First step toward the Excel-like editable-table the user asked for.
Architecture decisions in this phase came from a focused research
pass over Notion / Airtable / AG Grid / Handsontable / Glide / W3C
ARIA APG; the design notes are in this commit's predecessor as a
research synthesis. Five phases planned; this is phase 1 of 5 and
ships the cell-selection + keyboard-navigation + per-cell editor
mount-on-demand foundation. Edits in this phase live in a client-
side draft buffer only; row-level save + ETag conflict UX is
phase 3.
Scope:
- ARIA grid pattern verbatim (W3C WAI-ARIA APG): role=grid on the
table, role=row on rows, role=gridcell on cells, roving
tabindex (only one cell carries tabindex=0; arrows move it).
This makes the grid one tab stop in the page tab order — the
documented spreadsheet UX, and also the basis for screen-reader
correctness.
- Click selects a cell. Arrow keys move selection. Tab and
Shift-Tab move with row-wrap. Home / End jump within row;
Ctrl/Cmd+Home / End jump to grid corners. Enter, F2, double-
click, or any printable character all enter edit mode. In edit
mode: Enter commits and moves down (Excel convention), Tab
commits and moves right (with row-wrap), Escape cancels and
restores the prior value, blur commits.
- Mount-on-demand cell editor: one <input> at a time is
instantiated inside the selected cell. Survives 1000-row tables
without the focus-ring churn an always-editable design would
hit, and lets Phase 2 swap the input for schema-driven widgets
(number / date / select / etc.) without restructuring.
- Draft buffer at app.state.drafts keyed by row id (the row's
re-edit URL — stable across sort and filter). When a cell
commits with a value different from row.data, the draft entry
is set; render reads from the draft via effectiveCellValue() so
the visible cell content reflects unsaved edits. No-op edits
(commit returns the original value) clear any pending draft.
- Selection survives re-paints. Sort / filter / spec changes
trigger a re-render; the editor's setSelected at end of paint()
clamps to new bounds and rebinds tabindex. The user's cell
doesn't disappear when they sort the column they're editing.
- Numeric coercion fast-path: cells whose column declares
format=number/integer coerce the input string to Number on
commit. Phase 2 will generalize this to schema-driven coercion
for date, boolean, enum, etc.
UX consequence — single-click semantics change:
The pre-existing row-click-navigates-to-form-edit behavior is
gone. Single click now selects a cell (spreadsheet-native). The
"open this row in the form editor" affordance moves to phase 2
(an explicit "Edit…" button or an icon column). The row-click-
navigation tests in tests/tables.spec.js are replaced with seven
new tests covering the editor lifecycle.
What this phase does NOT do (and which phases own it):
- Phase 2: schema-driven editor widgets (right input type per
column). Server-side validation 422 → red-corner marks. Complex
types (object, generic array, oneOf) get an "Edit…" button that
opens the side-panel form-render mode the unified bundle
already ships.
- Phase 3: row-level save on row-blur via PUT + If-Match. Stale-
row badge with "Use mine" / "Reload" on 412. Outbox carries the
offline path transparently via the existing source.js layer.
- Phase 4: copy/paste from Excel/Sheets via TSV parser, spill-
from-anchor or fill-all into a selection range.
- Phase 5: undo (linear command stack, Ctrl+Z, session-local) and
multi-cell ops (range select, bulk delete, Ctrl+D / Ctrl+R fill).
Tests (tests/tables.spec.js, all 15 pass):
- clicking a cell selects it (replaces the old row-click-navigates
test; verifies single-click does NOT navigate)
- arrow keys move cell selection
- Tab and Shift-Tab traverse cells with row-wrap
- Enter enters edit mode; Enter commits and moves down (verifies
draft is applied to visible cell + selection moves)
- Escape cancels edit, restoring prior value (verifies no-op on
draft buffer)
- typing a printable char enters edit and replaces the value
- double-click also enters edit mode
- non-editable rows still get the readonly class (cosmetic guard
for an existing convention; phase 3 will gate write submission)
Files:
- tables/js/editor.js (new) — selection + keyboard handling +
edit-mode lifecycle + draft buffer.
- tables/js/app.js — state.selected / state.editing / state.drafts
fields.
- tables/js/render.js — ARIA roles + editor.attachToCell wiring;
cells render via editor.effectiveCellValue so drafts show.
- tables/js/main.js — paint()-end editor.attachToTable +
setSelected restore.
- tables/css/table.css — selected-cell focus ring (outline,
doesn't shift surrounding cells); cell-input bare-inside-cell
styling.
- tables/build.sh — editor.js in the concat list.
- zddc/internal/handler/tables.html — regenerated bundle.
Bundle size: 117 KB → 124 KB (+7 KB for editor.js + ARIA + draft
machinery). Well within the budget the library survey identified
(Tabulator would have been +100 KB; SlickGrid +34 KB; custom is
+7 KB and we keep the no-third-party-deps invariant).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two intertwined refactors that share too many files to split cleanly.
Both are described separately below.
PART 1 — in-dir convention for table+form spec files
Old layout had the spec at the parent and rows in a child:
archive/<party>/
mdl.table.yaml spec
mdl.form.yaml row-edit form
mdl/ rows-dir
row-001.yaml ...
URLs were /<dir>/mdl.table.html and /<dir>/mdl.form.html. Copying
mdl/ elsewhere lost the spec and form because they lived next door.
New layout collapses everything into the rows-dir:
archive/<party>/mdl/ self-contained
table.yaml spec
form.yaml row-edit form
row-001.yaml ... rows
URLs become /<dir>/mdl/table.html and /<dir>/mdl/form.html. The
"copying-the-folder-takes-everything" property the user asked for
falls out by construction; the row-edit URL /<dir>/<id>.yaml.html
keeps the same shape (spec is now in the same dir, not the
grandparent).
Server changes:
- internal/handler/tablehandler.go RecognizeTableRequest fires on
/<dir>/table.html when <dir>/table.yaml exists. The .zddc.tables
alias map is gone — pure presence-based discovery now matches
the form system's existing convention. Default-MDL fallback at
archive/<party>/mdl/ stays for the virgin-archive case (the
rows-dir need not exist on disk; the URL renders fully virtually).
- internal/handler/formhandler.go RecognizeFormRequest fires on
/<dir>/form.html and /<dir>/<id>.yaml.html with spec at
<dir>/form.yaml. specEligible accepts on-disk files OR the
default-MDL virtual path so an empty mdl/ dir still surfaces the
add-row form.
- internal/handler/tablehandler.go IsDefaultMdlSpec moves to
serving archive/<party>/mdl/{table,form}.yaml (5 segments after
ZDDC_ROOT). New isAtArchivePartyMdlLevel predicate; new
isAtArchivePartyMdlDir for directory-based recognition. New
IsDefaultMdlSpecAbs accessor for callers that hold an abs path
rather than a URL (formhandler).
- internal/handler/formhandler.go loadFormSpec(fsRoot, path) falls
back to embedded default-MDL bytes when os.ReadFile returns
NotExist AND the path matches the archive-party-mdl shape. Three
call sites updated to pass cfg.Root.
- internal/handler/formhandler.go serveFormCreate writes
submissions to filepath.Dir(req.SpecPath) — the spec, the form,
and rows all live in one directory. The submissionsDir creation
is idempotent (MkdirAll); cascade falls back one level for ACL
evaluation when the dir hasn't been materialized yet.
- internal/handler/tablehandler.go tableRowsRedirect now points at
/<dir>/table.html (was /<dir>.table.html) when the directory
request maps to a recognized table.
- cmd/zddc-server/main.go dispatch synth flips from
urlPath + ".table.html" to urlPath + "/table.html" for the
no-trailing-slash → tables-app routing.
- internal/apps/availability.go DefaultAppAt comment clarified
that the dir at archive/<party>/mdl/ IS the table (not a child).
Client changes:
- tables/js/context.js walkServer fetches <currentdir>/table.yaml
directly — no .zddc walk for table declarations. Rows are every
*.yaml in current dir EXCLUDING table.yaml and form.yaml. The
.zddc fetch-for-aliases is gated on file:// (online mode 404s
on .zddc reads via the dispatcher's reserve guard, so skipping
the request avoids browser console noise).
- tables/js/main.js add-row button links to relative form.html
(same dir).
- tables/js/render.js + filters.js: every column's autofilter is
uniformly a text-contains input, even enum columns — keeps the
filter row visually consistent and doesn't constrain users to
the enum vocabulary.
PART 2 — unified table+form HTML bundle
The form-render and table-render code paths share field schemas,
the cell editor for excel-mode IS a form widget, and the form
system's POST-back / validation already exists. Combining the two
HTMLs eliminates duplicating jsyaml/jsonschema/theme/source-
detection/.zddc-parsing across two single-file tools.
- tables/template.html grows two top-level mode containers:
#table-mode (toolbar + sortable table) and #form-mode (form +
submit button). Both hidden at parse time; the dispatcher
unhides one. The shared #form-context placeholder was added
here so the server's existing injectFormContext target
resolves.
- tables/js/mode.js (new) sets window.zddcMode synchronously
based on URL pattern: /form.html or /<id>.yaml.html → form,
/table.html → table, else inline-context fallback for
file:// (whichever context blob is non-empty wins). Unhides
the matching container at DOMContentLoaded.
- tables/js/main.js init() and form/js/main.js boot() each guard
early when mode isn't theirs. Both apps live on different
globals (window.tablesApp vs window.formApp) so module
registration doesn't collide.
- form/js/main.js title write falls back from #form-title to
#table-title (the unified bundle's shared header element)
when the dedicated id isn't present.
- tables/build.sh concatenates form modules (widgets, render,
object, array, errors, post, serialize, util) and form CSS.
No new external deps. Bundle grows from ~95KB to ~120KB.
- internal/handler/formhandler.go drops the //go:embed form.html
directive; serveFormRender now writes embeddedTablesHTML via
a small formRenderHTML() accessor (var declared in
tablehandler.go, same package). The embedded form.html file
is removed.
