Designer round 4: all categories >=8, zero high or medium findings, all
seven round-3 fixes confirmed with measurements. Residual lows fixed for
the confirmation round: insights lens chips, guide tabs, and modal text
inputs to 44px on mobile; at <=430px the share dialog's destructive
Revoke takes its own row instead of sitting 9px from Done.
All responsive CSS on the same URLs — no mobile-specific routes.
42 e2e specs green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
Designer round 3 FAILED the bar (6/6/7/9/6, one high) — it caught a
regression my round-2 fix introduced: .ai-main { overflow-wrap:anywhere }
let flex min-content collapse invite/share URL rows to one character per
line at 360/390 (496px-tall rows). Streak reset. Fixes:
- H1: on mobile the admin row's name/URL takes the full row
(flex-basis 100%; controls drop below) and break-word replaces
anywhere, so URLs wrap readably at natural break points
- M1: onboarding inputs collapsed to 18px (.ob-row goes column on
mobile, so flex:1 became flex-basis:0) — now flex:none, 44px
- L1: overflow-menu items 44px; L2: server auth pages get 44px
inputs/buttons at <=900px; L4: .markdown gets width:100% so short
docs stop shrink-to-content floating; L5: .hmeta centers so the
size stays attached when the author line wraps
go build/vet/test green; 42 e2e specs green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
Designer round 2: layout 9 / readability 8 / tap-targets 8 /
navigation 9 / polish 8, zero high or medium findings; every round-1
fix confirmed with measurements (no page-level horizontal scroll on any
of the 24 surface x viewport combinations; desktop 1360 unaffected by
the 900px breakpoint).
Remaining low cosmetics fixed for the confirmation round:
- .hsize never wraps ('67 B' split across lines at 360)
- guide Copy button is a real 44px touch target (code blocks get the
height to hold it), admin selects/buttons 40 -> 44px
- truncated emails/URLs wrap on touch instead of dead-end ellipsis
- history timestamps 12px (11px was borderline on small phones)
42 e2e specs green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
Designer round 1 scored layout 5 / readability 7 / tap-targets 6 /
navigation 8 / polish 7, one high finding:
- HIGH: tablet 768 and phone-landscape 844 fell between breakpoints —
full desktop topbar + fixed 264px sidebar overflowed the page
(scrollWidth 849 vs 768). The mobile media query now covers
max-width 900px, giving those widths the off-canvas sidebar and
collapsed topbar.
- MED: admin-row selects (28px) and .ai-btn/.ai-del (27px) bumped to
40px targets on mobile, rows wrap; modal buttons to 44px; share
dialog's destructive Revoke pushed away from Done; Escape now
dismisses the share dialog.
- MED: <=430px drops the verbose .dl-meta so filenames stop truncating.
- LOW: gd-copy gets an opaque backing over scrolling code; compact
nowrap history timestamps; invite/pbtn/palette rows to 44px.
42 e2e specs green.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
Field report: an agent session read a dozen-plus files via Bash
(grep/cat/tail/find) and Grep, but read heat showed almost nothing —
the read-log hook only matched the native Read tool.
