Build status

Infra Build · BLD

Inputs unchanged at build

Current objective

Get the pipeline into production

Suni's definition of production, in his words: voice memo in, task completed and sent to email, zero prompts, STP. The bottom of the funnel is built and the loop is now proven end-to-end. What remains is the third criterion — a live surface showing what is moving through the pipeline and what is stuck.

2h–3h remaining

What done looks like

✓STP runs on the mini — memo in, artifact out, no human relayProven end-to-end 2026-07-30: memo → card → approve → execute → artifact on disk → email in inbox → terminal state. Every output verified by inspection, not from a log.
○No silent failures — a run that does nothing says soEntry markers landed 2026-07-30 (BLD-88), confirmed firing under real launchd. BLD-85 surface (2) is the open half and is Suni-observed-only — an agent cannot see whether a command was gated.
○A live dashboard showing what is in the pipeline and whereThis page is the build-status half. The pipeline-flow half needs state events emitted at each pre-Linear stage.

Decisions pending

Sandbox Linear target: a dedicated team, or dry-run mode?BLD-57's own open question, now entangled with the comms cutover — both create team structure in the same workspace. Decide together or we build two structures and reconcile later.
Does the WIP rule re-triage the existing backlog, or apply going forward only?Raised by the mini. Without a ruling it gets re-argued per card, which is the cost the rule exists to avoid.
Nomenclature: is the top block an OBJECTIVE, a BUILD, or a SPRINT?'Build' collides with the BLD- prefix; 'Sprint' implies a fixed timebox and ceremony we do not run. Currently rendering as OBJECTIVE — see the note in the memo.

Near-term roadmap

19h30m–32h30m

Retire the file bridge — comms move to Linear3h–5hOne root cause: two writers, no coordination primitive. Cut over while the ledger is empty.
Professional UI — real app, login, works on your phone6h–10hNext.js + Supabase + Cloudflare. This page's second version is the first screen of the product.
Vault off Obsidian Sync2h–4hDesktop-only, ratified. Upstream of the sandbox.
Overnight autonomous builds1h30m–2h30mAdopt the platform sandbox rather than building a wall. Containment instead of approval.
The three disposition rails — Todoist, Calendar, Gmail draft4h–6hThey share a shape, so it is one build rather than three.
Earnings-summary skill3h–5hThe demo skill for Investec. Stand up the skill-authoring process first.

Recently shipped

Derived Linear digest ("top of mind") — SessionStart orientation, cross-machine
Mini allow-list refinement — process a memo end-to-end with zero gates
Output executor layer — registry + the 3 safe executors (the missing bottom of funnel)
Durable headless Claude auth on the mini (setup-token → CLAUDE_CODE_OAUTH_TOKEN)
Extract handler intermittently logs nothing on a successful run (extract.log silent-success blind spot)
Prove one memo end-to-end (STP acceptance) + reconcile the stranded-memo inventory
Extraction / intake agent (voice memo → asks → Linear cards)
Allowlisted outbound email rail (automated household mail)
Guardrail drift-check: detect + reset accidental "always allow" permission drift
Mirror the session-Monitor reaper (suspenders half) from the mini

In Progress

31h20m–2h30m

BLD-100High

~/.claude/ is not under version control — hooks + settings are production infra with no history, no rollback, no parity audit

shared30m–1h

Drift drop landed and confirmed clean 2026-07-30. Remaining: the laptop half, and a detector so drift is noticed by the machine rather than by us.

BLD-113High

TERMINAL state map is hardcoded mini-side and names states that don't exist on routable teams — completed cards land in Backlog

mini30m–50m

Needs the real state list from a live run; mini-local. Completed cards currently land in Backlog — a correctness defect in the production loop.

BLD-111Medium

Graduation convention: a module must bring its test suite in the same drop (outputs.py landed without its 51 cases)

shared20m–40m

Convention, not code — needs a decision on what 'brings its tests' binds.

Up Next

64h10m–8h20m

2BLD-77High

Migrate vault sync: Obsidian Sync → git-backed (kill concurrent-edit clobber)

shared2h–4h

Ratified: vault goes desktop-only. Upstream of BLD-57. Open: whole vault vs. design tree, and attachment strategy.

