A long archive of filed records, every decision kept where it can be found again

Build history · archived through 23 August 2026

How Supernova grew from project records to reviewed AI work.

Supernova connects project documents, client conversations, and delivery plans in one workspace. This archive covers its development from 27 July to 23 August 2026: what teams could do, what problems we found, and what changed. Its 765 saved code changes describe that period, not the current release.

Archive period

27 July–23 August 2026

Code changes in this snapshot

662 non-merge changes

Detailed product changes

59 recorded improvements

Five problems that changed the product.

These incidents explain why the product gained separate calculations, review checks, and access controls. Each lesson describes a specific improvement.

  1. Incident 01

    The estimate looked precise.

    An AI-generated estimate did not account for the amount of work in the specification.

    Problem: a precise number with incomplete inputs.

    What changed

    The estimator separates AI classification from calculations in code and shows missing inputs.

  2. Incident 02

    The reviewer corrected it once.

    A reviewer corrected a use case, but related stories and tasks still described the old behavior.

    Problem: related specifications disagreed.

    What changed

    The correction workflow identifies affected records and proposes updates for review. Coverage checks help reviewers find what remains unresolved.

  3. Incident 03

    The writer missed its own mistake.

    Asking the same agent to reread its output could repeat the original assumption.

    Problem: self-review could preserve an error.

    What changed

    Powell became a separate checking agent for generated use cases, recording findings for a human reviewer.

  4. Incident 04

    The website and product lists disagreed.

    Separate agent lists let the public description drift away from the product.

    Problem: manually maintained descriptions diverged.

    What changed

    The website and console began reading the same agent registry.

  5. Incident 05

    A hidden link looked like access control.

    Removing a navigation link did not itself stop someone opening the page's address.

    Problem: menu visibility was mistaken for authorization.

    What changed

    A shared page-permission registry and checks on the server were added. Write actions and agent tools require their own authorization checks.

The archived chapters

Read the outcomes. Expand the details.

Counts include every code commit reachable in the 23 August snapshot. Merge commits combine branches; the separate change count excludes them. Each chapter includes a summary, implementation details, and selected original commit records. Historical feature and agent counts refer to that chapter's date.

Oldest first

12 chapters · 765 commits

  1. Chapter 01 · 27–28 July

    21 commits · 17 non-merge changes

    Teams could manage a project in one workspace

    The first two days brought people, project documents, milestones, and work logs together behind sign-in.

    Read implementation details and saved changes
    What changed
    The application gained sign-in, project teams, notifications, delivery checkpoints, work logs, and developer profiles. AI could draft milestones from contract documents.
    How it was built
    Uploads were labelled by document type so milestone extraction could use the contract. Shared date formats, action feedback, and activity records made the pages consistent.
    Why it mattered
    The product first needed reliable records of who was involved, what was agreed, and what work had happened.

    Detailed changes

    Recorded in this period

    3 improvements

    1. 01
      Project-scoped console

      Authentication, project navigation, contextual headers, team management, and role visibility established the operating shell.

    2. 02
      Documents became inputs

      Typed uploads separated proposals, agreements, and supporting files so extraction could read the right authority.

    3. 03
      Delivery evidence

      Milestones, checkpoints, work logs, action feedback, notifications, and audit events gave decisions a durable record.

    Read 3 selected original commit records
    1. 383b481Build Supernova RMS console: landing page, auth, project-scoped dashboard
    2. 08e5df4Extract milestones from project documents with Claude
    3. d676a76Add checkpoints, work log, developer profiles with role visibility; borderless design
  2. Chapter 02 · 29–31 July

    63 commits · 45 non-merge changes

    Specifications gained links and review steps

    Use cases, user journeys, roles, and the data model were connected so reviewers could see which decisions later work depended on.

    Read implementation details and saved changes
    What changed
    Discovery, the part of Supernova used to specify a project, gained source references, a diagram of data relationships, journey maps, and approval steps.
    How it was built
    Generated records linked to their source records. Reviewers could approve individual items and discuss changes in a shared feedback thread.
    Why it mattered
    Related specifications can disagree. Visible links and review steps help teams catch those disagreements before building from them.

    Detailed changes

    Recorded in this period

    4 improvements

    1. 01
      Complete use-case descriptions

      Use cases gained a structured description of the actor, goal, normal flow, exceptions, and outcomes, with links to project scope and milestones.

    2. 02
      Approval steps between specifications

      Later specifications began waiting for review of the records they depended on.

    3. 03
      Review where the work lives

      Per-page feedback evolved into one global refinement thread with item context and explicit operator approval.

