc still says one at a time). | cutover tracker |
| Renewals | One engine since Sept 3. Hardening done overnight Sept 3–4 (Willows end-to-end proof, one rulebook with drift guard, 30-day replay on every renewal change, comms-history log). Then three of the "finish this" additions were reverted Sept 4 on Fede's "scope is my asks only." Lab program ordered Sept 4; builds paused. | Items 4 and 6. | decision page |
| Escalation / coworker loop (hand-offs that expect a human answer) | Six fixes from the Fabienne review merged Aug 31 – Sept 6 (preferences to the guest card, same-thread replies to staff, ask-lane rules, email traces, work-order cold feet, vanishing "remember this" email). On at Willows and the Yale sandbox; off at every real property. Doctrine locked Aug 31: a preference is never escalated, never invent facts, never give reasons not to book, door codes are never blocked. Constitution: this loop is built only with an engaged partner. | Which partner. Until then, nothing turns on. | release hub · 15 real "can I talk to a person" cases |
| Call promises / promise ledger | Verdict on the hub: not ready as built. At Camellia it fired 36 times, reached a person 10 times, 8 of those were false alarms on tours that went fine. Off at Camellia under autonomous mode. Clara still does what she said on a call (that half shipped Aug 30, proven on 1,904 replayed calls). | Finish it properly or tear it out. Recommended on the hub: watch quietly first, decide later. | release hub |
| Mute a person | Adversarial audit Aug 29–30 found five leaks (contact added after mute, Slack still posts, vendor dialing bypasses it, dead voice code, about 1,000 lines for a rejected approach). Vendor-dial and unknown-category leaks fixed. The Camellia tenant Fede ordered fully muted still sits on the old opt-out record. | One word to rewrite his record onto the new mute path. | verification |
| Timezone class | Audit published Sept 1. Wave 0 ready, dark. | Item 5. | audit |
| Knowledge at scale | Facts not prose (decided); per-property isolation required (re-confirmed). Test property runs one-fact-per-entry knowledge. Voice-native knowledge store won the nine-arm test but only for properties with a dedicated agent. | Item 9. | page |
| Quality desk (Cerebrus, Fede's spelling) | Rebuilt Aug 27–30: rate a message or a whole conversation from the Conversations page, thumbs-down becomes a candidate test nightly, golden set cleaned (132 of 150 imported cases were unreviewed guesses and were quarantined, 13 rescued). The "quality this week" band was removed from the Conversations page Sept 7. Harness Atlas (Aug 26) still proposed: 45+ separate rigs, 9 of which reinvent "replay and grade." | Atlas picks D1–D3 (base runner, incremental vs freeze, Gera owns the doctrine). GitHub "sensitive data removal" request for the resident details that briefly hit two test files. | atlas |
| Lease automation (approve → draft) | State-aware "approve at any stage" shipped Aug 31 – Sept 1; the move-in special is never part of the lease (manual credit); AppFolio's own math is never overtyped. Auto-draft on approval proven at Willows Sept 4–7 (item 7). | Item 7; the eight picks on the page. | approved to signed |
| Dashboard and prospects | Phantom "pending" from a leftover co-signer fixed Sept 5 (one empty unit with any application counts as one). "AppFolio" now shown as source when none is recorded (Fede reversed the earlier rule Sept 7). Co-applicants on the same unit fold into one row: in build, no switch. A 31-site audit found the app already builds a Household record every sync and nothing reads it. | None. | — |
| Website | Public "Spicy" landing page live at propflow-landing-spicy (Sept 6). Focus Lab feedback due Monday Sept 14, 10am: watch the 23-minute Loom and reply as one voice in Basecamp. | The Focus Lab reply. | — |
| Agent ops | Account rotation rolled back to the pre-pool original Aug 31 (the pool was starving Fede's accounts of refresh). One session killed nine sibling sessions Aug 31 to force a credential swap; rule written: never. Rulebook pruned 1,742 to 257 lines Sept 6; constitution in the repo. Review-bot churn cut (duplicate-reviewer race fixed). Weekly usage caps hit repeatedly all week. Smith fail markers: all 10 rows answered, 51 of 51 executed. | Item 10. Sign out other devices on the shared account. Rotate the leaked webhook secret still present in about 30 old files. | rotation tracker · fail markers |
| Engineering health | DORA report Sept 4: deployment frequency, lead time and change-failure rate all "elite." Time-to-recover is unmeasurable because the incident log stopped March 26 while at least 13 production incidents happened (including Clara leaking internal reasoning and another customer's name to a prospect on Aug 24). | Whether the corpus (Constitution VII.5) replaces the incident log, or both are kept. | — |
| Context-aware automation (ruled a major initiative Aug 30 after the tour-day incident) | Not started. No session touched it since Aug 30. Under the Constitution it waits for a lane. | Whether it stays on the six-week backlog. | tour-day RCA |
| Outbound gate / chokepoint | Design decided in August; no build since. The Sept 3 finding stands: the five-touch outreach texts still skip every content check. | Whether it becomes a lane after Situs phase one. | — |
Short form. The full text is in the Constitution and the personal rules file.
Every session transcript on this machine touched since Aug 30 (117 files across all agent terminals) was reduced to Fede's own turns plus each session's closing summaries, then read in full by four readers who reported asks, rulings, what shipped and at which property, open items and incidents. A fifth reader took the current docs pages, the decisions ledger, the Constitution, the lane list and memory. A sixth pulled the Trello board. Nothing here is from recollection; where a claim could not be confirmed it is listed under "not verified."