- build script: cp form/dist/form.html → internal/handler/form.html
step is gone (file no longer exists in the source tree). cp
tables/dist/tables.html → internal/handler/tables.html now
runs unconditionally rather than only on beta/stable cuts —
the renderer is a fixed binary component and dev iteration
needs the embedded copy refreshed every build. Channel-cascaded
apps (internal/apps/embedded/) stay channel-gated as before.
- form/dist/form.html still builds for standalone offline-only
use (downloadable from /releases/), but no longer goes into
the binary.
Tests:
- internal/handler/tablehandler_test.go and formhandler_test.go
rewritten for the in-dir layout. New test
TestRecognizeFormRequest_DefaultMdlAtArchiveParty covers
empty-form, create POST, re-edit row, and the negative cases
(Working/, non-mdl name) where the fallback must NOT fire.
- internal/handler/directory_test.go updated for the new
/<dir>/table.html redirect target.
- cmd/zddc-server/main_test.go TestDispatchSlashRouting Location
expectation updated.
- tests/form-safety.spec.js loads tables/dist/tables.html
(named form.html in the temp dir to trigger form-mode in the
dispatcher) so it tests the same bytes the server returns.
Title-element selector switches to #table-title.
- tests/tables.spec.js updates the status-filter test for the
uniform text-input filter.
Docs:
- AGENTS.md form-data system rewrites the URL conventions and
storage layout for in-dir; gains a Tables system section
parallel to forms describing the self-contained-directory
property; subfolder rules ("one table per folder by
construction; subfolders allowed and silently ignored as rows
— legitimate uses: nested sub-tables, per-row attachments,
drafts, future history sidecars") so we don't re-derive this.
Not included (deferred):
- ACL gating on cell-level writes — not relevant until Phase 3.
- Editable cells UI — separate commit (Phase 1).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Sub-threshold finding from a focused security review of the CI URL
work — defense-in-depth even though it sits inside the documented
"trust upstream" boundary.
The mirror walker's purgeOrphans deletes local files that aren't in
the upstream's listing. It walked a dirPath built recursively from
upstream-supplied entry names and called os.Remove on the resolved
local path with no containment check. A hostile or compromised
upstream returning ".." in a directory listing could steer the
walker out of cache.root and into the parent — deleting whatever
matches the upstream's "expected to be there" filter in the wrong
directory.
A healthy master never produces such entries (listing.FromDirEntries
filters dot-prefix names), so the bug only fires under an actively
malicious or MITM'd upstream — confidence stayed below the report
threshold. But the fix is small and the cost of being wrong is real
deletion of files outside the cache, so it's worth doing.
Two layers:
1. walker.go walkDir filters upstream listing entries with name ==
"" / "." / ".." or containing "/" / "\" before recursing. Logs
a WARN with the dropped name so an operator can see if their
upstream is misbehaving.
2. purgeOrphans verifies the resolved localDir is contained under
s.cache.root (HasPrefix(root + sep) || == root) before
ReadDir+Remove. Logs a WARN and bails on mismatch.
Either layer alone would fix the original vector; both together
match the defense-in-depth pattern cachePathFor already follows for
single-file writes (line 506).
New TestWalker_HostileUpstreamCannotEscapeCacheRoot constructs a
fake upstream that returns a "../" entry in its listing, places a
sentinel file in the parent of cache.root, runs a mirror walk, and
asserts the sentinel survives. Both filter and containment guard
fire; the sentinel stays put.
Existing mirror tests unchanged — the filter only drops names that
shouldn't appear in healthy listings.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
URLs are now case-insensitive against the on-disk casing under
ZDDC_ROOT, with a lowercase-wins tiebreak when sibling case variants
exist. File and folder names preserve case on disk — the change is a
pure URL→FS-name mapping; nothing renames anything.
internal/fs/resolve.go ResolveCanonical walks segments left-to-right
under fsRoot. Per segment: try lowercase first (canonical / cheap
lstat fast-path), then exact-case, then readdir+CI scan with the
all-lowercase variant winning the tiebreak. Walk stops at the first
segment that doesn't exist on disk so virtual prefixes (.archive,
.profile, .tokens, .auth) and 404 paths flow through with their tail
preserved verbatim. Path-escape safety check on the resolved abs
path matches the existing safeJoin pattern.
Wired in at the top of cmd/zddc-server/main.go dispatch(), which
rewrites r.URL.Path before any handler runs. Downstream handlers
(plus their existing safeJoin calls and the cascade walker) pick up
canonical case automatically — no per-handler changes. The ACL
cascade benefits from this for free since EffectivePolicy is keyed
by the now-canonical absolute path.
internal/handler/middleware.go AccessLogMiddleware snapshots the
as-typed URL path before the rewrite. The audit log's `path` field
records what the client actually sent; a `resolved_path` field is
added only when canonicalization changed it. Operators reading the
log can see both the raw request and what was served.
Lowercase as the project-wide canonical convention is already
honoured by the auto-created folders in internal/zddc/ensure.go
(working/, staging/, archive/<party>/incoming/) and the server's
own state dirs (_app/, .zddc.d/tokens/, .zddc.d/outbox/,
.zddc.d/logs/). Operators who drop a Mixed-Case-Folder/ on disk
keep that casing — the resolver finds it via the readdir tier.
Performance: the lowercase-first lstat is one syscall on the hot
path. Only mismatches (mixed-case URL where on-disk is also
mixed-case) pay the readdir+EqualFold scan, and Linux page-caches
small-dir readdirs aggressively. Apache mod_speling uses the same
"try then fallback" pattern.
Tests:
- internal/fs/resolve_test.go — 9 unit tests: exact-case,
mixed-case-URL-with-lowercase-on-disk, mixed-case-URL-with-
mixed-case-on-disk, both-cases-exist-lowercase-wins, nonexistent
segment preserves remainder, file-segment terminates walk, escape
rejection, trailing-slash normalization, root.
- cmd/zddc-server/main_test.go TestDispatchCaseInsensitiveURL —
end-to-end through the dispatcher with sibling Archive/ and
archive/ on disk; all four URL casings of the same path serve the
lowercase variant's content (proves the tiebreak fires through
every layer).
- Full Go suite green.
Docs: AGENTS.md gains a "URL handling" subsection in the
zddc-server section; ARCHITECTURE.md security-model table gains a
"URL canonicalization" row.
Out of scope (separate decisions, can revisit if needed):
- ACL glob CI-matching. If .zddc rules use mixed-case URL globs,
they won't match the canonical lowercase URL. Workable today by
writing rules in lowercase. Touches a different package.
- Redirect-to-canonical (303). Server serves under whichever case
the client used; canonicalization is internal. Could 301 to
canonical for SEO/bookmark hygiene as a follow-up.
- Client-mode (proxy/cache). Only master mode is wired so far.
Cache-handler CI lives in internal/cache/cache.go cachePathFor
and is a separate code path.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
A focused security review of phases 1-4 surfaced one MEDIUM finding
(confidence 9/10): in client mode (--upstream set) the cache layer
forwards the configured bearer to upstream on every incoming request
without authenticating the local caller, AND --addr defaulted to
:8443 (all interfaces). Together those mean a CLI user running
`zddc-server --upstream https://master --bearer-file ~/token` on a
laptop on hotel/cafe Wi-Fi exposes an open-proxy confused-deputy:
any attacker on the same L2 connects to https://<laptop-ip>:8443,
accepts the self-signed cert, issues GETs (or PUTs/DELETEs that
queue in the outbox), and the cache laundries each request through
upstream with the engineer's bearer. The full cached subtree leaks.
Two layers of defense in config.Load:
1. Loopback default in client mode. When cfg.Upstream is set and
neither --addr nor ZDDC_ADDR was passed explicitly, --addr
downgrades to "127.0.0.1:8443" (vs ":8443" in master mode). CLI
users on a laptop get safe-by-default. Operators who want a
non-loopback bind opt in explicitly.
2. Refuse non-loopback bind + bearer-file without acknowledgement.
When cfg.Upstream is set, BearerFile is non-empty, the chosen
addr is non-loopback, AND --insecure-direct is not set, the load
fails with an error that names the bind, the threat (open-proxy
confused-deputy laundering bearer credentials), and the
acknowledgement flag. The helm zddc-server-cache/ chart already
sets ZDDC_INSECURE_DIRECT=1 and relies on Kubernetes-namespaced
pod networking for the gating, so the chart path is unaffected.
The guard is bearer-file-conditional because proxy mode without a
bearer doesn't have a credential to launder, and refusing it
would needlessly block proxy-without-auth deployments.
Tests in internal/config/config_test.go lock down all four cases:
- --upstream with no explicit --addr → 127.0.0.1:8443
- --upstream + non-loopback --addr + --bearer-file (no IDirect) → refuse
- --upstream + non-loopback --addr + --bearer-file + --insecure-direct → ok
- --upstream + non-loopback --addr + NO bearer → ok (no credential to leak)
Doc updates: zddc/README.md client-mode "Flags" section gets a
WARNING block describing the loopback default + insecure-direct
escape hatch. AGENTS.md ZDDC_UPSTREAM row mentions the addr
downgrade. ARCHITECTURE.md gains a "Confused-deputy guard at
startup" subsection under "Master + proxy/cache/mirror" with the
two-layer defense rationale. helm/zddc-server-cache/values.yaml.example
adds an inline note next to addr: ":8080" explaining why the chart
sets ZDDC_INSECURE_DIRECT=1 and what the consequence is of removing
either side of the gating.
Master mode is unaffected — the client-mode validation block is
gated by `if cfg.Upstream != ""`. All existing tests pass.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 3 + 4 live two-instance smoke tests against the synthetic
~/zddc-test-data fixture surfaced three real bugs that the unit
tests missed. All three are fixed in this commit.
