Changelog

The scheduled sync works again

Upload a resume to someone on the bench

• ›The file is stored and its text read, the same way applicant resumes are, so
• ›The app then suggests skills from it, each with the line of the resume it
• ›The card's Resume link opens the stored copy.

The plugin works in claude.ai

Scorecards save again

Two steps in the general process are optional

• ›Step 5, Criteria Corp — decided during Role Setup. Where it is used it
• ›Step 7, Round 3 — renamed from "Round 3: Hiring Manager" to **Final

Ask the assistant who to interview first

The assistant counts open jobs

See where each applicant came from

• ›Applied via on each candidate's page: ZipRecruiter, Indeed, LinkedIn,
• ›The candidate list shows it under each name, for example "via ZipRecruiter".

The process section says what changed

The home page opens with where hiring stands

Only admins can change scoring weights

Strengths read as sentences again

The overview sheet, rebuilt as an infographic

Scoring and MCP tokens are reachable

Set up a role's interview questions in the app

• ›Draft from the job description. Claude drafts the decision question, the
• ›Edit anything, then Approve. Guides stay off until a person approves,
• ›The hiring manager is filled in from the role request when the app can match

Interview guide buttons say what they are

A one-page overview you can send to anyone

Score every new applicant in one click

Resumes are read again

"Open in Recooty" opens the candidate

Scoring still waits for a click

New applicants show up again

New jobs appear on their own

A failed sync says so

The JD generator works again

Open an archived resume from a search result

Archive results are readable again

Act on a candidate without leaving the page

Recent applicants come first

Narrow a search without leaving the box

• ›since:2024 - applied that year or later
• ›in:michigan - city or state, quote it if it has a space
• ›role:"data scientist" - the role they applied for
• ›exp:5 - at least that many years of experience

Search everyone we have ever seen, from the pipeline page

The assistant can search resumes, and knows what we hold

The assistant is on the pipeline page

Archive results say who someone is, not just that a word matched

Every archive result carries its date

Ask who knows a skill, in Claude

Search by the jobs people have held

Find people by where they worked, and where they are

Searching stays fast as the archive grows

Answers that admit what they did not search

The archive now holds every old resume, and counts them honestly

Search inside resumes, not just names

Three thousand old resumes, searchable

Compare candidates side by side

The site fits on a phone

• ›The home page header put the wordmark and four buttons in one row that
• ›The changelog was widened by long file paths in code spans, which are a
• ›Staffing Requests squeezed its title, filter tabs, the request button and
• ›Ask about hiring shrinks to its icon on a phone. At full width the pill
• ›"Before You Start" and its subtitle shared one line and collided.

Open roles on the home page

Recruitment Analysis no longer asks anyone to refresh anything

Checklist bullets that look clickable now are

The CGO candidates are findable again

Who needs what account

Quote a bill rate range, not a single number

The assistant remembers what it was asked

Newer models where the work is judgement

The assistant answers from live data

• ›Who on the bench knows Zoho? Bench people by skill, with hours available,
• ›Who are the top scoring candidates? Pipeline candidates with their score,
• ›Where does hiring stand? Counts by job and stage.

What it will not do

• ›No scoring, no guide generation. Both cost real money and both have
• ›No salary or internal cost rates. Bench rate is what we would bill, and
• ›No inventing. If a resume is not stored here it says so. Most bench

The bench tells you who to chase

Searching the bench by what people actually do

Sign-in required

Report a bug from wherever you found it

Fill in every rate in one pass

Cost and margin, for admins only

The bench tells you what to do about it

Edit without leaving the page

A list, not twenty cards

Sort, and find who needs chasing

Keep-warm in one click

The process guide reads as one process again

• ›The steps are numbered from 1 again. The General Process opened at step 3,
• ›"Before You Start" is said once, not three times. The section now carries
• ›The request form is a link. The card that tells you every hire starts with