3BLD-76High

End-to-end pipeline canary (synthetic transaction health check)

mini1h–2h

A synthetic transaction through the live pipeline. Wants the disposition rails to exist first to be meaningful.

4BLD-60Medium

Bridge ledger-parser hardening (header-less bullet drop)

laptop30m–1h

Largely dissolved by BLD-120 — the parser goes away with the mailbox. Do not harden what is being retired.

5BLD-74Medium

Scope the ACTION consumer to disposition-labeled cards (don't fire on build cards)

mini40m–1h20m

Scope the consumer to disposition-labelled cards.

BLD-122High

Hosted migration — Phase C: gateway + dashboard publisher + queue boundary onto a VPS

BLD-123High

Pin `--bare` at every `claude -p` call site before it becomes the default

Blocked

93h45m–6h30m

BLD-107High

Tailscale remedy — real-outage verification (the half a stub cannot prove)

shared15m–30m

Blocked: Requires Suni physically present in DC — the remedy must be proven against a real dead tunnel, and running it remotely severs the session that would observe it. Unblocks on his return to DC.

1BLD-120High

Retire the file bridge — move laptop↔mini comms onto Linear

shared3h–5h

Cut over while the ledger is empty. Decide the OPS state set alongside BLD-57's sandbox target — both create team structure.

BLD-43High

Gateway honey-do route → create Linear issue

BLD-44High

iOS Shortcut on Kristin's phone (dictate → POST → confirm)

shared30m–1h

Blocked: Needs Kristin's phone in hand to install and test the Shortcut.

BLD-50High

Reactor upgrade: draft-only → process-within-gates

BLD-24Medium

Mini notify → gated ACT (autonomy ladder, "last hop")

BLD-45Medium

Robustness: email fallback + confirmation UX

BLD-99Medium

DC location profile — capture display UUIDs, wire `home` + cockpit context

BLD-35Low

GitHub branch protection (Gate 1)

Triage

4615h50m–28h20m

BLD-102High

Audio profile enforcement — per-location lock + anti-drift + United WiFi/AirPods case

laptop1h–2h

Per-location lock plus anti-drift.

BLD-114High

Bridge id allocator races — the reactor and the interactive session both mint M-#### from the same file

shared

DISSOLVED by BLD-120 — server-side ids cannot collide. Left open deliberately; closes when BLD-120 closes. Do not build.

BLD-115High

Research-capable executor rail — the vault executor has no network, so a research-backed deliverable can only block or fabricate

mini1h30m–2h30m

A network-capable unattended agent writing to the vault. The safety posture IS the build. Deferred behind reliability.

BLD-16High

P0 — coexistence prototype (Obsidian Sync vs git, one folder)

BLD-17High

Calendar client (calendar disposition)

BLD-18High

Gmail client (draft-email disposition)

BLD-19High

Todoist client (task disposition)

mini

Rolled into the BLD-33 rail set — do not estimate twice.

BLD-32High

Comment-as-instruction trigger

mini1h–2h

Comment trigger. Partially subsumed by BLD-59.

BLD-33High

Phase-2 disposition clients (Todoist / Gmail / calendar)

mini4h–6h

Todoist + Calendar + Gmail-draft share a shape — one build, not three. They block loudly with a named dep today, which is the correct failure.

BLD-41High

Build: Honey Do iPhone intake for Kristin

mini1h–2h

Honey-Do intake for Kristin. Rail exists; needs the disposition clients.

BLD-46High

Stand up the family intake rail (intake.thakorhome.dev)

mini1h–2h

The intake rail. Becomes the capture screen of the app.

BLD-57High

Sandbox wall — capability-isolation walled garden (mini)

mini1h30m–2h30m

RE-SCOPED: adopt native sandboxing rather than building the wall. Blocked by BLD-77.

BLD-59High

Seamless-comment tooling — comment trigger + present-engine routing + Design State + wake-reconcile

mini1h30m–3h

Comment-as-instruction plus present-engine routing.

BLD-89High

Dispatcher serial execution starves action-extract under bulk riff-synthesis load (fairness gap)

mini40m–1h10m

DEPRIORITIZED — real and unblocked, but nothing user-visible waits on it. Keeps its priority; does not get a session.