    4. 04
      An explorable data model

      A diagram showed data entities and their relationships, with pan, zoom, search, and export controls.

    Read 3 selected original commit records
    1. 3585ffdImplement the gated Discovery flow: Documents -> Use Cases -> Personas -> Journeys
    2. 0f87035Build the Ontology stage as a real ERD and extend the gate chain to six stages
    3. 18bef23Hold the console chrome still: fixed shell, one scroll container, no nav animation
  3. Chapter 03 · 1–2 August

    40 commits · 36 non-merge changes

    Sales and delivery could use the same project records

    Delivery boards, sales work, and the catalogue of past projects gained connected views of the same evidence.

    Read implementation details and saved changes
    What changed
    Backlog and sprints became a real swimlane board, project dossiers gained work and team evidence, sales demand became actionable, and capabilities became a navigable Company Brain.
    How it was built
    Pages began sharing data. Lists gained pagination, project links opened the appropriate view for their status, and views reflected each person's role.
    Why it mattered
    The same project must make sense to the team delivering it, the seller representing it, and the leader looking for reusable proof.

    Detailed changes

    Recorded in this period

    4 improvements

    1. 01
      Backlog and Sprints

      Stories, dev tasks, sprint assignment, acceptance criteria, comments, capacity, and swimlanes became one delivery board.

    2. 02
      Project reference pages

      The catalogue connected clients, projects, industries, and delivered capabilities so teams could find relevant examples of past work.

    3. 03
      Sales demand loop

      Requirements, fulfilment stages, interviews, client ownership, and project handoffs joined sales to delivery.

    4. 04
      Hard product laws

      Pagination replaced silent caps, project status determined the primary view, and commercial amounts left every console role.

    Read 3 selected original commit records
    1. cfdf7a5Move Backlog & Sprints into its own nav item, spanning every milestone, with Dev Tasks built in
    2. a0f73a7Pagination everywhere, role/status-shaped project views, no-revenue rule
    3. 9ef72feCatalogue Domains + Capabilities pages with dossiers on the imported tables
  4. Chapter 04 · 3–4 August

    82 commits · 68 non-merge changes

    The console connected to the work

    Teams could consult customer records, email, meetings, and code history alongside their project work.

    Read implementation details and saved changes
    What changed
    Zoho, Microsoft Graph, ProShort, and Git data began feeding Daily, dossiers, meetings, inbox, deals, leads, accounts, and work logs.
    How it was built
    Connectors imported new records and could resume interrupted imports. Ownership checks controlled access, and integrations retained links to the source records.
    Why it mattered
    Teams and AI assistants needed the conversations and activity behind a summary to understand commitments and changes.

    Detailed changes

    Recorded in this period

    5 improvements

    1. 01
      Zoho CRM

      Incremental owner pulls, related lists, stage history, contacts, attachments, and account records landed in one database.

    2. 02
      Microsoft Graph

      Mail and calendar gained paged backfill, delta sync, resumability, attachments, thread grouping, and mailbox ownership.

    3. 03
      ProShort calls

      Recordings and transcripts became first-class evidence for meetings, deals, briefings, research, and commitments.

    4. 04
      Git as delivery evidence

      Project repositories, author mapping, commit heatmaps, searchable messages, and audit linkage joined work logs.

    5. 05
      Daily dossiers

      Inbox, meetings, deals, leads, and accounts moved from flat mirrors to connected timelines and next-action surfaces.

    Read 3 selected original commit records
    1. 695c2abMicrosoft Graph connector: email and calendar, delta-synced
    2. 7e0b2ebProShort sync service: recordings land in nexus-native tables
    3. 365e74eOne canonical scope-item list now backs both Milestones and Interfaces
  5. Chapter 05 · 5 August

    81 commits · 59 non-merge changes

    A sales handoff could lead to a saved build plan

    The handoff from sales to delivery and the specification workflow gained saved records, generation, and review controls.

    Read implementation details and saved changes
    What changed
    Both halves of handoff went live. Use cases, personas, ontology, design language, roles, backlog, journeys, canvas, and dev tasks stopped falling back to mocks.
    How it was built
    Generators followed the order of their source dependencies. Database checks enforced review steps, and a separate review checked that results were saved and controls performed their stated actions.
    Why it mattered
    A team must be able to reopen a plan and find the reviewed records it approved.

    Detailed changes

    Recorded in this period

    4 improvements

    1. 01
      Deal to kickoff

      Sales preparation, transcript checks, meeting scheduling, Tech Lead notification, acceptance, and project creation became one persisted handoff.

