How it works underneath — architecture, APIs, schema, and operations, with the code and the files each claim comes from.
27 posts · release summary · RSS
Three findings in twenty-four hours — 'approve' bound to the wrong record, a card narrated as a sentence, a card lost in the auto-file round trip — were each a seam between the pieces that make a card. What shipped: the card persisted on the turn and replayed to the model as the tool call it was, written down before it is filed. What is next: one typed card contract in core. vocion-core#690, #694.
Measured: 20m37s per deploy, 12m16s of it Next.js building on a 2-vCPU box. Now the image is built once on a CI runner, tagged by the core commit, skipped when the tag exists, and pulled by the box. A workspace-only deploy is 6m27s end to end; a core change about 17. Plus the twenty-minute rollback that taught the pipeline to refuse a superseded commit.
Handoff gates: a grammar declared on an object type's transition, a deterministic check that refuses and returns weak work to the seat that produced it, and a judge with the seat's own rubric that passes, returns with a reason code, or escalates to a person. Why the next seat is not the judge, why there is no second review product, and what the first reference runs measured. vocion-core#672, #683.
Per-product reference sets: synthetic inputs written the way a request actually arrives, real shipped commits as the gold standard, a standards page the seat reads and a dataset it never sees. The first set's baseline was 2 of 16 — and why that is the right number to start from.
The data-room collector shipped, ran after every sync, and filed nothing. One AND in a where-clause had pinned it to the sixty seconds around the run it followed — which made a documented constant unreachable and the dedupe code around it dead. vocion-core#463.
The red team's first live run returned nothing parseable twice, and the agent read the document by hand instead and said so. The cause was a length budget, not a reasoning failure — and the fix was to stop asking a model for JSON in prose and declare a tool it has to call. vocion-core#457.
The new code started serving 40 seconds before the backup finished and 59 seconds before the last migration ran — and the swap that put it there killed every open stream. The rewritten flow migrates first, drains first, proves the new container is serving its own pin, and prints a receipt for each phase. It also found nine index builds that had never been applied.
A delegation that blew up rendered as "Delegating to the research specialist" forever, or as "the research specialist finished". The trace reader handled neither shape: a caught tool failure arrives as an ordinary end event carrying an error status, and an uncaught one arrives as a different event the reader had no case for. Four surfaces now say so — including after a reload, which used to rewrite every failure as a success. Release notes 2026-09-16 — vocion-core#368.
A decorative <kbd> showing a shortcut letter sat inside every decision button in the dashboard. Nothing marked it decorative, so it joined the accessible name: the Snooze button was called "Snooze s". A screen reader announced that, and a browser test clicking by exact name never matched — which is how a missing Snooze verb hid behind a failing test for a day. Release notes 2026-09-16 — vocion-core#367.
Eight fixes to the web connector and the candidate extractor, each found by running a real sync and reading the log. Listing-first crawl ordering, feed discovery scoped to the listing path, caps sized for one model call per document, a dry-run line that carries the dedup key, and a record that can no longer flag the card it is refreshing. Release notes 2026-09-16 — vocion-core#356, #362, #364, #365.
One table behind a LangGraph BaseStore, scope carried in the key, a read-only filesystem mount, and a middleware that puts approved rules on the system message. Plus the migration renumbering the release train forced, and an honest account of the one write path the gate does not structurally cover. Release notes 2026-09-16.
A voice guide that asks for a register in the abstract is an instruction a model satisfies by performing the register instead of writing in it. That is where "Curious about…", "No pitch, just…" and "Quick question" come from — the guide's own words. The fix is not better wording: it is a schema refinement on the field the model fills in, a pure lint at three seams, and a retry budget of one. Release notes 2026-09-16 — vocion-core#370.
Mail the workspace and the lead answers in the thread. The inbound webhook verifies the raw body before parsing it, threading walks Message-ID then References then subject, and a stranger's mail becomes an ask instead of an agent turn.
Three archetypes cover every page the dashboard has. Here is what each one is for, how to tell a Ledger from a List, the anatomy of each, and the rule that new pages compose them instead of hand-rolling a layout.
Rendered output stops being text in a transcript and becomes a row: an artifact table, render_* tools, first-party cards, and one org-checked route. Plus the operational lesson from a default that served files without asking who you were.
Release notes 2026-09-15 — vocion-core#318 (merged): approving or rejecting an agent-proposed action no longer writes a generated rule into the agent's learning step; it queues a candidate for the feedback classifier, deduplicated, for a person to accept.
Dump before you migrate and abort if the dump is empty. Log migrations one file at a time and then check that the tables they declared actually exist. Bump the pin by pull request, not by hand. The five practices, and the reasoning behind each.
Every decision a person makes in the review queue is written to a ledger. A six-rung ladder per action kind reads that ledger, and a promotion is refused unless the numbers are there. The thresholds, the schema, and how it maps onto trust.yaml.
Release notes 2026-09-15 — vocion-core#319 (merged): a knowledge source may declare a processor that runs once per changed document inside the sync, with its own budget and failure scope; the web connector reads feeds and URL lists; sources and budgets get v1 routes.
There is no agent picker. A workspace has one lead, the specialists are a delegation roster derived from the org chart, and an @agent or @team tag routes a single turn without changing who you are talking to.
The agent rail is one surface on every page, fed by one typed event union over SSE. Here is the AgentEvent contract, the stream route, how a dropped connection resumes, and how a trace row gets drawn.
Release notes 2026-09-15 — docs/MANIFESTO.md (vocion-core#322, merged): seventeen principles, the loop Outcome → Work → Measure → Learn → Improve → Automate, the Outcome Contract every agent should carry, and a twelve-question test a change must pass before it ships.
Every turn on a workspace 400'd because the harness still sent temperature. The one-line guard shipped in an hour. Writing it up found two defects in it, one of which silently made every default call non-deterministic. Release notes 2026-09-15 — vocion-core#346, #355.
A server component imported a helper from a 'use client' module. It crashed the page — but only after somebody answered a question, because that was the only state in which the component that used it rendered at all. Three bugs found by using the product. Release notes 2026-09-15 — vocion-core#349.
There is a clean checkpoint between the tool node and the next model call. You still cannot put a person's words through it, because of what the message list looks like at exactly that moment. The investigation, so nobody has to repeat it. Release notes 2026-09-15 — vocion-core#352.
Every dashboard URL was /dashboard/… and the workspace was a cookie, so a mailed deep link opened whichever workspace the reader last had active. The fix is a /w/<slug>/… entry route, and the plan after it is /{account}/{workspace}.
Release notes 2026-09-15 — vocion-core#316 (merged): web search, page fetch, crawling and image generation bill the workspace's own Tavily, Brave, Firecrawl or OpenAI key when one is stored, and the server's key only when none is. One definition of spendable; null and failure are different answers.