The process guide matches the 2026 process document

• ›Role Setup comes before Sourcing. The app had it the other way round. You
• ›The public posting decision is stated once, in Sourcing and Entry Points,

Ask about hiring, back on the home page

Interview guides open with a brief, not a score

The rubric belongs to the role

Questions say what a weak answer sounds like

Guides say what they are and link to their sources

Closing the round

The talent bench lives in the app

• ›Augusto people with spare capacity — Mike Perry, Aaron Brander, David
• ›Strong interviews waiting for the right work — Brent Scarafone, Reece
• ›People added by hand, who exist in no other system

The process guide matches the process

Four stages that were missing

• ›How a Role Gets Opened — the recruiting request. Every hire starts here,
• ›Sourcing and Entry Points — bench first, then referrals, then a public
• ›Offer, Onboarding, Close-the-Loop — including turning the posting off
• ›Bench and Keep-Warm — the keep-warm message and the 4-to-6 week check-in

Action buttons appeared on the wrong stages

Freshteam is out of the app

• ›Live job postings is gone from the home page. It listed Freshteam
• ›The chat box is gone from the home page and the four pipeline pages. It
• ›FRESHTEAM_API_KEY is no longer read by anything.

Connecting Claude, written down

The Google Doc export works

Test submissions can be kept out of the pipeline

Recooty base URL

The Google Doc export actually works

Create the Google Doc without scrolling

Failed exports say what to do about them

Sections are easier to find

Word resumes are read properly

• ›**Candidates who sent a .docx were being scored as if they had submitted a
• ›Both have been re-read from the copies already in storage and re-scored.
• ›A resume in a format we still cannot read says so in the log rather than

Housekeeping

• ›Removed /api/debug-freshteam, a debug endpoint that was reachable in
• ›scripts/score-all.ts accepts candidate names, so a fix affecting two people

Guides and JDs now land in the Recruiting shared drive

• ›Calls to Drive now pass supportsAllDrives. Without it Drive leaves
• ›The service account has to be a member of the shared drive, not merely

The candidate page tells you where things stand

How good is this score, really?

Interview guides go to Google Drive

Fathom recordings are visible, and you can go find them

Easier to read

• ›The page is wider — 1400px, matching the job pipeline — so evidence lines stop
• ›Section headers show their points as a share of what was available:

Connect Claude to recruiting

• ›/connect — setup instructions written for someone who has not added an MCP
• ›/admin/mcp — who has connected, what they have called, and a revoke button
• ›Admin is an email allowlist in ADMIN_EMAILS — an env var so the list changes

Interview guides close the loop

• ›Scorecards are captured in the app — rate against the same CV1–CV5 rubric
• ›Tick questions off as you ask them, and the guide shows how much was covered
• ›Later rounds know what earlier ones asked. A deep-dive guide generated
• ›Shareable interviewer link — a clean URL with the guide and scorecard only,

augusto-interview skill

• ›New skill for the augusto-tools marketplace. Uses the MCP server for live

MCP server

• ›Five tools: search candidates, get full candidate detail with evidence,
• ›OAuth 2.1 with PKCE, per the MCP specification. Clients register
• ›Augusto-branded consent screen naming the client and exactly what it will
• ›Every tool call is attributed to a person and a client, so an action on
• ›Nothing here spends ATS credits: tools read stored data and can re-score, but

Tune scoring weights from the UI

• ›Sliders on /pipeline/scoring with a live preview showing exactly how
• ›The preview rescores from stored evidence, so dragging a slider costs
• ›Saving creates a new profile version. Existing scores keep the weights

Scoring from the UI

• ›A candidate with no score now has a Score this candidate button on their
• ›When details have not been fetched, the button says so and names the cost:

Scheduled sync

• ›Corrected: Cloudflare Pages has no Cron Triggers — they are a Workers

Scheduled sync

• ›POST /api/cron/sync — pulls the roster and replays pending webhooks.
• ›Roster sync is free, so it can run every 15 minutes. Candidate detail stays