1. walker: filenames with spaces/parens land on disk percent-encoded
walkSubtree was passing the URL-encoded child URL (built via
url.PathEscape) to fetchFileIfNeeded → cachePathFor, so a file
named "Foo (IFI) - Bar.md" landed at <root>/.../Foo%20%28IFI%29
%20-%20Bar.md on disk. Then purgeOrphans iterated os.ReadDir
(which sees the encoded names) and compared against upstreamNames
(decoded names from the listing JSON). Every fetched file was
classified as an orphan and immediately deleted: a 180-file walk
produced "fetched=180 purged=111" with only 70 files remaining.
Fix: walker now maintains two parallel path strings — dirURL
(URL-encoded for HTTP requests) and dirPath (decoded for disk
keys). fetchFileIfNeeded, fetchListing, persistOnly, and
purgeOrphans all take the decoded path. listingCachePathFor
gets dirPath too. Smoke confirmed: dirs=29 files=180 fetched=179
purged=0 (one file already cached from the user's GET that
triggered the walk).
2. outbox: replay loop sleeps 5min after eager startup pass
RunReplayLoop's idle-poll interval is 5min. After the eager
startup pass with 0 entries, the loop sleeps 5min — even if a
PUT-while-offline arrives 1 second later, replay won't fire for
~5 min. The cache returned 202 promptly but the queued write sat
on disk until either a 5min nap elapsed or another PUT happened.
Fix: Outbox gains a wake chan (buffered=1, drop-on-full).
Enqueue posts to it after writing meta.json. RunReplayLoop selects
on wake alongside the timer, so a new offline write triggers an
immediate replay attempt. Smoke confirmed: PUT queued at T+0,
master back at T+3, replay completes at T+3 (was previously a
30s wait through the timer-based poll).
3. master: PUT/DELETE didn't honor If-Unmodified-Since
The cache's outbox sends If-Unmodified-Since: <cached-mtime> on
replay so the master can reject conflicting writes with 412. The
master's checkIfMatch only evaluated If-Match (ETag-based), so
the cache's mtime-based precondition was silently ignored. Result:
an offline PUT staged before an external mod would clobber the
newer external content on replay — silent data loss in the exact
scenario the outbox is designed to detect.
Fix: checkIfMatch now also evaluates If-Unmodified-Since per
RFC 7232 §3.4, returning 412 when the file's current mtime is
strictly later than the header value (1-second resolution to
match HTTP-Date precision). Smoke confirmed: cache GET → external
mod via direct file write → cache offline PUT → master back →
replay sends IUS → master 412 → outbox entry renamed to
<id>.conflict-<RFC3339>/ → master content preserved (the
external mod, not the stale offline write).
Also added an info-level "outbox: replay attempt" log to tryReplay
so an operator watching the cache logs sees the replay loop is
alive even when every entry defers (transport error). Previously
the loop was silent unless a replay actually completed (200) or
conflicted (412).
go vet + go test ./... + go test -race ./internal/{cache,auth,handler}/...
all green. Synthetic ~/zddc-test-data fixture (553 files, 144 PDFs)
exercises the walker against realistic ZDDC filenames including
spaces, parens, and accented characters that the unit tests'
"a.txt" / "b.txt" inputs never hit.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
PUT / POST / DELETE in client mode now work end-to-end. Online: the
cache layer forwards to upstream and (on success) drops any cached
entry for the path so the next read fetches fresh. PUT/DELETE include
If-Unmodified-Since derived from the cached file's mtime so the master
can reject conflicting writes with 412 Precondition Failed.
When upstream is unreachable, the request is captured in the outbox
at <root>/.zddc-outbox/<id>/ — directory per queued write, mode 0700,
containing meta.json (method, RawURI, Content-Type, base mtime,
queued-at) and body.bin (request body, capped at 256 MiB). The client
gets 202 Accepted + X-ZDDC-Cache: queued and a JSON envelope.
A background replay loop started by runClient processes the queue:
- 2xx → delete entry; drop cached path so next read fetches fresh
- 412 → rename to <id>.conflict-<RFC3339>/ for manual reconciliation
(body + meta intact for inspection or re-submit)
- 4xx other → drop (retry won't help; logged at WARN)
- 5xx / transport error → leave for next pass
Replay schedule: eager at startup, then 30s while pending falling
back to 5min while idle. Loop honors graceful-shutdown context.
Disabled in --mode=proxy (proxy persists nothing by design — offline
writes return 503 instead of queueing).
Outbox IDs are <unix-nano-base16>-<hex-random> so lex-sort = queue
order; concurrent enqueues never collide. Conflict-rename appends a
4-char random suffix on the unlikely same-second collision.
The local cache is intentionally not updated for offline writes:
until upstream confirms the user reads still see the upstream-cached
version (or 503 if uncached). Trade-off: no "did my queued write
actually win?" ambiguity, at the cost of not seeing one's own
offline edits immediately. Phase 5 will surface .conflict-<ts>/
directories in browse views.
Tests (20 new in outbox_test.go, 5 new in cache_test.go covering
the write path): NewOutbox creates 0700 dir, Enqueue persists meta
+ body, Pending returns lex-sorted entries excluding conflicts,
Replay deletes on 2xx / renames on 412 / leaves on transport error
/ leaves on 5xx / drops on 4xx-other, IUS sent only for PUT/DELETE
with base mtime, query string preserved, ServeHTTP online write
forwards + evicts cache, ServeHTTP offline write queues with 202,
ServeHTTP offline + no outbox returns 503, ServeHTTP PUT sends IUS
from cached mtime, oversize body rejected, IDs lex-sortable,
RunReplayLoop stops on context cancel, concurrent Enqueue 30×
no collisions. Full suite + go vet clean.
Doc updates: zddc/README.md gains a "Writes (online + offline
outbox)" subsection covering both paths and replay outcomes;
"What client mode is NOT, yet" now lists only conflict UI and
multi-tenancy. AGENTS.md client-mode pipeline gains writes +
mirror-mode bullets. ARCHITECTURE.md adds a "Writes: outbox +
offline replay" subsection with the trade-off rationale and the
phase-5-deferred conflict UI hand-off.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
--mode mirror layers an access-triggered walker on top of the cache
pipeline. When an incoming request's URL falls under one of the
configured --mirror-subtree paths, the scheduler kicks off a recursive
walk of that subtree iff (a) no walk for that subtree is in flight and
(b) now - last_walk_at >= --mirror-min-interval (default 1h). Walks
run in a goroutine; the user's request never blocks on scheduling.
Why access-triggered: a naive "walk on a fixed timer" would produce
thundering-herd polls on a master from many vendor mirrors most of
which are idle most of the time. Demand-triggering means idle mirrors
generate zero upstream traffic until someone hits them; active
mirrors stay current as a side effect of normal use.
The walk:
1. Recursively fetches JSON listings under the subtree, persisting
each at <dir>/.zddc-listing.json so directory browsing works
offline for walked subtrees.
2. For each file, fires a conditional If-Modified-Since GET (bounded
parallelism; default 4 concurrent) — 304 no-op, 200 overwrites,
403/404 purges the local cache.
3. After enumeration, per-directory orphan purge: local files absent
from upstream's filtered listing are removed (handles upstream
deletes + ACL revocations).
State persists at <root>/.zddc-mirror-state.json as
{subtrees: {<path>: {last_walk_at}}}. In-flight tracking is in-memory
only — a crash mid-walk lets the next access retry without manual
cleanup. Subtree path matching is longest-prefix-wins; "/" is a
catch-all (full mirror, the default when --mode=mirror is set without
explicit --mirror-subtree).
The cache layer also gained directory-listing caching (independent of
mirror mode but enabled by it). Directories are now stored at
<dir>/.zddc-listing.<html|json> sidecars, varied by Accept header.
Hit/miss/offline semantics mirror the file pipeline. Phase 2's
limitation that directories always proxied live (no offline browse)
is now resolved for any directory the user has visited or that mirror
mode has walked.
Mirror scope falls out of auth: the walker uses the local instance's
bearer, so it sees exactly what the user can see at upstream. Admin
bearer → full mirror; vendor bearer → vendor's permitted subtree;
no code distinguishes the cases.
New flags (also as ZDDC_* env vars), ignored when --mode != mirror:
- --mirror-subtree <csv> — repeatable subtrees (comma-separated);
empty + --mode=mirror = "/" (full mirror)
- --mirror-min-interval <duration> — default 1h
Tests (15 new in walker_test.go, 3 new in cache_test.go): subtree
normalization, longest-prefix matching, root-as-catch-all, walk
fetches all files in scope, out-of-scope URLs are no-op, rate-
limiting prevents double-walks within min-interval, walks re-fire
after interval elapses, orphan purge removes local-only files,
state file survives restart, concurrent triggers don't double-walk,
end-to-end ServeHTTP-kicks-mirror-on-access, listing format varies
by Accept, listing offline serves stale, persisted state atomic
write + corrupt-input handling. Full suite + go vet clean.
Doc updates: zddc/README.md flags table gains the two new entries
plus a "Mirror mode (access-triggered subtree walker)" subsection
with trigger semantics and properties; the "What client mode is NOT,
yet" list shrinks accordingly. AGENTS.md env-var table gains the
two new entries. ARCHITECTURE.md "Master + proxy/cache/mirror"
section now documents the walker scheduler / walk algorithm / state
file in a "Mirror walker (access-triggered)" subsection.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
zddc-server can now run as a downstream client of another zddc-server.
Set --upstream <url> and the master-side machinery (archive index, apps
server, watcher, OPA decider, ACL middleware, token store) is bypassed
entirely; cmd/zddc-server/main.go short-circuits to runClient(cfg)
which uses zddc/internal/cache/Cache as the entire request handler.