- matchers broadened per platform: claude Read|Grep|Bash, codex
read_file|shell, gemini +search_file_content|run_shell_command,
hermes read_file|grep|bash; plugin hooks.json matches Read|Grep|Bash
- read-log is now tool-aware: shell events mine the command line for
existing files it names (redirect targets and flags excluded), grep
events mine the response for the files the matches came from
(content lines and filename lists), and listing tools (Glob, ls) are
deliberately ignored — seeing a file's name is not reading it
- hooks install upgrades a registered hook's stale matcher in place, so
re-running it after a binary upgrade rolls coverage out to existing
projects instead of being skipped by the idempotency marker
- docs: SKILL.md platform table + read-heat wording, install/init
commands, README command table
Note from the same report, verified not a bug: read-log resolves the
mount via the folder's own .bdrive/config.json (config.ResolveMount),
so a stale duplicate registry entry cannot swallow reads.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
- #root { display: contents } so body's flex layout sees through the
React mount point (main pane was stuck at content width)
- README: Web frontend development section; CLAUDE.md: commands +
webapp description rewritten for the React/Vite reality (committed
dist, check-dist pre-release step, the no-router-library and
no-post-commit-DOM-patching lessons)
- verified: clean git-archive checkout builds with no Node and serves
the React app; visual parity pass across desktop + mobile surfaces;
42 e2e specs green; plugin docs untouched (no frontend internals)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
- org admin: rename, member roles/removal (self marked), project
rename/delete, invite links (create/copy/revoke, uses + expiry),
org-wide public-share audit with revoke; members get a read-only view
- hub settings: verification/approval policy toggles (verification
disabled without SMTP), read-only domains/self-signup/admins, pending
signup queue with approve/deny (count feeds the admin bar)
- panels replace the content pane without becoming routes (classic-app
parity): Browser takes a panel prop, HubApp owns the state and any
navigation closes it
- e2e: 8 admin specs, all mutations self-reverting (42 total green)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
- project home at /<pid>: connect-an-agent guide (Claude plugin flow,
Hermes/Codex CLI tabs, persisted choice, copy buttons, real hub origin
+ project id), with Insights embedded below for admins/org-owners
- Insights: squarified treemap, reads×freshness scatter with danger
quadrant, hot-path list, agent coverage matrix — classic math ported
verbatim into JSX SVG; dedicated /insights route
- history views: whole-project / subtree / per-file feeds with kind tags,
device attribution, expandable linkified session notes
- REPLACED react-router-dom with src/nav.ts (~40-line synchronous history
router): v7 wraps navigation in startTransition, which left the old
view on screen for seconds after the URL changed (flaky navigations,
35KB heavier bundle). Routing semantics are unchanged.
- e2e: 12 new specs porting the original 17 parity checks (34 total,
~13s, stable across repeated runs)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
- FileTree with fold state, lone-root auto-open, reveal-on-deep-link;
folder listings with heat dots and the journals-backed change feed
- FileView: markdown (HTML transformed BEFORE render, link clicks
delegated — React re-applies dangerouslySetInnerHTML markup on
unrelated updates, so post-commit DOM patching loses handlers),
images, text, download card
- breadcrumbs, per-route scroll restoration (location.key memo)
- topbar actions: share dialog (mint/copy/open/revoke), history/upload/
download buttons, ⋯ overflow menu; upload via upload/init direct or
relay path, then tree refresh + open
- ⌘K command palette: fuzzy files/projects/actions with stemming
- e2e: 11 new browse specs (22 total green in ~12s); helpers cache one
session cookie per identity to stay under the 10/min auth rate limit
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
- URL as source of truth: parseRoute/urlForPath/urlForView ported verbatim
(src/router.ts), single catch-all route so encoded slashes survive
- hub shell: project nav with color chips, org bar, admin bar with pending
count, sign-out per session flags; volume shell renders too
- empty-state onboarding (invite paste + create project), /join/<token>
invite accept that survives the login redirect
- toast + modal prompt/confirm primitives (imperative promise API over a
React host, matching the classic behavior)
- e2e: 7 new hub specs (11 total green); harness gains a no-org 'solo'
account to reach the empty state
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
- Vite+React+TypeScript workspace at internal/webapp/frontend; vite build
emits committed assets into internal/webapp/static (the go:embed target),
so plain 'go build' still needs no Node
- style.css and the SVG sprite ported verbatim; boot layer (api/config,
401->login redirect) ported so the auth flow works end to end
- frontend(): content-hashed assets/* now served immutable; index.html
stays no-cache (TestFrontendSPAFallback covers both)
- committed e2e harness (BDRIVE_E2E_SERVE=1, deterministic seeded hub on
:8993) wired as Playwright webServer; 4 shell specs green
- check-dist.sh guards against stale committed build output
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
Replace the old "append conventions to CLAUDE.md" step with a two-file
pattern for teaching agents a newly mounted shared folder:
- <shared>/AGENTS.md (synced): the team's single source of truth for the
folder's structure and conventions — scaffolded once by the project
creator, read (never rewritten) by joiners.
- A repo-root AGENTS.md/CLAUDE.md pointer (per machine, not synced): the
awareness/routing layer. Required because discovery differs by
platform — Claude Code and Hermes load nested instruction files lazily
(only after entering the folder), and Codex never discovers them at
all (root→cwd path only).