BLD-93High

voice-pipeline is not under version control — production code with no undo

mini30m–1h

voice-pipeline is the top of the funnel and has no version control at all.

BLD-103Medium

Consolidate polled health checks into one registry (incl. location-services status)

BLD-109Medium

/allow — one-token permission-rule add + automatic cross-machine parity note

BLD-110Medium

📋 ALLOWLIST QUEUE — standing container, never closes (one comment per ask)

shared

Standing container, never closes — not a build.

BLD-119Medium

Durable staleness sweeps: closed blockers + a repo-ownership pre-check before editing

mini40m–1h10m

Must sweep BOTH relation and prose blockers — BLD-85's was prose, so a relation-only sweep returns clean on the case that motivated it.

BLD-121Medium

Log actual time against estimates so Suni-time calibrates instead of guessing

BLD-20Medium

Vault-vs-folder structure design (intake vs knowledge/output)

BLD-23Medium

Naming cleanup — Ubuntu→placeholder + "Personal Ops"→"Ops"

BLD-27Medium

Cross-machine security parity (deny-list / baseline allow-list → mini)

BLD-28Medium

Laptop-closed → phone push (from the mini)

BLD-29Medium

New inbound-gateway source routes (/gmail, /todoist, /calendar, /feeds)

BLD-30Medium

Daily health-check digest

BLD-37Medium

Linear → Todoist time-block sync agent

BLD-38Medium

Build-plan responder: card → In Progress → draft plan + notify

BLD-39Medium

New-card auto-triage: estimate + gating + suggested sequence

BLD-40Medium

Contractor / property dashboard generator (published read-only)

laptop1h–2h

Published read-only contractor view. Same renderer as this page.

BLD-47Medium

Register Mini + Laptop as Linear agents (assignee = execution router)

shared30m–1h

PROMOTED by BLD-120 — machine-as-assignee is what makes 'who owes this' a query rather than a parse.

BLD-54Medium

Mini: pending-review queue + SwiftBar (parity)

BLD-55Medium

Session lifecycle: chats-as-portals + work-type profiles + auto-orient/closeout

BLD-58Medium

Linear state machine + card description template

BLD-64Medium

Bridge mailbox rotation — cap append-only growth (reconcile + review cost)

BLD-86Medium

Backlog sweep — 21 unsynthesized riffs + the `status: you` strand bug

BLD-87Medium

Auto-close card when its Gmail draft is sent (task-completion = email-sent)

BLD-90Medium

Port the guardrails allow-list drift-check to the mini (L-0131 parity)

BLD-91Medium

Session-handoff collides between concurrent chats in the SAME folder

BLD-14Low

P2 — widen toward whole-vault-on-git (end-state)

BLD-21Low

Delete dormant ST-Vault bridge mailboxes

BLD-22Low

Ops-vault move (Phase 2 step 2)

BLD-25Low

Tag-drift reconcile pass (scheduled agent)

BLD-75Low

park-idea minted direct-to-Parked skips the Parked Ideas.md append

BLD-80Low

Refine VS Code cockpit layout — multi-folder / multi-machine panels

BLD-97Low

Obsidian Sync is high-latency and unobservable (was: "silently not propagating" — that diagnosis was wrong)

Estimates are Suni-time — wall-clock for Suni at the terminal responding in real time, to done. Excludes approval lag and context-switching, so it is the focused-session floor, not a forecast. Column totals sum the cards shown in that column.
Owner — laptop architecture and design · mini execution and always-on work · shared both machines. Every card assumes Suni is in the loop; owner names which machine does the work, not who decides.
Numbered cards in Up Next render in proposed execution order. Click any card to read it here; the link inside opens Linear.
This page is read-only by design. Moving a card between columns or answering a decision in place are writes, and a static page can only write by shipping an API token to the browser. Those two asks are the trigger for the app version, not a gap in this one.
Derived view — Linear and the Bridge are canonical; if a regeneration disagrees with this page, this page is wrong. Judgment layer hand-maintained in focus.json, updated 2026-07-30. Refreshing the browser does not rebuild — re-run build.py.