Three modes via --mode <proxy|cache|mirror>:
- proxy: forward upstream live, no disk persistence
- cache (default): persist responses on access; subsequent hits serve
from disk + background If-Modified-Since revalidate
- mirror: accepted but currently behaves like cache; the access-
triggered walker lands in phase 3
Cache directory layout is intentionally a normal ZDDC root: a file
fetched from <master>/foo/bar.txt is stored at <root>/foo/bar.txt with
no sidecar metadata. The local file's mtime is set to the upstream's
Last-Modified header so revalidation reflects the master's notion of
file age, not local fetch time. Running zddc-server --root <cache-dir>
without --upstream serves the cached files as a plain master — useful
for portable offline snapshots. A small .zddc-upstream marker is
written once on first persist for provenance.
Pipeline (GET/HEAD only — writes deferred):
- Hit → http.ServeContent serves directly (range-aware, 304-aware) +
background revalidate (304 no-op, 200 overwrite, 403/404 purge)
- Miss → forward to upstream with the configured bearer; tee response
body to client + tmp-file atomically renamed into the cache
- Network error + cached → serve stale + X-ZDDC-Cache: offline
- Network error + no cache → 503 + X-ZDDC-Cache: offline
- Directories always proxy live (no listing cache yet — phase 3)
- Cache-Control: no-store / private and non-200 responses bypass cache
Range requests work end-to-end (Range/If-Range headers forwarded on
miss; http.ServeContent handles them natively on hit). Hop-by-hop
headers per RFC 7230 §6.1 are dropped from forwarded responses.
New flags (also as ZDDC_* env vars), all ignored when --upstream is
empty (so master deployments are untouched):
- --upstream <url>
- --mode proxy|cache|mirror (default cache)
- --bearer-file <path> (0600 file with the master-issued token)
- --skip-tls-verify (separate from --no-auth; for self-signed dev)
Validation: --upstream must be http(s)://...; trailing / is trimmed.
Mode validated to one of the three known values. The startup
no-root-.zddc check is skipped in client mode (the cache directory
starts empty by design). The plain-HTTP-on-non-loopback check is also
skipped (the local instance never reads the email header to decide
anything; auth is forwarded to upstream as a Bearer).
Tests: zddc/internal/cache/cache_test.go runs httptest.NewServer as
the upstream and covers miss-then-hit, proxy-mode-no-persist,
directory-never-cached, HEAD-no-body, offline-with-cache,
offline-no-cache → 503, bearer forwarding, query-string preservation,
no-store bypass, path-traversal rejection, error-status forwarding,
revalidate-on-403/404/200/304, range-on-hit, concurrent-same-URL,
cache-path boundary cases. 23 new tests, full suite + go vet clean.
Live two-instance smoke verified: master at 127.0.0.1:18443, client
at :18444 with --mode cache, miss→hit→hit transitions work, file
materialises under cache root with parent dirs created, marker file
written once, range-on-hit returns 206, master sees background 304s
on every hit, killing master leaves cached files serving from disk
and never-cached files returning 503 + offline header.
Doc updates: zddc/README.md gains a "Client mode" section with the
modes table, flag reference, pipeline summary, two-instance recipe,
and explicit list of phase-2 limitations; AGENTS.md adds the four
new env vars to the reference table and a "Client mode" subsection
with smoke-test recipe and a pointer to the cache package;
ARCHITECTURE.md adds "Master + proxy/cache/mirror" before "Bearer
token issuance," covering the topology, the persist/warm switches,
the cache-IS-a-ZDDC-root invariant, the request pipeline, and the
v1-out-of-scope multi-tenancy note; CLAUDE.md's zddc/ entry
expanded to mention both deployment shapes so future agents pick it
up by default.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
zddc-server now issues its own bearer tokens for non-browser callers
(CLI tools, scripts, downstream proxy/cache/mirror instances). No
external IDP, no JWKS rotation. Self-service flow: sign in via the
browser, visit /.tokens, click "Create token," paste the resulting
plaintext into a 0600 file, and pass --bearer-file <path> to whatever
calls back into the server.
Storage is <ZDDC_ROOT>/.zddc.d/tokens/<sha256-hex>, YAML per token
with email/created/expires/description. Filename is the *hash* of the
plaintext, never the plaintext itself — a leak of the tokens
directory exposes hashes, not credentials. Mode 0600 / 0700, atomic
writes via temp+rename. Already shielded from public serving by the
existing dot-prefix guards in dispatch and fs.ListDirectory.
ACLMiddleware now recognises Authorization: Bearer <token>. On valid
token, sets the request email from the token file and falls through
to the existing ACL chain. On any failure (unknown / expired / store
unavailable / Bearer with no validator), returns 401 — no silent
fallback to anonymous, so a misconfigured client fails loudly.
JSON API at /.api/tokens (GET list, POST create, DELETE /<id> revoke)
backs a small inline HTML self-service page at /.tokens. Users can
only see and revoke their own tokens; cross-user revoke returns 404
to avoid leaking ownership.
--no-auth (ZDDC_NO_AUTH=1) skips ACL enforcement entirely on this
instance. On master: anyone reads everything (dev / trusted-LAN /
public-read deployments). On a downstream proxy/cache/mirror: trust
upstream's filtering, don't re-evaluate ACLs locally. Implemented as
a swap to policy.AllowAllDecider; all existing handlers keep calling
AllowFromChain unchanged. Distinct from --insecure, which only
relaxes the no-root-.zddc startup check. WARN-level startup log when
--no-auth is active so accidental enablement is visible.
33 new tests covering token storage, validation/expiry/revocation,
the JSON API end-to-end, the HTML page, and the middleware-Bearer
integration including the case-insensitive prefix and expired-token
paths. Full suite + go vet clean.
Doc updates: zddc/README.md "Authentication" rewritten to cover both
auth paths and the token UI/API; AGENTS.md gains ZDDC_NO_AUTH and a
"Bearer tokens" subsection flagging the dot-prefix-shielding pre-
condition; ARCHITECTURE.md adds "Bearer token issuance" and
"--no-auth" subsections under "Server security model" with the
hash-as-filename rationale and dispatch-shielding regression-
sensitivity called out; CLAUDE.md adds a one-line summary of the new
auth topology so future agents pick it up by default.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
When an HTML GET hits a directory that's the rows-dir of a registered
table — i.e. parent declares `tables: { <name>: ... }` with a valid
spec, OR the default-MDL fallback applies at archive/<party>/mdl/ —
ServeDirectory now 302s to <parent>/<name>.table.html so users land
on the table view instead of a bare browse listing of the row-yaml
files. JSON GETs on the same URL fall through unchanged so the table
client can still enumerate row files.
Detection reuses RecognizeTableRequest: synthesize the equivalent
.table.html URL from the directory request and let the existing
recognizer apply its operator-vs-default-vs-missing-spec rules. No
duplicated validation.
Updates main_test.go's TestDispatchSlashRouting to expect the new
behavior on archive/<party>/mdl/.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add a virtual-URL alias so the existing form-based .zddc editor is
reachable at the natural directory location (<dir>/.zddc.html) in
addition to the legacy /.profile/zddc/edit?path=<dir> entry. Both
flow through the same renderZddcEditor body — same template, same
gate, same form-posts-to-/.profile/zddc semantics.
Wiring:
- IsZddcEditorRequest(urlPath) reports whether the URL ends with
the .zddc.html leaf (case-fold not needed; .zddc is itself case-
sensitive on disk).
- ServeZddcEditorAtPath strips the leaf, resolves the parent dir,
asserts the dir exists, gates on hasAnyAdminScope, calls the
shared renderer.
- The dispatcher routes IsZddcEditorRequest URLs BEFORE the dot-
prefix segment guard (which would otherwise 404 the .zddc.html
leaf). The route is method-gated GET-only; mutations still go
through PUT/POST/DELETE on <dir>/.zddc via the file API.
Permission model unchanged from the /.profile entry: hasAnyAdminScope
gates visibility of the editor itself; CanEditZddc decides whether
the form is interactive or read-only at the requested directory.
Subtree admins can still inspect ancestor cascade ACLs (intended
since the cascade is what determines their authority).
Test (TestDispatchZddcEditorAtPath): root admin opens project /
working/ / deployment-root editors; non-admin and anonymous both
404; missing directory 404; trailing-segment-after-leaf 404.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
URL convention for directories under a project:
- <dir>/ (with trailing slash) → browse (the directory view; same
behaviour as today)
- <dir> (without trailing slash) → the canonical default tool for
that directory's context, served
inline (no 301 hop)
Tool mapping via the new apps.DefaultAppAt(root, dir):
- working/... → mdedit
- staging/... → transmittal
- archive/ → archive
- archive/<party>/ → archive
- archive/<party>/incoming|received|issued/... → archive
- archive/<party>/mdl/... → tables (the per-party MDL grid editor)
Directories outside the canonical layout (project root, scratch
folders) keep the legacy 301-to-trailing-slash redirect since no
default tool fits.
This generalises and replaces the bespoke
"GET archive/<party>/mdl/ → 302 mdl.table.html" redirect added in PR4.
The new dispatcher rule serves the table app inline at the bare-mdl
URL by routing through RecognizeTableRequest with the canonical
.table.html suffix appended; relative fetches resolve identically
because both URLs share the same parent directory.
Tests: TestDefaultAppAt covers all canonical positions plus
case-fold and out-of-tree edges. TestDispatchSlashRouting (replacing
the now-obsolete TestDispatchMdlRedirect) verifies the slash-vs-no-
slash distinction at every canonical folder + non-canonical
fallback.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Address two follow-ups from the security review of feat/zddc-inherit-directive:
1. file.go's Inherit docstring previously claimed "the internal decider
treats it as inherit:true and emits a warning at evaluation time" —
the decider does the first part but the warning was never wired up.
Strike the over-promise; point operators at the cascade tracer
(`/.profile/effective-policy`) which surfaces both `cascade_mode`
and `chain.visible_start` so a fenced configuration that's being
ignored under strict mode is visible.
