You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This folder is BioRouter's archive. Every document in it describes work that is finished: designs that were built, plans that were executed, review campaigns that closed, stress tests that ran to completion, and a small number of ideas that were abandoned or later removed. It is kept for the record and for provenance: read it to find out why a subsystem is shaped the way it is, never to find out what the code does today. The material spans May 2026 to October 2026, and almost all of it shipped.
The three exceptions are worth knowing up front, because they are the ones most likely to mislead. Dashboard mode was built through four generations and then removed from the desktop app on 2026-07-18 — nothing in dashboard-mode/ describes a feature that still exists. MCP apps was never a feature this project chose: it was present from the first commit and was removed on 2026-09-08, so mcp-apps-removal/ is likewise an archive of something gone. The extension trait design in legacy-architecture/ was never shipped either; extensions are MCP servers built on the rmcp SDK instead. Everywhere else, a document that has been overtaken says so in its own header and names its successor.
Come here when you are tracing a decision, reconstructing what landed in a release, or trying to understand why an identifier like BR-43 or H7 appears in a commit message. If instead you want the current behaviour of a subsystem, you are in the wrong folder and should leave quickly: the live working documents live in docs/apps-sdk/, docs/agent-loop/, docs/knowledge-base/, docs/design/ and their siblings. Several of those READMEs point back here deliberately, for exactly the provenance question this folder answers.
How to check a document's standing. Every file here opens with a context header carrying a Status: line that states whether the work shipped, was superseded, or was removed, and — where it was superseded — links the document that now holds the truth. Trust that line over the filename.
No documents sit directly in this folder apart from this index. All content sits in the topic subfolders below. One of them, agent-loop-campaign/, carries a sub-area of its own, cross-platform/ (the campaign's Windows and Linux arm), indexed by that campaign's README as well as here.
The method statement and results of a campaign that drove a real model to build 100 diverse agentic apps with the Agent Drafter. Ran to completion; the eight H1–H8 drafter fixes it produced shipped on 2026-07-11 in commit 679deed5.
The archived evidence set for a separate test drive against a 100-app specification corpus — per-app rubrics, three cross-cutting audits, and a six-wave remediation plan. Authoring stopped at spec 025 of 100; the remediation shipped on branch feat/apps-sdk-v2.
The June 2026 strategy RFC that benchmarked OpenAI's Agents SDK and proposed a layered App SDK, plus its code-level companion implementation design. Largely built, and superseded by the Apps SDK v2 spec of 2026-07-12.
October 2026 local-inference validation concluded 2026-10-12 UTC (2026-10-11 PDT), recording native Ollama context/tool-call fixes, managed Llama Server hardening, scoped checks and retained model-quality failures. Implementation is in unmerged PR #393; no merge or release is claimed.
The executive report of the agentic-loop review — a walkthrough of the loop, answers to 14 review questions, a comparison against eight open-source coding agents, and 27 sub-reports. A point-in-time snapshot read against commit 24cdc3a2 on 2026-07-12, now superseded by the campaign it triggered.
The plan of record for the BR-1…BR-70 fix campaign that implemented the review's findings: wave conventions, regression gates, and a dated log of every merge decision. Finished and merged to main across 86 BR-prefixed commits.
That campaign's Wave 3 cross-platform arm, kept as a sub-area: the parity audit that coined the GAP-N findings, the BR-68 command-safety and BR-70 CI-gate designs, the superseded BR-64 macOS Seatbelt design, and the cluster's verification report. The cluster merged Gate 3 GREEN; three lower-ranked GAP findings appear to remain open.
The July 2026 campaign that closed every open GitHub issue in one pass — the plan of record for issues #18–#43, the GUI vision pass that verified the UI-visible fixes over CDP, and the parallel stress test run against the result. All closed except workspace control (#30), which was plan-only then and landed later as its own campaign.