Webhook visibility

• ›Duplicate deliveries are now counted rather than silently discarded, and the

Webhooks no longer lose events

• ›Unmatched events now stay pending instead of being marked processed
• ›Every roster sync replays them, so a candidate ingested later still picks up
• ›Replay records a transition only when the stage actually differs, so it never

Webhook hardening

• ›Secret token in the URL. Register the webhook as
• ›Workspace check. Events whose team.id is not Augusto's are rejected —
• ›Event allow-list. Only subscribed types drive state changes; anything
• ›Structural limits. POST only, JSON only, 256 KB cap — so one large POST

Security

• ›Every API route now verifies the caller. Cloudflare Access signs a JWT on
• ›12 tests covering the ways verification could be fooled: forged signatures,
• ›Verification is skipped when CF_ACCESS_TEAM_DOMAIN / CF_ACCESS_AUD are

Interviews on the candidate profile

• ›Linked recordings with date, length, and a link to watch, plus the AI summary
• ›Score movement shown side by side — on paper vs. after interview, with the
• ›"Re-score with interviews" button

Interview transcripts as scoring evidence

• ›Fathom recordings are matched to candidates and their transcripts stored
• ›Scores now carry a phase — pre_interview (resume and assessments) or
• ›The JD dimension no longer requires a resume: a transcript alone is enough

Matching

• ›A recording links automatically only when the candidate's email is in the

Interview guides

• ›Generate a guide per candidate for the Scan, Deep dive or Final round
• ›Built from the candidate's own score: the gaps scoring could not verify
• ›Core values questions target the values with weak evidence, drawn from the
• ›Every question carries a "listen for" so a non-expert interviewer can judge
• ›Time-boxed to the round, with a rationale explaining why each section is
• ›Printable, with ruled space for notes and a 1–5 scorecard covering role fit

Candidate profile

• ›Verdict summary at the top — strongest and weakest dimensions, the most
• ›Sections are now collapsible cards with tinted headers, so a long profile
• ›Dividers between dimension blocks inside each section

Pipeline

• ›Rank markers replace the plain grey numerals — the top three read as filled
• ›The Augusto wordmark now links home from every pipeline page
• ›Recruiting assistant available on all pipeline pages, not just the guide

Readability

• ›Larger type throughout the pipeline — names, titles, evidence and score bars
• ›Brighter evidence colors: green + for what earned points, red − for gaps,

Candidate profile

• ›Grouped into sections — Role fit, Culture & values, Background, Assessments —
• ›Core values grouped by value, each headed by its name and statement, with the
• ›Skills from the ATS parse, with JD matches promoted and the rest collapsed.

New

• ›/pipeline/scoring — what each dimension measures, where its evidence comes

Pipeline UI

• ›/pipeline — open roles with applicant counts, scored counts and flags
• ›/pipeline/[jobId] — candidates ranked by score, filterable to scored or
• ›/pipeline/candidate/[appId] — per-dimension breakdown with the evidence and
• ›Coverage is shown beside every score, and flagged when thin: a high composite
• ›Unscored candidates sort last rather than as zeros — unknown is not bad
• ›"Sync from ATS" pulls the roster only, so it never spends candidate credits

Fixes

• ›Rubrics are inlined at build time (npm run build:rubrics). Reading them with
• ›scores.coverage is stored rather than re-derived at read time, where a real

Database

• ›Supabase Postgres schema: 15 tables, row-level security on every one, 28 policies
• ›Scoring weights live in scoring_profiles as data, so one engine serves
• ›Scores record model, rubric, weights and JD version, so results stay
• ›Interviewers cannot read composite scores — seeing one before interviewing

ATS integration

• ›Provider-agnostic ATSAdapter interface with declared capabilities()
• ›Recooty adapter, verified against the live API: jobs, paginated
• ›Webhook receiver for Job Created / Application Received / Parsed / Moved —
• ›Stage transition history, since the ATS has no stage API to replay from
• ›ats_detail_fetched guards Recooty's per-candidate view credit