2. AllowedAtLevel hardcodes ModeDelegated. Safe today (1-level
synthetic chain, no ancestors) but a footgun if anyone migrates
the shim to a real PolicyChain later. Add a `// Deprecated:`
marker pointing at GrantedVerbsAtLevel for fence-aware paths.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The cascade tracer's JSON response now carries:
- Top-level `cascade_mode` (string): the active mode (delegated /
strict). Helps reviewers correlate the visible_start with the mode.
- Top-level `chain.visible_start` (int): chain.VisibleStart(leaf, mode)
— the lowest level whose grants the leaf can see, accounting for any
inherit:false fence in delegated mode (always 0 in strict mode).
- Per-level `inherit` (*bool, omitempty): the level's explicit inherit
value, nil when absent. A reviewer can scan the levels and see which
one fences ancestors.
The level's `exists` flag now also fires for `permissions:` and
`inherit:` entries (previously it only checked Allow/Deny/Admins),
so the response correctly reflects modern .zddc files that use the
permissions map.
Test: TestServeProfileEffectivePolicy_InheritFence builds a vendor-
folder layout, asks the tracer about a my-company user, confirms
decision=false, visible_start=1 (fence at /Vendor/), leaf.Inherit=
&false, root.Inherit=nil.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
A .zddc may now declare `acl.inherit: false` to fence off ancestor
grants and roles from the descendant subtree — the "complete reset
plus add back" pattern operators want for vendor folders and other
narrowly-scoped subtrees. The cascade walker honors the deepest fence
in [0, toIdx] when evaluating any level at-or-below it, both for
GrantedVerbsAtLevel/EffectiveVerbsRange and for role lookup
(RoleMembers / lookupRoleMembers).
Federal/strict cascade mode IGNORES the fence — required by
NIST AC-6 ("ancestor deny is absolute; no leaf-level override"). So
inherit:false has no effect under strict mode and ancestor grants
remain visible. Operators running the federal Rego preset get the
same behaviour from external policy enforcement.
API surface: ACLRules.Inherit (*bool, nil = unset = inherit-true);
ACLRules.InheritsAncestors() bool; PolicyChain.VisibleStart(toIdx,
mode) int. The mode parameter is now threaded through
GrantedVerbsAtLevel, MatchesPrincipal, MatchingPrincipals,
RoleMembers, and lookupRoleMembers so role resolution is fence-aware.
Tests:
- file_test.go: parser round-trip for absent / true / false inherit
- inherit_test.go: VisibleStart (no fence, fence clamps, nested fences,
strict-mode override), EffectiveVerbs (fence hides ancestor grants,
strict-mode keeps them), RoleMembers (ancestor roles hidden by fence,
local redefinition still works)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three pieces wire the per-party Master Deliverables List as the default
view at archive/<party>/mdl/:
1. **Dispatcher redirect.** GET (and HEAD) on
<project>/archive/<party>/mdl/ (case-fold on archive and mdl) now
302 → <project>/archive/<party>/mdl.table.html. Non-archive paths
and deeper mdl/ paths fall through unchanged.
2. **Default-spec fallback in RecognizeTableRequest.** When a request
matches archive/<party>/mdl.table.html and no operator-supplied
tables: { mdl: ... } declaration covers it, the handler returns a
recognised request anyway. Operator declarations still win — and a
typo'd declaration pointing at a missing file yields 404 (not a
silent fallback).
3. **Static-file fallback for the spec yaml.** GET archive/<party>/
mdl.table.yaml and archive/<party>/mdl.form.yaml return embedded
default bytes (default-mdl.{table,form}.yaml in the handler package)
when no operator file exists at that path. Operator files always
win because the dispatcher's os.Stat finds them before reaching the
IsDefaultMdlSpec branch.
The defaults use ZDDC vocabulary: tracking, title, discipline, type,
plannedRevision, plannedDate, status (DFT/IFR/IFA/IFC/AFC/AB), owner,
notes. Operators override per-party by writing
archive/<party>/{mdl.table.yaml,mdl.form.yaml} and a tables: { mdl: ... }
entry in the party's .zddc.
Tests:
- 4 dispatcher redirect cases (success, case-fold mdl, case-fold archive,
deeper-path skip, non-archive skip)
- 6 tablehandler cases (default fires at archive/<party>/, operator
override wins, scope check, embedded yaml served, operator yaml wins,
scope check on yaml fallback)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ListDirectory now appends a synthetic <viewer-email>/ entry when the
listed path is exactly <project>/working/ (depth 2, case-fold) and no
real directory there matches the viewer's email under any case.
The entry has IsDir=true and a new Virtual=true flag on
listing.FileInfo (omitempty in JSON so existing clients that don't
know the field continue to render it as a regular folder). A first
write to that path materialises a real folder via the existing
auto-own pipeline (EnsureCanonicalAncestors → WriteAutoOwnZddc),
after which subsequent listings drop the synthetic entry naturally.
Anonymous viewers, listings outside working/, and listings inside a
deeper working/ subdirectory all skip the synthetic entry.
Six tests cover: appears-when-missing, suppressed-when-real-exists
(case-fold), anonymous-no-entry, staging/-no-entry, deep-working-no-
entry, and pre-existing-PascalCase-Working/ still triggers it.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
When a folder is created under <project>/staging/ whose name parses as a
ZDDC transmittal folder (YYYY-MM-DD_<tracking> (<status>) - <title>) and
whose tracking number contains -TRN- or -SUB-, also create the same-
named folder under <project>/working/ as a drafting space for staff.
The mirror is one-way and one-shot: created at staging-mkdir time only.
Renames and deletions of either side are not propagated. The
transmittal client orchestrates cleanup at issue time (move files to
archive/<recipient>/issued/, then delete both staging and working
siblings) — the server stays out of that decision.
-MDL- tracking deliberately skips the mirror; MDL deliverables live in
archive/<party>/mdl/ rows, not via the working↔staging pairing.
Implementation: mirrorStagingToWorking() in fileapi.go, called after a
successful serveFileMkdir. EnsureCanonicalAncestors handles working/'s
own auto-own .zddc; the mirror folder gets its own creator-grant on top.
Six new tests cover -TRN-/-SUB- mirror, -MDL- skip, non-transmittal
name skip, deep-path skip, and idempotency over a pre-existing sibling.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
New helper pair:
- ResolveCanonicalPath(fsRoot, target) — case-fold path resolution, no side effects
- EnsureCanonicalAncestors(fsRoot, target, email…) — case-fold + MkdirAll + auto-own .zddc seeding
For each canonical position along the requested path the helpers
substitute on-disk casing (so /Project/working/foo lands in an existing
Working/ rather than a new sibling) and materialise missing
working/staging/archive/<party>/{mdl,incoming,received,issued}/ folders.
working/, staging/, and archive/<party>/incoming/ get a creator-owned
.zddc seeded automatically; received/, issued/, and mdl/ are created
without auto-own (WORM and data-store concerns respectively).
reviewing/ is rejected — purely virtual, never on disk.
Wired into the file API:
- serveFilePut — resolve before auth, ensure after auth
- serveFileMkdir — resolve before auth, ensure after auth, with
two auto-own checks (target-is-canonical OR
parent-is-canonical)
- serveFileMove (POST) — resolve src+dst, ensure dst before rename so
a move from working/<draft> →
archive/<recipient>/issued/<draft> creates
the per-party folders on the way in
7 new unit tests in zddc/internal/zddc/ensure_test.go cover lazy
creation, case-fold reuse, per-party incoming auto-own, WORM no-auto-own,
empty-principal skip, reviewing rejection, and traversal rejection.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The transmittal-folder grammar was duplicated as a private regex inside
the archive package. Replace the local regex with calls to the shared
parser in zddc/internal/zddc/folder.go so the grammar lives in one
place and the upcoming staging→working mirror logic can reuse it.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
BREAKING CHANGE. Project-level Issued/Received/Incoming folders no
longer carry special semantics. WORM enforcement and auto-ownership
move to the per-party canonical layout:
- WORM mask now triggers on archive/<party>/received/ and
archive/<party>/issued/ (any case, any party)
- Auto-own .zddc writes on first mkdir under working/, staging/,
or archive/<party>/incoming/ (any case)
Predicate API:
- IsAutoOwnPath(parentDir, fsRoot) — replaces IsAutoOwnParent(name)
- IsWormPath(requestPath) — same name, new pattern
- WormFolderLevelIndex unchanged signature, new pattern
Legacy SpecialFolderNames / AutoOwnFolderNames / WormFolderNames /
IsAutoOwnParent are deleted (no Deprecated: stubs — early-development
project, no back-compat to preserve).
Tool availability (apps/availability.go) is case-fold throughout:
- mdedit: descendants of working/
- transmittal: descendants of staging/
- classifier: descendants of working/, staging/, or
archive/<party>/incoming/
Working/, WORKING/, working/ all match identically.
Test fixtures rewritten:
- special_test.go: covers IsAutoOwnPath / IsWormPath /
WormFolderLevelIndex / ResolveCanonical / canonical lists
- availability_test.go: per-party rules, case-fold scenarios
- fileapi_test.go: rolePermissionsTestSetup now seeds
Project-X/archive/Acme/{incoming,issued,received}/ rather than
Vendor/{Incoming,Issued,Received}/ at the project root
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Pure refactor. The mkdir post-hook in handler/fileapi.go duplicated
zddc-package types; lifting the body into the package itself lets the
upcoming EnsureCanonicalAncestors helper share it without re-exposing
the file API's internals.
No behaviour change. The grant shape (creator email → rwcda + CreatedBy
audit field) and the atomic-write path through zddc.WriteFile are
unchanged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Introduce the lowercase canonical folder model that the new auto-create
feature will key off:
- ProjectRootFolders = [archive, working, staging, reviewing]
- PartyFolders = [mdl, incoming, received, issued]
- AutoOwnCanonicalNames = [working, staging, incoming]
- VirtualOnlyCanonicalNames = [reviewing]
ResolveCanonical(parentDir, logical) does a case-fold lookup against
os.ReadDir(parentDir) so a manually-created Working/ is reused rather
than shadowed by a new working/ sibling.