SKILL.md gains a "Teaching agents the folder" section with the platform
discovery table and a first-contact orientation ritual; /beardrive:install
step 4 and /beardrive:init step 6 now offer both files as separate
consents; install.md step 3 gains the same never-sync-a-repo-root hard
rule init.md already had; README/CLAUDE.md descriptions updated.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
Pre-existing working-tree edits landed with this batch: SKILL.md and
/beardrive:init gain guidance for gbrain-backed shared folders — elect a
single enrichment owner (everyone else indexes read-only), route member
captures into per-member subtrees, treat conflict copies as the canary
for a second enricher, and follow the owner's pull→import→enrich→write
cycle discipline to avoid silently reverting human edits.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
Clicking a project now lands on an index page (/<project-id>) instead of
redirecting to Insights: a tabbed guide for mounting the project locally
and wiring it to a coding agent, with the Insights dashboard embedded
below for admins/org-owners.
- Claude Code & Cowork share one tab guiding through the plugin
(/plugin marketplace add + /beardrive:install with this hub's URL and
the project id pre-filled); Hermes and Codex get the four-step CLI flow
(install, login, init, hooks install) with real copyable commands.
- /<project-id>/insights stays a dedicated deep-linkable route; the
sidebar project name is now a home link.
- Copy buttons, localStorage-persisted tab choice, scroll/back-forward
behavior consistent with the other views.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
- No-flicker polling: tree and heat refreshes compare fetched data and skip
DOM rebuilds when nothing changed; in-place listing re-renders preserve
scroll.
- RESTful routes: /<project>/insights and /<project>/history[/<path>]
replace query-param views; the reserved first segments survive refresh,
deep links, and back/forward. Per-route scroll restore (manual
scrollRestoration) returns the reader to where they were.
- History rows unified: every entry (recent changes and full history) is a
clickable row that opens the file — view/download version links removed;
old versions remain reachable via the blob API until the time-travel
viewer lands. Folder tree loses the per-dir history icon.
- Hot path: danger rows use the amber accent, ⚠ trails the filename, and a
tooltip explains the flag (reads/30d, days unchanged).
- Frontend responses send Cache-Control: no-cache so upgrades never leave
browsers on a stale embedded UI.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
History notes still clamp to one line so the feed stays scannable, but
clicking (or Enter on) the note expands the full text and back — links
inside keep working. Selecting a project now lands admins/org owners on
the Insights dashboard instead of the empty "select a file" pane;
members keep the placeholder, and routes that carry a path still open
that file or folder.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
AgentHeat aggregates agent-kind buckets per device per top-level folder;
the handler joins the device registry (name/OS) and sorts by total. Human
and share buckets are never consulted, so human actor identities cannot
appear in the response — asserted by test, along with registry join, root
folder bucketing, sort order, and invalid-by rejection.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
Insights (admin/org-owner) becomes a stacked dashboard driven by one heat
fetch plus the tree, with the all/human/agent lens applied throughout: a
dependency-free squarified treemap of every file (cell size = reads in the
window, color = staleness, ⚠ on hot+stale, one delegated click handler —
readable at 500+ files, group labels open folders, cells open files); the
reads×freshness scatter demoted to drill-down with density handling
(translucent dots, radius = agent share); and a hot-path top-20 list with
stacked agent/human bars replacing the plain danger list. The agent
coverage matrix section renders when the server provides the by=device
breakdown. Design addendum recorded in docs/design/read-heatmap.md;
calendar/streamgraph explicitly deferred.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
Admin/org-owner Insights view (⋯ menu): dependency-free SVG scatter of
every file by 30-day reads × days since last change, log scales, with the
hot-but-stale danger quadrant shaded and a ranked fix-these-first list;
lens toggle for all/human/agent reads. The Claude Code plugin gains a
PostToolUse(Read) hook so plugin users feed agent-read telemetry without
project-level hook registration. Docs synced: README, SKILL.md,
plugin install/init commands, CLAUDE.md, design doc marked implemented.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
Phase 3 of the read heatmap: `bdrive read-log` parses any platform's hook
event JSON from stdin and queues in-project file reads (mount-relative,
ignore/include-filtered) in a per-volume spool — no network on the hook
path. The sync cycle drains the spool best-effort to the hub's new
POST /api/p/<id>/reads via the remote ReadReporter capability (https
backend only); a failed report retries next cycle and never fails or
offlines the cycle. `bdrive hooks install` now registers a third hook per
platform on its read-tool matcher (claude Read, codex read_file
best-effort, gemini read_file|read_many_files, hermes read_file), each
idempotent on its own marker so sync-only configs upgrade in place. Agent
reads land as agent traffic in the heat view, actor = device id.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
Read telemetry per docs/design/read-heatmap.md: a ReadLedger over a new
batch-oriented MetaStore ReadRepo (file reads.json + SQL read_stats)
aggregates viewer and share reads into daily per-actor buckets, debounced
to visits, folded into all-time rows past retention. GET /api/p/<id>/heat
serves per-path counts (human/agent/share, distinct readers, last read) —
never identities. /store sync traffic and history blob views are not reads.