    2. 02
      Nine generated artifacts

      Interfaces through dev tasks were generated from the contract in dependency order and persisted behind review gates.

    3. 03
      Honest system state

      Local-storage gates, mock fallbacks, placeholder task data, and fake loading states were replaced with database truth or explicit emptiness.

    4. 04
      Adversarial integration

      Parallel builders were followed by dedicated review passes for broken persistence, misleading controls, and surfaces that only looked connected.

    Read 3 selected original commit records
    1. eb91bb3Give the Tech-Lead-owned half of the Deal-to-handoff flow a real backend
    2. 7d9a2a8Make Journeys + Canvas real, the last mock Discovery artifact
    3. 96e7f6aDev tasks are real, and every mock fallback becomes an honest empty state
  6. Chapter 06 · 6 August

    95 commits · 79 non-merge changes

    Reviewers could check what the specifications covered

    Generated specifications gained links to source requirements, coverage checks, and editing tools.

    Read implementation details and saved changes
    What changed
    Use cases gained progressive elaboration, journeys gained step-level traceability, the ontology grew schema discipline, and reviewers could edit or refine the generated record directly.
    How it was built
    Coverage checks linked requirements to their generated records. Large jobs ran in smaller batches to avoid cutting output short, and data entities linked back to the use-case steps that needed them.
    Why it mattered
    A long specification can still omit a required behavior. Reviewers needed to see those gaps directly.

    Detailed changes

    Recorded in this period

    5 improvements

    1. 01
      Review before adding detail

      Reviewers first checked a use-case outline. Further generation added normal steps, rules, outcomes, failure paths, and links to the related stories.

    2. 02
      Coverage as data

      Scope items, use-case steps, entity nouns, screens, and journeys gained explicit claims and audits for anything left uncovered.

    3. 03
      Ontology v2.1

      Entities gained table names, scope roots, volume classes, closed enum rules, field inventories, and step back-links.

    4. 04
      Journey v3

      Stages gained emotion rationale, source-grounded pains, service-blueprint backstage actions, entity citations, and step-level evidence.

    5. 05
      Multimodal refinement

      Reviewers could ask questions, edit records, and attach PDFs or images to a versioned cascade over the full project memory.

    Read 3 selected original commit records
    1. affc4f9The use-case tier becomes progressive elaboration
    2. 2c2d265Feedback now makes real, versioned changes across the whole Discovery pipeline
    3. a3896a5The ontology lands at v2.1 with every mechanical audit clean
  7. Chapter 07 · 7–9 August

    129 commits · 115 non-merge changes

    Specifications helped teams decide what to build first

    Screen flows and shared development tasks connected the plan to implementation. Research agents gained access to source conversations.

    Read implementation details and saved changes
    What changed
    Journey diagrams began showing screens and the steps between them. Developer plans separated shared infrastructure from feature tasks, and the agent directory showed recorded work.
    How it was built
    Shared screen and entity registries removed duplicate invention. Agents read source mail and transcripts, returned schema-constrained output, wrote telemetry, and were blocked from persisting incomplete runs.
    Why it mattered
    Knowledge becomes valuable when it changes a decision: what to build first, who must respond, which risk is real, and what evidence supports the next move.

    Detailed changes

    Recorded in this period

    5 improvements

    1. 01
      Screen-flow diagram

      A shared list of screens linked use-case steps to the pages and data they needed, making missing coverage visible.

    2. 02
      Foundation-aware dev plans

      Platform tasks owned shared entities once, story tasks depended on them, and audits refused duplicated migrations or unowned schema.

    3. 03
      Grounded research agents

      Al, Brain, and Byerley read real mail, transcripts, CRM history, commitments, and sourced web evidence before recommending a move.

    4. 04
      Agent operations

      The Robot Universe exposed the real roster, run counts, model telemetry, output guards, and links back to the work each agent produced.

    5. 05
      Proposals and case studies

      Deal evidence could become a reviewed proposal, and completed delivery evidence could become a reusable case-study deck without exposing revenue.

    Read 3 selected original commit records
    1. b74300aJourneys become wireflows
    2. 5fb7197Dev tasks get a foundation pass
    3. ea3f7d5Byerley joins the roster and the Robot Universe gets a page
  8. Chapter 08 · 10–11 August

    130 commits · 122 non-merge changes

    Salespeople could see what needed attention

    Sales pages began leading with overdue commitments, stalled work, and suggested next actions.