Pure addition. The existing SpecialFolderNames / AutoOwnFolderNames /
WormFolderNames are kept (now Deprecated:) so dependent packages keep
compiling until the predicate rewrite lands.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Extracts the YYYY-MM-DD_<tracking> (<status>) - <title> grammar into a
reusable parser in the zddc package, and exposes a tracking-type
predicate for -TRN- / -SUB- (case-fold). The transmittal-folder regex
was previously only inside archive/index.go where it captured just the
date; the new ParseTransmittalFolder also returns tracking, status, and
title so handlers can recognise transmittal envelopes for upcoming
staging↔working mirror logic.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The .archive virtual prefix is now project-scoped at exactly one URL
depth: any /<project>/<sub>/.../.archive/... gets a 301 to the
canonical /<project>/.archive/.... The dispatcher does this before
calling the handler; query strings are preserved (the browser handles
the fragment automatically). .archive is also GET/HEAD-only — anything
else returns 405 with Allow: GET, HEAD, ahead of the file API.
Why: offline-built HTML files reference siblings as
"../.archive/<tracking>.html" from arbitrary depths. All of those refs
should converge on a single stable URL per (project, tracking) so
external links and bookmarks don't fork by entry point.
Permissions now follow the resolved file, not .archive itself.
.archive is a virtual surface — it has no on-disk directory and no
.zddc of its own, so gating it as if it did is wrong. Two gates only:
- Resolve: only the per-target file's ACL chain decides. A user
explicitly allowed at one transmittal folder but denied at the
project root can still fetch tracking numbers that resolve there.
Per-target denial returns 404 (not 403) so existence doesn't leak.
- Listing: filter entries by per-target ACL. If the project bucket
has zero indexed entries → 404 (unknown / empty project, indistinguishable
from a probe). If the bucket is non-empty but the caller can read
no entries → 403 (existence-leak guard: don't confirm an inaccessible
project's archive exists). Otherwise → 200 with the filtered subset.
The listing endpoint is now content-negotiated like ServeDirectory:
Accept: text/html serves the embedded `browse` SPA bytes (with the
embedded ETag and X-ZDDC-Source: embedded:browse); Accept:
application/json returns the JSON entry array (with content-hash ETag
and 304 short-circuit). Vary: Accept set on both. The browse SPA's
auto-detect path-fetch then renders the archive entries as a sortable,
filterable flat list at /<project>/.archive/.
ServeArchive's signature is now (cfg, idx, w, r, project, filename) —
the dispatcher hands the normalized project string in directly, so
projectFromContextPath is gone. Old behavior was to derive project
from contextPath inside the handler; with the upstream redirect that's
redundant and the handler's preconditions are simpler.
Tests: archivehandler_test.go rewritten around the new semantics;
added per-target-only resolve, project-root-deny + per-target-allow
rescue, listing 403/404 distinction, JSON/HTML content-negotiation,
and conditional GET. main_test.go gains TestDispatchArchiveRedirect
(deep paths, query preservation, already-canonical no-op) and
TestDispatchArchiveMethodGate (PUT/POST/DELETE → 405).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The fsnotify watcher only sees events the local kernel generates, so on
SMB/CIFS-backed roots (Azure Files) writes from any other client are
invisible — the archive index would silently miss them until pod
restart. Add two backstops:
1. Periodic full re-walk via Index.Rebuild on a configurable interval
(--archive-rescan-interval / ZDDC_ARCHIVE_RESCAN_INTERVAL, default
60s, 0 to disable). Atomically swaps ByProject under the existing
RWMutex; concurrent reads stay safe.
2. Admin-only POST /.profile/reindex that triggers an immediate rebuild
and returns {duration_ms, project_count, tracking_count}, for the
"I just dropped 50 files and don't want to wait" case. Gated by
IsAdmin with the same 404-on-non-admin pattern as the other admin
sub-resources.
Tests: TestRebuild_PicksUpAddsAndDrops covers add+drop semantics and
returned counts; TestServeProfileReindexPOST covers the happy admin
path; matrix entries cover the gate (anonymous/non-admin → 404, admin
GET → 405 method-not-allowed since the route is POST-only).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Refreshes //go:embed bytes off fe28a73. Dev image now ships the new
tables renderer (handler-local embed at handler/tables.html) plus
build-label refresh on the six cascade-served tools.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Resolved `.archive/<tracking>.html` URLs now serve the target file's
bytes inline via http.ServeFile with Cache-Control: no-cache, replacing
the previous 302 redirect to the per-transmittal URL.
Why: external links like `.archive/<tracking>.html#section` are meant
to track the latest revision. A redirect exposes the snapshot URL — any
forwarded link then pins to that snapshot instead of "latest." Serving
in-place keeps the `.archive/` URL stable as the resolver's "current"
target moves over time.
Cache-Control: no-cache is intentional. Each load revalidates against
the on-disk file's Last-Modified/ETag, so when a new revision lands the
resolver picks it and the browser refetches transparently.
ACL is unchanged: enforced on both the `.archive` context directory and
the resolved target file (per-target denial returns 404, not 403, to
avoid disclosing that a tracking number exists in a hidden subtree).
archivehandler_test.go status expectations updated 302 → 200; fixture
bodies adjusted for body-content verification of the in-place serve.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Tables is the eighth HTML tool: a read-only tabular view over a
directory of YAML files declared via `tables:` in `.zddc`. Anchor use
case is the Master Deliverables List, where each row is one
`<tracking>.yaml` under `Archive/<Party>/MDL/`. Rows click through to
the existing form renderer for editing.
Schema (zddc/internal/zddc/file.go)
- New `Tables map[string]string` on ZddcFile. Map key becomes the URL
stem (`tables[MDL]` → `<dir>/MDL.table.html`); the value is a path
relative to the .zddc pointing at a `*.table.yaml` spec describing
columns + the rows directory. No upward cascade in v1 — each
directory hosting a table declares it directly.
Server handler (zddc/internal/handler/tablehandler.go)
- `RecognizeTableRequest` matches GET `/<dir>/<name>.table.html`
against the cascade's `tables:` declarations. Dispatch routes
table requests before the form-system intercept.
- `ServeTable` ACL-gates with `policy.ActionRead` and serves the
embedded `tables.html` template; client walks the directory itself
via the listing JSON or FS Access API.
- tables.html embedded via //go:embed — same pattern as form.html.
Frontend (tables/)
- Vanilla JS: app/context/util/filters/sort/render/main modules.
- Reads spec + row YAML files via window.zddc.source (HTTP polyfill
or local FS handle); js-yaml 4.1.0 vendored in shared/vendor for
client-side parsing.
- Sample fixtures under tables/sample/ for local testing.
Build + CI
- Lockstep build registers tables alongside the other 7 tools (HTML
output, embed mirror, versions.txt, release-output, tags).
- Playwright project added; `npx playwright test --project=tables`
is part of `npm test`.
Drive-by: rename mdedit Playwright selectors `#select-directory` →
`#addDirectoryBtn` to fix three pre-existing failing tests.
Drive-by: ignore locally-built `zddc/zddc-server` binary so it doesn't
get accidentally staged.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Refreshes the //go:embed bytes off 3115e38. Dev image (ZDDC_REF=main)
now ships the file API and verb-based RBAC.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Replaces the binary acl.allow/deny model with five permission verbs
(r/w/c/d/a) and first-class roles, and adds an authenticated file API
(PUT/DELETE/POST move/mkdir) so the HTML tools can edit-in-place over
HTTP. Closes the AC-3(7) and AC-6 federal-readiness gaps.
File API (zddc/internal/handler/fileapi.go)
- PUT <new> → action c
- PUT <existing> → action w
- PUT <.zddc> → action a (CanEditZddc strict-ancestor rule)
- DELETE → action d
- POST mkdir → action c (auto-writes creator-owned .zddc when the
parent is Incoming/Working/Staging)
- POST move → action w on src + c on dst, atomic via os.Rename
- Optional If-Match for optimistic concurrency, --max-write-bytes cap,
audit log emits a structured file_write event per operation.
Permission model (zddc/internal/zddc/{acl,file,roles,cascade_mode}.go)
- acl.permissions: { principal → verb-set } map; principals are email
patterns or role names. Empty verb set is an explicit deny.
- roles: { name → members } definitions, available at the level they
declare and all descendants. Closer-to-leaf shadows ancestor.
- Legacy acl.allow/deny still work; they fold into permissions at
parse time (allow → "rwcd", deny → "").
- Cascade walks leaf→root; first level with any matching entry wins;
the union of matching verb sets at that level decides.
- --cascade-mode=strict adds a root→leaf ancestor-deny pre-pass so an
ancestor explicit-deny is absolute (NIST AC-6). Default delegated
preserves the existing commercial behavior.
Special folders (zddc/internal/zddc/special.go)
- Incoming / Working / Staging: mkdir auto-writes a .zddc into the new
subdir granting created_by + that email rwcda directly. Same form
operators write by hand; creator can edit it later to add others.
- Issued / Received: server-enforced WORM split. Cascade grants
inherited from above the WORM folder are masked to r only; grants
placed at-or-below the WORM folder retain r,c. Operators grant
write-once (cr) to the doc controller via an explicit .zddc at the
Issued/Received folder. Admins exempt — only escape hatch.
Browser polyfill (shared/zddc-source.js)
- HttpDirectoryHandle + HttpFileHandle implement the FS Access API
surface (values, getFileHandle, createWritable, removeEntry,
queryPermission/requestPermission) over zddc-server's listing JSON
and file API. Existing tools written against showDirectoryPicker
work unchanged.
- detectServerRoot() returns { handle, status }: tools auto-load on
HTTP, surface a clear "no permission to list" message on 403, and
fall back to the welcome screen on 0.
- classifier renames take the atomic POST move path on HTTP-backed
handles; mdedit and transmittal route reads/writes through the
polyfill so prior FS-API code paths cover both modes.