The six-surface privacy audit run against branch feat/privacy-tiers on 2026-08-06 and its complete synthesis, re-run 2026-08-07 with the two payloads the first synthesis never received and re-checked against the tree at bbcfdb06. The audit finished; issue #56 was still in flight. Sixteen findings had already been closed by then, twenty-three remained open, and three were produced by the fixes themselves. Its register uses C-nn / O-nn / N-nn identifiers.
A whole-app latency review against v1.86.0 and its implementation log, plus a second, independent effort comparing BioRouter against the third-party jcode harness (analysis, report). Nine fixes merged 2026-06-23 and the jcode waves 2026-06-24; the review's Tier 2 and Tier 3 items were never scheduled.
The 2026-07-18/19 campaign on branch feat/streaming-tool-call-ui: an investigation into why a tool card appeared late and already finished, the streaming work it produced across fourteen previously non-streaming providers, and the three QA rounds over the result. The streaming track merged to main at 78471bdc; the QA-campaign fixes were held for a separate merge decision. Its findings register uses R1-NN/R2-NN/R3-NN identifiers.
Defect notes left by the August 2026 documentation pass over BR-71 workspace control — findings that belong to source rather than prose, recorded because that pass was documentation-only and fixed no Rust. The first, the CLI plural alias defect, was still open on 2026-08-03.
The campaign that designed, built and tested Crew, from the first research on 2026-09-21 to its merge into main on 2026-09-27 (PR #366): the implementation plan and its §16 redesign and naming scope, the status ledger and handoff, the source investigations and independent reviews, four live QA rounds, the bounded evidence runs behind each acceptance claim, and the merge QA and three fix waves that followed (PR #377). Shipped. Its designs are living documentation under crew/design, where the rows later fixes changed are marked in place.
The complete paper trail of dashboard mode — four generations of design specs and plans for spawning many chat windows on a free-floating canvas — and the record of its removal. The feature was deleted from the desktop app on 2026-07-18; the whole folder is an archive.
The record of MCP apps, a feature that let a third-party MCP server run its own interactive HTML in a sandboxed iframe inside BioRouter, and the survey it was removed from. Present since the first commit on 2026-01-20 and never requested; removed on 2026-09-08 after all six ecosystem extension repositories were confirmed not to use it. Not to be confused with Agent Drafter's Built apps, which stay.
The adversarial judgement of three competing chat-groups designs and the reduced plan authorised on 2026-07-16. Its scope was overtaken: the branch went on to ship both the tabs and the split with drag and drop zones that this document deferred.
The design for the custom Electron application menu and the system tray's click behaviour, with its implementation plan. Written 2026-05-07 and implemented as buildApplicationMenu(); the live menu has since drifted from the listing.
The implementation plan for the four fixes batched into v1.72.1 — runtime enforcement of disabled skills, plus a drop-zone-only redesign of three import modals. All four shipped in May 2026.
The design for NotificationSurface, a shared primitive owning the layout of both transient toasts and inline alerts. The primitive shipped; the follow-up it deferred — migrating roughly 40 ad-hoc inline banners — should be assumed still open.
Two records from one QA session run on 2026-06-24/25 against build 1.86.1: an item-by-item issue tracker and a regression pass over the 133 commits of the preceding eight days. Neither is a checklist to re-run.
The two view-level redesigns from the July 2026 desktop overhaul — the Home page and the Knowledge view — each with an HTML companion carrying its mockups. Both signed off (2026-07-08 and 2026-07-10) and shipped; the overhaul's app-wide specification and its still-open status record stayed in docs/design/ui-overhaul/.
The 100-request scenario corpus for stress-testing the Auto Visualiser's render_dashboard composite-report tool, with per-run outcomes. Executed against the dev app driven by GPT-5.5 and completed 100/100 on 2026-07-11.
The design for CodeGraphAgent — a .brxt extension wrapping a vendored CodeGraph engine — and the two plans that built it: a foundation plan and a bio-language extractors plan. Shipped; the extension now covers 23 languages including R, Julia, MATLAB and Perl. Note that the code lives in a separate repository.