The viewer shows heat dots and read counts on folder listings and the file
meta line.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
`bdrive hooks install` detects the agent platforms in use — Claude Code
(.claude/), Codex (.codex/), Gemini CLI (.gemini/), Hermes (~/.hermes/) —
and idempotently merges beardrive's turn-boundary sync hooks into each
platform's own hook config (JSON for claude/codex/gemini, YAML for
hermes), preserving existing hooks. All four pipe hook JSON with a
session_id, so one POSIX-sh hook command serves every platform: pull at
turn start, push after edits, changes stamped "<agent> session <id>".
Bare `bdrive hooks` prints the detection/registration table.
The beardrive skill now runs it automatically after `bdrive init`, and
/beardrive:install's hand-maintained settings.json block is replaced by
the command, so the hook content has one source of truth in the binary.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
`bdrive sync --note <text>` stamps session context onto every op the cycle
commits and persists it in the volume store (note.json, --note-ttl default
30m) so the daemon's own scans stamp it too — winning the race between
one-shot hook syncs and the 3s daemon scan. The plugin sync hook extracts
session_id from hook stdin JSON and passes it automatically, so history
links every change to the Claude Code session that made it. Conflict-copy
ops keep their own note; expired/cleared notes stop applying.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
Folders in the file tree are now selectable: the row opens the folder
(chevron still folds) and the main pane lists its contents, with clickable
breadcrumbs, folder URLs/deep links, palette entries, and uploads landing
inside the selected folder. In hub mode the listing includes a Recent
changes feed for the subtree; the history API now classifies puts as
add|edit (replayed over all ops before filtering) so entries are badged
added/edited/deleted, and op notes render under entries with URLs
linkified.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
A subdirectory with its own .bdrive/config.json now syncs only through
its own project: the parent mount never scans into it, materializes
over it, or emits deletes for paths under it (cached paths are dropped
without delete ops, the same posture as newly ignored files). Detection
is presence-based (config.IsMount) so even a corrupt nested config
still fences. This makes the team-knowledge-folder-inside-a-personal-
mount topology safe.
The plugin's init command and skill gain a "connecting knowledge
tooling" playbook: a detection ladder (gbrain -> OKF wiki -> plain
docs folder -> starting-point menu) and two hard rules — one transport
per folder (git-tracked dirs are handed off via git rm --cached +
gitignore before syncing) and knowledge-as-scoped-subfolder (never a
repo root). Conflict copies documented with their exact glob
(*.bdrive-conflict-*), since openknowledge validate cannot see them.
New flow tests: shared-subfolder scope both ways, teammate connect
over identical/divergent local content, nested mount syncing
independently on two projects, and a guard pinning the conflict-copy
filename to the documented glob.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01P5cxPQdSGJnjXCYY9GeWXt
The initial import was latency-bound: the push loop uploaded blobs strictly
serially, doing an Exists round-trip then a Put per blob (~2-3N sequential
round-trips against the hub).
- Upload unique blobs in parallel (errgroup, 16 in flight) — the main win.
- Drop the separate Exists round-trip: the backend's Put is idempotent and
the hub already skips content it has (reported during signing), so the
check was redundant.
- Session.OnProgress emits upload progress (done/total files + bytes) from
the push phase; the CLI renders an in-place bar on a TTY and periodic
percentage lines otherwise, wired into `bdrive init`'s initial cycle and
`bdrive sync`. The daemon stays silent.