Tests
- zddc/internal/zddc/{cascade_mode,roles,special,acl}_test.go cover
delegated vs strict, role membership / shadowing / legacy fallback,
WORM split semantics, verb-set parser round-trip.
- zddc/internal/handler/fileapi_test.go now also covers role-based
vendor scenarios, WORM blocking vendor & doc controller writes,
explicit Issued .zddc unlocking the cr drop-box, admin bypass,
auto-ownership on mkdir, and strict-mode lockouts.
Docs
- ARCHITECTURE.md + zddc/README.md document the verb model, role
syntax, special-folder behaviors, cascade-mode flag, and full file
API surface. Federal-readiness gap analysis strikes AC-3(7) and
AC-6.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Adds a second way to configure the apps signing pubkey alongside the
existing --apps-pubkey / ZDDC_APPS_PUBKEY (path-to-PEM-file) form: an
inline PEM block under apps_pubkey: in the root .zddc file. Resolution
order:
1. --apps-pubkey / ZDDC_APPS_PUBKEY (path) ← env/flag wins
2. apps_pubkey: inline PEM in root .zddc ← second
3. nothing ← URL fetches refused
Honored only at the root .zddc — same trust-anchor treatment as the
existing admins: field. Subtree write authority cannot re-anchor
trust because subtree apps_pubkey: entries are ignored. (Same
unmarshal pattern as the rest of ZddcFile; the root-only enforcement
is in setupApps where we explicitly read filepath.Join(cfg.Root,
".zddc") rather than walking a chain.)
Why offer both: env/flag fits k8s + systemd deployment shapes where
the operator already manages a config volume and prefers env-based
plumbing. Inline-in-.zddc fits the "everything in one config file"
mental model and matches how operators already think about admins:
and acl:. Either ships a working URL-fetch-verify story; the choice
is operator preference.
Logged differently per source so operators can grep for which path
populated the key:
apps signing pubkey loaded source=env/flag path=/path/to/pubkey.pem
apps signing pubkey loaded source="root .zddc apps_pubkey"
Smoke-tested end-to-end: a root .zddc with inline apps_pubkey: PEM
block + apps: archive: <upstream-URL> + ZDDC_APPS_PUBKEY unset —
the server logs "loaded source=root .zddc apps_pubkey" at startup,
fetches the URL, verifies the .sig against the inline key, caches.
Tampering still rejects; missing .sig still rejects; everything that
worked yesterday still works.
Docs: env-var tables in zddc/README.md and AGENTS.md note the
inline alternative; the federal-readiness gap analysis subsection
on code signing now lists both paths in its resolution order; the
release-page "Verify your downloads" section mentions both for
operators.
Production binary unchanged at ~13 MB. All 11 Go test packages green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two interlocking pieces shipped together:
1. Strict Ed25519 signature verification on URL-fetched apps artifacts.
Every URL the apps cascade resolves must publish a corresponding
<url>.sig (raw 64-byte Ed25519 signature). The fetcher rejects on
any failure (sig 404, transport error, wrong key, tampered body)
and the resolver falls back to the embedded copy.
The trusted public key is OPERATOR-CONFIGURED via --apps-pubkey /
ZDDC_APPS_PUBKEY (PEM file path). No baked-in default — same posture
as TLS certificates. Operators using zddc.varasys.io's canonical
channels download pubkey.pem from there and configure the local
path. Operators with their own signing infrastructure pass their
own public key.
Build pipeline (./build) gains sign_release_artifacts: walks
dist/release-output/ after promote and produces an Ed25519 .sig
alongside every real file. ZDDC_SIGNING_KEY=~/.config/zddc-signing/
key.pem (mode 0600). Symlinks skip — the .sig at the symlink
target is what counts.
Test coverage: parse-PEM round-trip, malformed/wrong-type PEM
rejection, valid-signature accept, tampered-body reject, wrong-key
reject, malformed-signature reject, end-to-end fetch+sign+verify,
fetch-rejects-tampered, fetch-rejects-missing-sig, fetch-rejects-
wrong-key. Existing fetch tests updated to use signed-fixture
helpers.
2. Dev Helm chart mounts production data READ-ONLY and layers an
OverlayFS writable scratch on top. Prod data is the lowerdir;
dev's writes (form submissions, archive index state, .zddc edits)
land in upperdir; main container sees the merged read-write view
at $ZDDC_ROOT. Setup runs in a privileged init container; main
container runs unprivileged. Solves the dev-replica-on-shared-
dataset problem at the filesystem layer with no zddc-server code
change.
Docs: env-var tables in zddc/README.md and AGENTS.md gain a
ZDDC_APPS_PUBKEY row. The Federal-readiness gap analysis "Code-signed
apps: URL fetches" subsection is rewritten as "what's currently in
place" instead of "what would need to be added," with a forward
pointer to per-entry signed_by: (multi-key) and Sigstore as the
federally-acceptable evolution.
The website "Verify your downloads" section + the embedded pubkey
gone — but the website needs separate updates landing in zddc-website
to publish pubkey.pem and add the verify section. Pending in that
repo's commit.
Production binary unchanged at 13.1 MB. All 11 Go test packages green.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Ship a second parity-tested Rego policy that flips the cascade's
leaf-allow-overrides-parent-deny rule for NIST AC-6 conformance.
Standard cascade (existing access.rego, mirrors internal Go evaluator):
Bottom-up walk; first explicit match wins; deny-first within a level.
A leaf-level allow CAN override an ancestor's deny. This is the
cascade's intentional delegation property — a project-owner who
re-allows a previously-denied collaborator works as expected.
Federal mode (new access_federal.rego):
Any deny anywhere along the chain is absolute. An allow only matters
if no level (any depth) has denied the same email. Required by
NIST AC-6 default expectations: a central admin's deny at the root
must be unbypassable by a tenant who controls a subtree's .zddc.
Operators run real OPA with this Rego and point ZDDC_OPA_URL at it;
the internal Go evaluator stays on the commercial cascade. The
toggle is "which policy does your OPA evaluate," not a knob inside
zddc-server.
Surfaced via --print-rego flag:
zddc-server --print-rego # standard (default)
zddc-server --print-rego=standard # same
zddc-server --print-rego=federal # AC-6 strict variant
Parity test (federal_parity_test.go) compiles both Regos and asserts:
* They AGREE on every cascade scenario where no ancestor-deny
intersects a leaf-allow (most cases).
* They DISAGREE — by design — on the three scenarios where the
AC-6 rule differs:
- "leaf allows what parent denied" → standard allows, federal denies
- "deep leaf re-allows after middle deny" → same
- "glob deny at root + specific allow at leaf" → same
Cross-checks the divergence flag explicitly so any future change that
accidentally collapses the two policies fails the test.
Closes the AC-6 row of the federal-readiness gap analysis (now marked
"partially complete" in zddc/README.md — the full bullet would be a
built-in --policy-mode=federal toggle that also flips the in-process
Go evaluator).
Production binary unchanged at 13.1 MB (Rego files embedded as bytes;
OPA library remains test-only).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Eliminates the manual cascade-trace ritual when debugging "why can't
alice see /Project-X" reports. New endpoint returns the resolved
policy chain plus the active decider's verdict in JSON:
GET /.profile/effective-policy?path=/Project-X/sub/&email=alice@…
Response shape:
{
"path": "/Project-X/sub/",
"email": "alice@…",
"decision": true,
"decider_kind": "*policy.InternalDecider",
"chain": {
"has_any_file": true,
"levels": [
{"index": 0, "zddc_path": "/.zddc", "exists": true,
"acl": {...}, "admins": [...],
"matches_email": false, "decision_at_level": "no_match"},
{"index": 1, "zddc_path": "/Project-X/.zddc", "exists": true,
"acl": {...}, "matches_email": true, "decision_at_level": "allow"}
]
}
}
Per-level email matching reuses the same MatchesPattern code the live
evaluator uses, so the trace can never disagree with the actual
verdict — and when ZDDC_OPA_URL points at an external OPA, the
decision goes through that OPA, making the endpoint a useful smoke
test for OPA wiring too.
Admin-only via the existing /.profile gate (404 to non-admins).
Required params; 400 if either is missing or path doesn't escape ROOT.
Test coverage:
* TestServeProfileGateMatrix: anonymous → 404, non-admin → 404,
admin without params → 400 (gate cleared, validator rejected)
* TestServeProfileEffectivePolicy: full payload-shape assertion
against a worked-example fixture (closed project where alice is
allow-listed but bob is not)
Also fixes pre-existing doc drift: README's "Admin Debug Page"
section referenced /.admin/whoami|config|logs but the actual code
mounts /.profile/* (the rename predates this PR; the doc was stale).
Closes the "/.admin/effective-policy debug endpoint" item from the
federal-readiness future-work list.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The previous keyForURL stripped default ports (:443 for https, :80
for http) and omitted the scheme, so:
http://example.com/x.html ──┐
https://example.com/x.html ──┴──→ same cache entry (collision)
https://example.com/x.html ──┐
https://example.com:443/x.html ──┴──→ same cache entry
This was a defensible HTTP convention but a real correctness issue
on reverse-proxy stacks where http and https legitimately serve
different bytes for the same path, or where two upstreams share a
host but answer on different default ports.
New layout: <scheme>/<host>[:<port>]/<path>. Full origin tuple in
the key, no port stripping, scheme segregation. Examples:
https/zddc.varasys.io/releases/archive_stable.html
https/example.com:8443/x.html
http/example.com/y.html (distinct from https/example.com/y.html)
Operators retain the "ls _app/ to inspect what's cached" affordance
they relied on; they just see one extra directory layer (scheme
first, then host).
Tests:
* Updated TestKeyForURL to assert the new layout for every
previously-covered case
* New TestKeyForURL_NoCollisions explicitly asserts that the
dimensions previously collapsed (default-port↔bare, http↔https,
different non-default ports) now produce distinct keys
Doc references to the cache layout under <ZDDC_ROOT>/_app/ updated
in zddc/README.md (3 mentions).