Scoring engine

• ›CGO-Job's model ported from Python to TypeScript
• ›9 deterministic dimensions (location, leadership, AI depth, core values,
• ›3 AI dimensions via Claude with structured output — JD match, ICP alignment,
• ›Every dimension returns evidence and gaps, never a bare number
• ›Missing data is not bad data: a dimension with no input is excluded from
• ›18 tests

Candidate ingest

• ›Two-tier ATS sync: roster sync is free, candidate detail is opt-in and capped,
• ›GET /api/ats/sync previews what a full ingest would cost before spending
• ›Recooty's viewed_at is trusted over our own flag, so candidates opened
• ›Resume pipeline: download → private Supabase Storage → text extracted once
• ›Stage changes seen during sync are recorded as history
• ›vendor_viewed_at records the ATS's "already viewed" signal separately from

Core values scoring

• ›docs/reference/Augusto_Core_Values_Reference.md is now the source of truth,
• ›Per-value scores (CV1-CV5) so you can see which values a candidate shows
• ›Warning signs deduct and surface as concerns — "no evidence" and "evidence
• ›Removed the keyword core-values pass. The rubric is behavioral, so matching
• ›Rubric version is recorded in every score's provenance, so editing the

Changelog & releases

• ›In-app changelog at /changelog — the version badge in the header and footer
• ›GitHub Releases published automatically on merge to main, with notes taken

Docs

• ›docs/ARCHITECTURE.md — system design and the reasoning behind it
• ›docs/SETUP.md — env vars, Supabase, Google sign-in, Cloudflare flags
• ›docs/BACKLOG.md — planned work
• ›docs/reference/ — process playbook and verified Recooty API shapes

Talent Bench

• ›Bench stat bar on dashboard — shows count of candidates on bench and total candidates, links to the bench sheet
• ›New /api/bench route reads the "bench recruting" tab from the talent bench spreadsheet

Staffing Requests

• ›Open/All filter toggle — defaults to open requests, toggle to see completed (dimmed)
• ›Compact row layout — tighter, more scannable
• ›"Request a Role" button links directly to the Google Form
• ›Assigned To shown inline on collapsed rows

Staffing Requests panel

• ›Pulls open staffing requests from the "Recruiting Requests" Google Sheet tab in real time
• ›Status dropdown (New / In Progress / Job Posted / On Hold / Complete) writes back to the sheet on change
• ›Assigned To field writes back on edit
• ›Expandable rows show full details — submitted by, duration, start date, rate, comments
• ›"Generate JD" shortcut on each row pre-fills the role title in the JD Generator

Header

• ›Sign out button (Cloudflare Access logout)

Recruiting Assistant (chat)

• ›Floating chat panel — ask questions about the recruiting process or open jobs in plain language
• ›Live Freshteam candidate search — type a name or email and get candidate cards inline with job title, stage (color-coded), email, and a direct ATS deep-link
• ›Markdown rendering in chat responses (bold, bullets, links)
• ›Suggested quick questions on first open

JD Generator improvements

• ›AI pre-fill suggestions — tab out of Role Title and Claude suggests must-have skills, nice-to-haves, and responsibilities tailored to Augusto's actual role types (AI engineers, designers using AI, n8n/LLM tooling)
• ›Suggestions are Augusto-specific: n8n, Claude, MCP, RAG, Loveable.dev, healthcare/finserv context
• ›"Why Join Us" and EEO statement automatically appended to every generated JD

Process guide improvements

• ›Static CSS swimlane diagram — replaced React Flow with a readable lane layout (no zoom/pan complexity)
• ›Larger fonts throughout — stepper, prereqs card bullet list, stage cards
• ›Actionable CTAs on key steps — Generate JD, Post Job on Freshteam, Schedule Interview buttons

Live job postings

• ›Freshteam API integration — live job panel with applicant counts, Job Posting and ATS links per role