    Read implementation details and saved changes
    What changed
    Tasks could be completed from the workspace. Account and deal pages led with a useful summary, while maps showed the broader pipeline.
    How it was built
    Rules selected the next action, supported edits updated the source system, and missing data was labelled. Reviews included the running pages in Chrome.
    Why it mattered
    People needed a clear route through what was late, blocked, promised, or ready to advance.

    Detailed changes

    Recorded in this period

    5 improvements

    1. 01
      Needs-you-now

      Deals, leads, accounts, meetings, and the home briefing began leading with overdue promises, stalled work, and the next deterministic action.

    2. 02
      Write-through work

      Completing tasks, fulfilling commitments, linking CRM records, refreshing deals, and sending approved drafts updated their source systems.

    3. 03
      Grounded assistance

      Norby and the research workers answered from the current dossier, confessed missing evidence, redacted generated amounts, and logged every run.

    4. 04
      Leadership maps

      Pipeline dots, owner-status lead bubbles, trends, scrum, and KPIs turned thousands of rows into explainable operating pictures.

    5. 05
      Regeneration governance

      Expensive reruns became captured requests with a named requester, reason, project, operation, approval state, and audit event.

    Read 3 selected original commit records
    1. 680d101The Daily task view became a to-do list
    2. a5bc92aReference-app parity lands with grounded research and Ask Antino
    3. f6e6114The AE estate finishes with verdict-first dossiers and needs-you-now leads
  9. Chapter 09 · 12–15 August

    19 commits · 18 non-merge changes

    Reviewer corrections could update related developer tasks

    The correction workflow expanded to developer tasks, while role and story relationships became more explicit.

    Read implementation details and saved changes
    What changed
    The role model was rebuilt, dev tasks gained complete story relationships, and the feedback cascade expanded through its ninth artifact type.
    How it was built
    Reviewers settled conflicting notes and agreed shared terms before applying updates. Existing task identities were preserved, and integrity checks found broken links.
    Why it mattered
    Correcting a use case is incomplete if its old assumption survives in related stories, the data model, journeys, or tasks.

    Detailed changes

    Recorded in this period

    4 improvements

    1. 01
      Role-based access

      VP, PM, sales, CXO, and developer reach was rebuilt around actual work, with Discovery genuinely read-only where required.

    2. 02
      Many-to-many traceability

      Dev tasks could cover many stories, every task regained a real parent, and finished work named its sprint and completion state.

    3. 03
      Updates to related records

      Reviewers resolved conflicting notes and agreed terminology before the correction workflow updated dependent specifications.

    4. 04
      Dev tasks join the cascade

      The ninth artifact type became update-only, preserving task identity while applying reviewer consequences to what developers build.

    Read 3 selected original commit records
    1. 9461694Ship the dictated role-based access rework
    2. 584af4bEvery dev task now has a real parent story
    3. ab1ae38Add dev-task cascading as the missing ninth artifact type
  10. Chapter 10 · 17–20 August

    35 commits · 35 non-merge changes

    Agent runs gained shared checks and clear outcomes

    Verification, tool permissions, run records, and feedback review became reusable parts of the agent infrastructure.

    Read implementation details and saved changes
    What changed
    Powell checked generated use cases against source material. Donovan proposed reusable lessons from corrections. The roster used one registry, and generation jobs gained explicit outcome reporting.
    How it was built
    A separate agent checked generated work. Code enforced tool access and run limits, incomplete output was distinguished from success, and proposed lessons were reviewed against previous decisions.
    Why it mattered
    Returning text does not establish success. Teams need to know whether a run completed, what it changed, and what still needs review.

    Detailed changes

    Recorded in this period

    5 improvements

    1. 01
      Powell verifies

      A separate checker compared use cases with real source material and wrote every finding to a review ledger before human approval.

    2. 02
      Corrections become lessons

      Repeated, reviewer-confirmed mistakes could become proposed checking rules only after replaying them against historical decisions.

    3. 03
      Shared agent infrastructure

      Agents gained enforced tool permissions, run limits, activity records, output counts, and explicit completed, failed, or incomplete outcomes.

    4. 04
      GENESIS and CASCADE

      The one-time build and recurring feedback operation received named runbooks, thematic batching, coverage accounting, and final integrity checks.

    5. 05
      Canonical roster

      Marketing and operations began reading one agent registry, removing invented workers and catching implemented agents missing from the ledger.

    Read 3 selected original commit records
    1. 935bbc6Powell: a verifying agent that checks use cases against real source material
    2. 090e0d9A shared harness for the plain-SDK agents
    3. e1780e8Review feedback now reaches the dev tasks a developer actually builds from
  11. Chapter 11 · 21–22 August

    49 commits · 47 non-merge changes

    Developers could use Supernova from their coding tools

    A connector brought project context into developer tools, while the estimator separated AI classification from calculations in code.