NOTE: existing _app/ caches under the old layout will be ignored
on first request after upgrade — entries will be re-fetched and
written to the new path. Operators can `rm -rf <ZDDC_ROOT>/_app`
during the upgrade window if they prefer not to have orphans.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The TLS configuration was using Go stdlib defaults — secure for typical
commercial use, but federal evaluators need an explicit cipher
allowlist they can map to a FIPS-validated implementation. Pin the
cipher and curve lists to NIST SP 800-52 Rev. 2 § 3.3 conformant
values:
Ciphers (TLS 1.2):
TLS_ECDHE_ECDSA_WITH_AES_128_GCM_SHA256
TLS_ECDHE_ECDSA_WITH_AES_256_GCM_SHA384
TLS_ECDHE_RSA_WITH_AES_128_GCM_SHA256
TLS_ECDHE_RSA_WITH_AES_256_GCM_SHA384
TLS_ECDHE_ECDSA_WITH_CHACHA20_POLY1305
TLS_ECDHE_RSA_WITH_CHACHA20_POLY1305
Curves: X25519, P-256, P-384
MinVersion: TLS 1.2 (already set; 1.3 used when negotiated)
TLS 1.3 cipher selection is not operator-controllable in Go stdlib
(the runtime picks from a fixed AEAD-only set); all of those
already meet the federal bar so no change needed there.
Also adds HSTSMiddleware emitting `Strict-Transport-Security:
max-age=31536000; includeSubDomains` when zddc-server is itself
terminating TLS (ZDDC_TLS_CERT != none). Behind an upstream proxy
terminating TLS the proxy is responsible for HSTS, so the middleware
only wraps the chain when useTLS=true.
Test coverage:
* TLSConfig(none) returns nil + useTLS=false
* TLSConfig(selfsigned) sets the exact NIST allowlist
* Negative test asserting weak ciphers (CBC, RC4, 3DES, RSA-key-
exchange) are NOT in the list — guardrail against regressions
Federal-readiness gap analysis updated: this control is now partially
complete. OCSP stapling and CT-log inclusion remain on the list for
full DoD STIG conformance.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Phase 2 enhancements to the policy decider, plus listing-level ETags
that benefit every deployment regardless of decider mode.
Reference Rego policy
---------------------
internal/policy/rego/access.rego mirrors InternalDecider's semantics
exactly — bottom-up walk, deny-first within a level, default-deny when
HasAnyFile=true, glob matching with @-boundary semantics (special-cased
bare "*" because OPA's glob.match treats empty delimiters
inconsistently for that pattern).
Embedded into the binary via go:embed; --print-rego dumps it to stdout
so federal customers standing up an external OPA can use it as a
parity-tested baseline:
zddc-server --print-rego > /etc/opa/policies/zddc-access.rego
Parity test runner
------------------
parity_test.go imports the OPA Go module as a TEST-ONLY dependency
(github.com/open-policy-agent/opa@v0.70.0). Every fixture from the
internal Go evaluator's test set runs through both implementations;
any divergence fails CI. The test-only import means production
binaries (built by `go build ./cmd/zddc-server`) stay OPA-free —
release-flag binary size unchanged at ~13 MB.
The parity test caught a real bug on first run: bare "*" patterns
didn't match through OPA's glob.match with empty delimiters. Fixed
in access.rego with a special-case rule. This is exactly the kind of
subtle drift the parity guard exists to catch.
External-mode decision cache
----------------------------
HTTPDecider is now wrapped in a cachingDecider with a default 1s TTL.
Bursty patterns like .archive listings (one OPA round-trip per entry
before, one per (email, decision-input) tuple per TTL window after)
amortize cleanly. Verified: 20 identical /D/ requests produce 1 OPA
hit with cache, 40 hits without (each listing makes 2 ACL queries).
ZDDC_OPA_CACHE_TTL knob (default 1s) lets operators tune. 0 disables.
1s matches the fsnotify watcher debounce window — staleness is
bounded the same way other policy-edit propagation already is.
Internal mode unchanged; the in-process Go evaluator is already
cheaper than a cache lookup would be.
Listing ETags
-------------
GET / (project list) and GET /<dir>/ (directory listing JSON) now
carry content-hash ETag + Cache-Control: private, max-age=0,
must-revalidate. SHA-256 of the rendered JSON, truncated to 16 hex
chars (64 bits — collision risk on a listing of any realistic size
is vanishingly small).
Server-side caching deliberately not added: it would require
mtime-based invalidation, and Azure Files SMB mounts (a common
deployment substrate) don't support fsnotify reliably. The
content-hash ETag delivers the bandwidth savings (304 on identical
fetches) without depending on watcher correctness — the hash is the
actual response, so it can't lie about staleness regardless of
underlying watcher behavior.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add an internal access-decision boundary that all handlers go through
instead of calling zddc.AllowedWithChain directly. Two implementations
ship:
* InternalDecider — wraps the existing zddc.AllowedWithChain. The
default. No new dependencies, identical semantics to the legacy
code path. ZDDC_OPA_URL=internal (or unset).
* HTTPDecider — POSTs the canonical OPA wire format
(POST /v1/data/zddc/access/allow with {"input": {...}}, response
{"result": true|false}) over HTTP, HTTPS, or a Unix-domain socket.
For federal customers running their own audited Rego policies
alongside zddc-server. ZDDC_OPA_URL=http(s)://… or unix:///….
External-mode failure semantics: unreachable / non-2xx / malformed
response → fail closed (deny) by default with a WARN log. Operators
who prefer availability over correctness flip with ZDDC_OPA_FAIL_OPEN=1.
The decider is constructed once at startup, plumbed through ACLMiddleware
into the request context. Handlers retrieve it via DeciderFromContext;
non-request callers (fs.ListDirectory, EnumerateProjects, enumerateAccess)
take it as an explicit parameter.
zddc.ZddcFile and zddc.ACLRules gain JSON tags so external Rego authors
get idiomatic input shape (acl.allow, admins, …) instead of Go field
names (ACL.Allow, Admins, …).
Test coverage:
* InternalDecider parity tests against zddc.AllowedWithChain (every
documented cascade scenario: empty chain, leaf-allow-wins, leaf-
deny-beats-parent, leaf-allows-what-parent-denies, deepest-match-
wins, etc.)
* HTTPDecider happy-path test (canonical wire format)
* Fail-closed / fail-open / malformed-response tests
Production binary size unchanged (no new deps; HTTP transport is
stdlib net/http). 11 ACL call sites migrated. End-to-end verified
against the worked-example layout in zddc/README.md.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two safe-by-default flips, both opt-out via explicit acknowledgement.
1. --insecure / ZDDC_INSECURE=1: zddc-server now refuses to start when
no <ZDDC_ROOT>/.zddc exists. With no .zddc anywhere in the chain,
AllowedWithChain falls through to "HasAnyFile=false → allow" and
the tree is publicly accessible to anonymous callers — almost never
what an operator wants on a fresh deployment, and previously a
silent footgun. The flag is the escape hatch for deliberately-
public archives (no .zddc anywhere by design).
2. ZDDC_CORS_ORIGIN now defaults to empty (CORS disabled) instead of
the canonical "https://zddc.varasys.io". The embedded-tools install
path serves tools and data same-origin, so the default never needed
to permit cross-origin XHRs from a third-party host. Every deployment
was implicitly trusting zddc.varasys.io to make authenticated XHRs
on behalf of every logged-in user; if that origin were ever
compromised, the blast radius extended to every customer server.
Operators who deliberately use the CDN-bootstrap pattern or self-
hosted tools at a different host now set the value explicitly.
Helm chart values updated accordingly: prod default is empty; dev
keeps localhost:8000 for tool-iteration workflows. Existing deployments
that depended on the old defaults will need to either set the value
explicitly or pass --insecure.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The recent history rewrite (squash of 4 thrash CI commits) made the
chart's previous appVersion (0.0.16-beta-8df0def) reference a now-
dangling commit. The dev pipeline failed clone "remote ref not
found" until we re-bumped to a SHA in the new history.
Re-cut beta with the new HEAD parent (ae75855) so notify-chart-dev
rewrites the chart's appVersion to a SHA the BMCD dev pipeline can
actually fetch. Combined with the Dockerfile clone-via-fetch fix in
tnd-zddc-chart 86c5758 (handles bare SHAs), the dev pipeline should
build cleanly.
Two related operational improvements:
1. HTTP timeouts on http.Server (ReadHeaderTimeout 10s, ReadTimeout +
WriteTimeout 60s, IdleTimeout 120s). Caps slow-client connection
hold time; closes the slowloris vector. Listing + tool-HTML
responses complete in milliseconds even with gzip, so 60s is
generous for legit traffic.
2. --access-log defaults to <ZDDC_ROOT>/.zddc.d/logs/access-<host>.log
instead of stderr-only. The server auto-creates the parent tree
(mode 0750), so a fresh deployment gets an audit trail without
operator setup. Every JSON record carries a `host` field (from
os.Hostname) — multi-replica deployments share the .zddc.d/logs/
directory but write to per-host filenames, and downstream
aggregators can disambiguate via the host field.
Opt-out: --access-log= (explicit empty). Distinguishing "unset"
from "set to empty" follows the same pattern config.go already
uses for --cors-origin.
Live verification:
$ zddc-server -root /tmp/r -addr 127.0.0.1:8765 -tls-cert none -insecure-direct
$ curl http://127.0.0.1:8765/
$ ls /tmp/r/.zddc.d/logs/
access-bizon.log
$ tail -1 /tmp/r/.zddc.d/logs/access-bizon.log
{"time":...,"level":"INFO","msg":"access","host":"bizon",...,"email":"anonymous","method":"GET","path":"/","status":200,...}
$ zddc-server -root /tmp/r ... -access-log= # opt-out
$ ls /tmp/r/.zddc.d/ # empty: no logs/ created