The August 2026 campaign that retired the standalone biorouter-headless binary and its Linux tarball, folding browser access into biorouter serve and making the daemon serve the interface itself. Shipped; release assets dropped 11 → 10. Its decisions are living documentation under deployment. Also the record of three pre-existing defects it walked past, one of which — the Linux GUI package shipping Windows binaries — is still open.
The design and plan for the two UCSF Versa providers and for splitting the provider grid into labeled sections. Both providers shipped and exist in the tree; the section ordering specified here is superseded by the live ProviderGrid.tsx.
The design and plan for sending pasted or dropped images to vision-capable models as inline base64 blocks, adding a supports_vision flag and fixing Gemini's silent image drop. Shipped in v1.76.1.
Two design-and-plan pairs from 2026-05-07: .brxt bundled skills (plan), letting an extension package carry its own skills, and skill bundles (plan), treating a parent folder of sub-skills as one installable unit. Both shipped.
The origin design for the personal knowledge base feature — git-backed markdown KBs maintained by an LLM, credibility classification, graph derivation, the MCP and HTTP surfaces, the .brkb format. Approved 2026-05-30; the feature shipped.
The six implementation plans that built the feature out were executed in order against that design. That folder's own index describes each in full; they are summarised here because they are cited by number from commit messages and from the live Knowledge documentation:
The execution record of the July 2026 branch and pull-request merge campaign: decisions D1–D10, conflicts resolved, and a commit-level inventory of everything that landed. Completed 2026-07-13 with all nine examined pull requests merged.
The design and plan for merging the hand-written documentation/ folder and the Docusaurus-generated docs/ site into one plain-markdown tree, replacing the old pages' branding with BioRouter's own and renaming recipe to workflow. Executed in May 2026; docs/ was reorganized again in July 2026, so the target paths are a record of intent rather than links to follow.
Two superseded internals designs. The extension trait design was never shipped — extensions are MCP servers on the rmcp SDK instead. The agent error modelwas built and its two-tier policy still governs the loop, but AgentError no longer exists; the folder index points at the source for the current types.
Data files, scripts and rendered pages
This folder holds no non-Markdown files at its top level, but four subfolders carry machine-readable evidence, two carry scripts, and one carries rendered HTML. Each is described by its own folder's index.
agent-drafter-testdrive-100/data/ — the machine-readable authoring ledger, the platform-integration audit output, and the static evidence for the five layout probes. agent-drafter-testdrive-100/authoring-logs/ holds the per-app static-analysis JSON and browser issue logs from the same run. The sibling app-results/ and layout-probes/ folders are Markdown, not data: they hold the per-app rubrics and the layout probe write-ups.
agent-loop-review/generate_review_html.py — a Python script that stitches the review's Markdown corpus into one self-contained HTML report using pandoc. The generated review.html is not checked in; the Markdown corpus remains the source of truth, so you never need to run this to read the review.
biorouter-crew/*.json and biorouter-crew/smoke/: the raw results of the Crew build's institutional and AWS probe runs, and the small synthetic probe scripts that produced them and the later local runs. Not production Crew code.
ui-overhaul-2026-07/home-screen-redesign.html and ui-overhaul-2026-07/knowledge-view-redesign.html — the rendered mockups for the two archived view redesigns. These must be opened in a browser to be useful; each has a Markdown companion beside it carrying the same reasoning and values in text.
Related documentation
Crew design and reference: the living design documents whose build record is the biorouter-crew folder here.
Apps SDK — the current authority on app and SDK behaviour, which supersedes the June 2026 RFC archived here.
The agent loop — the live documentation for the loop that the review and fix campaign in this folder diagnosed and rebuilt.
Knowledge base — the live working documents for the Knowledge subsystem, whose build record is the knowledge-base-buildout folder here.
Design — the current design specifications, including the UI overhaul and chat-groups work that overtook the archived chat-groups plan.