    Read implementation details and saved changes
    What changed
    The MCP server gained real sign-in and role-filtered tools, the estimator separated classification from arithmetic, and the public application received its own layouts and documentation structure.
    How it was built
    The Model Context Protocol (MCP) connector applied permission checks and reused the application's record-saving code. Estimates matched a fixed effort catalogue, and documentation separated current behavior from historical changes.
    Why it mattered
    People using external tools need consistent access rules. People reviewing estimates need to inspect the inputs behind the total.

    Detailed changes

    Recorded in this period

    10 improvements

    1. 01
      MCP moves to Microsoft Entra

      The shared connect token was retired. Claude Code now signs in through Outlook, and the server validates tenant, issuer, audience, scope, signature, and immutable identity.

    2. 02
      MCP survives adversarial review

      Independent security, authorization, transport, deployment, and protocol reviews found rollout blockers, privilege gaps, malformed-body risks, and discovery failures before wider use.

    3. 03
      MCP matches console authority

      Tool discovery is filtered by role, project status rules match the browser, and whoami explains which capabilities are withheld instead of advertising tools that will fail later.

    4. 04
      Project context in developer tools

      Tech Leads could read specification status, use cases, data models, milestones, and teams from their coding tools. They could also generate additional use-case detail through the same record-saving code as the portal.

    5. 05
      Estimator separates judgement from arithmetic

      A model may classify scope; deterministic code owns quantities, learning curves, overlap, uncertainty, and totals. The two paths cannot silently exchange jobs.

    6. 06
      Three-pass classification

      Inventory, classification, and convergence became separate passes after one-pass runs answered the same document differently and trusted their own claim of completion.

    7. 07
      Fixed catalogue matching

      Open-ended feature invention was replaced by one match against a priced catalogue. Shared features are unioned once, not multiplied across capabilities.

    8. 08
      Scale becomes explicit

      Screens, integrations, workflows, agents, entities, and other countable signals override typical catalogue quantities, while every fallback remains visible as a scale gap.

    9. 09
      Evidence-backed estimate surfaces

      Uploaded scope files, the estimates list, project estimate pages, editable capabilities and scale, deterministic repeat runs, and downloadable documents became one product flow.

    10. 10
      Back-tests stay honest

      Delivered-project evidence exposed where the estimator was precise but inaccurate. The failure was recorded and the missing scale signal fixed instead of tuning away the discrepancy.

    Read 3 selected original commit records
    1. 47a0425MCP tools are now filtered to what the caller's role permits
    2. 49996f9Price a scope by matching a fixed catalogue
    3. c5a6930Documentation reads as a set of documents again
  12. Chapter 12 · 23 August

    21 commits · 21 non-merge changes

    The public site explained the product and its reviews

    Public pages gained walkthroughs of delivery and review, while signed-in navigation made each person's workspace clearer.

    Read implementation details and saved changes
    What changed
    Eight marketing routes gained distinct story structures, console URLs began naming their role space, sidebars became role manifests, and the estimate page led with the decision and its evidence.
    How it was built
    The existing Antino palette and type system stayed fixed while each public route gained a subject-specific visual metaphor. Chrome verification covered desktop and mobile with complete data intact.
    Why it mattered
    A Company Brain is difficult to believe when its website reads like a feature inventory. The public explanation needed the same provenance, consequence, and human accountability as the product.

    Detailed changes

    Recorded in this period

    5 improvements

    1. 01
      Eight public stories

      Home, Pipeline, Trust, Agents, FAQ, Changelog, Privacy, and Terms gained consequence-led narratives inside one Antino brand system.

    2. 02
      One real agent universe

      The public operating model leads into the complete 29-agent roster sourced from the same registry as the console.

    3. 03
      Evidence-first estimating

      Estimate detail leads with the decision, then keeps source-backed matches, inference, overrides, and every module row reachable.

    4. 04
      Role spaces

      URLs and sidebar manifests now name the viewer's space directly, while server-side page gates remain the actual authority.

    5. 05
      Distinct visual systems

      Each marketing route gained its own metaphor and pacing without introducing a new palette or abandoning the existing design universe.

    Read 3 selected original commit records
    1. c1bfd04The public site has pages for its agents, pipeline, questions and changelog
    2. ffa251aEvery console URL names the role space it belongs to
    3. 2e6be48Every marketing story now has its own visual system

See how a review works today.

Read a review example →