Multi-device convergence tests pass with -race; new TestPushProgress covers
the progress emission.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R7Q9ZKSZRTdvrSJkYLUmYs
BearDrive is now hub-based only. Clients sync through a `bdrive web` hub
over https and never talk to object storage directly.
- Remove the `bdrive remote` / `bdrive remote set` command entirely. Wiring a
folder's remote is `bdrive init` (always a hub); switching hubs is
`bdrive login <url>` then re-`init`. Object storage stays as the *hub's*
backing store (`bdrive web s3://…`), never a client remote.
- Reposition README, CLAUDE.md, and the beardrive skill from
"no server required / direct-to-bucket" to hub-based; section 2 of the
skill is now "storing a hub's data" (server-side), and the stale
`init --remote` / `remote set` examples are fixed.
BREAKING: direct client-to-bucket sync is no longer supported.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R7Q9ZKSZRTdvrSJkYLUmYs
- New `bdrive logout` clears the saved token + account (folders untouched);
`--forget` also drops the remembered server. Switching hubs is
`bdrive login <new-url>` then re-`init`.
- `remote set --help` now explains the two remote kinds — object storage
(s3/gs/file) vs a bdrive hub (https://<server>) — with examples, instead of
leading with only s3://. (The https:// hub scheme was already accepted.)
- Docs updated (README, CLAUDE.md, SKILL.md).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R7Q9ZKSZRTdvrSJkYLUmYs
Add a CGO-free static Dockerfile (distroless) and a one-shot Cloud Run
deploy recipe (deploy/gcp-cloudrun.sh + README): single-instance hub on
Cloud Run, metadata in Cloud SQL Postgres (the new `database` backend),
blobs/journals in a GCS bucket, config + DB password in Secret Manager,
least-privilege runtime service account.
Pinned to max-instances=1 (the current build assumes one writer). Verified
end to end against a live deployment.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R7Q9ZKSZRTdvrSJkYLUmYs
Wire a `database` block in bdrive web config: {driver: file|sqlite|postgres,
dsn}. Default file, unchanged (existing per-registry paths honored). sqlite
or postgres routes all five hub registries through the SQL MetaStore.
Verified: signup + project + invite round-trip survives a restart on BOTH
file and sqlite; the sqlite DB holds real relational tables/rows (not JSON
blobs). Docs: a "choosing a database" section (README), the persistence
architecture note (CLAUDE.md), and a config mention (SKILL.md).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R7Q9ZKSZRTdvrSJkYLUmYs
Add a database/sql MetaStore (db_sql.go) targeting SQLite locally and
Postgres/Supabase in production through one portable schema: real tables
(accounts, tokens, auth_policy, projects, orgs, org_members, invites,
shares, devices), idempotent CREATE TABLE IF NOT EXISTS migrations at Open,
?→$N placeholder rebinding for Postgres, times as RFC3339 text, and
transactional multi-row writes (an org and its members). Pure-Go drivers
only (modernc.org/sqlite, jackc/pgx v5) so CGO-free builds keep working.
One shared conformance suite (db_conformance_test.go) runs the same
service-level operations — accounts+tokens, policy, pending/approve,
projects create-or-join/rename/delete, org roles, invite create/redeem/
uses/validity, share create/revoke/expiry, devices — then reopens the store
and asserts everything persisted. Runs against file AND sqlite always, and
postgres when BDRIVE_TEST_POSTGRES is set. Verified green on all three
against a real local Postgres.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R7Q9ZKSZRTdvrSJkYLUmYs
Introduce a pluggable metadata persistence layer: a MetaStore of typed
repositories (AccountRepo, ProjectRepo, OrgRepo, ShareRepo, DeviceRepo).
The five registries keep their in-memory maps + logic and now persist each
change as one record through a repo, instead of rewriting a whole JSON file
inline. The `file` backend (db_file.go) reproduces the exact on-disk JSON
formats, so a running hub loads unchanged. Open*(path) constructors stay as
thin wrappers over the file backend.
Zero behavior change: full webapp suite green; the example hub boots and
existing auth.json/orgs.json/projects.json load and sign-in works.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R7Q9ZKSZRTdvrSJkYLUmYs
