ds-review-bot round 3. The seam package broke the service-package export
contract (named export only), so the config catalog filed it under Other
libraries and default imports failed; it now default-exports DirectoryPicker
like every abstract seam, and the regenerated catalog lists it as one.
The picker-kind effect also let a settlement from a superseded flow open
leak into the current one (close/reopen mid-describe, or a reconnect that
swaps the backend): the read now resets the affordance on every open and a
cleanup-toggled flag discards obsolete settlements, both directions pinned
by jsdom races.
ds-review-bot round 1: the seam documented a merge-extensible union but
shipped a closed alias, and the gateway schema rejected any kind beyond
dialog/browse — a third backend could neither implement the seam nor be
advertised. The union now derives from an augmentable DirectoryPickerCapabilities
map, host.describe.directoryPicker preserves unknown wire kinds, and the
browse fixture applies listDirectory's root special case so creating under
'/' no longer mints a '//name' identity.
Fetch-Metadata and Origin are only attached to trustworthy destinations, so
over plain HTTP a rebound page's same-origin GET (EventSource, images,
navigations) arrives with no browser markers and a readable response. Remove
the marker shortcut; non-browser clients pass the same fence via loopback,
the CLI-derived LAN IP literals, or a declared authority.
A dangling colon (harness.internal:) or zero-padded port parses cleanly while
WHATWG silently rewrites it, turning an intended exact-port grant into an
any-port grant. Replace the character blacklist with a round-trip check: an
entry must read back from parsing exactly as written (case aside), refusing
the whole rewrite class at load.
WHATWG parsing would quietly read a hostname out of harness.internal/path or
user@harness.internal, authorizing the typo's hostname; other typos would sit
silently ignored until requests 403. Refuse every URL part beyond host[:port]
at plugin load.
The web GUI's folder picking was hardwired to one interaction: a native
OS chooser compiled into the gateway, unusable for remote deployments
and swappable only by editing apiproxy source.
Directory picking becomes a three-package capability seam in
packages/host: ctx.directoryPicker returns a discriminated capability —
dialog (the extracted native chooser; host-display only) or browse
(new: one-level listing + child creation over Node stdlib, hidden flags
host-stamped, symlinks followed, ancestry crumbs; remote-capable). The
gateway injects the seam, advertises the kind via
host.describe.directoryPicker, serves host.listDirectory /
host.createDirectory under browse, and answers
directory-picker-unavailable across kinds. cordis.yml is the swap
point; apps/cli keeps dialog mounted, so behavior is unchanged until
the in-app browser PR flips the default. The connection fixture serves
a deterministic browse tree; WorkspacesService gains the browse calls
the browser UI will drive. Decision record:
.agents/notes/implemented/architecture/2026-07-28-directory-picker-capability-seam.md
Markerless requests pass on any Host (a non-browser sender is the principal
and forges headers anyway); browser Host matching gains port-less entries and
WHATWG normalization; dsh derives LAN IP-literal authorities for an
all-interfaces bind and web grows --trusted-host for named ones.
The only browser-trust guard covered host.pickDirectory, while the
consequential methods (session.prompt drives bash) accepted any Host —
open to DNS rebinding, where a rebound page reads and writes the API as
if same-origin and only the Host header betrays the attacker's domain.
The pickDirectory-specific loopback guard becomes a prefix-wide fence:
Host must be loopback or an exact host[:port] from the new trustedHosts
config, an attached Origin must equal that authority, and explicit
cross-site markers are refused; requests without browser markers (curl,
tests, native clients) pass, because without a browser there is no
confused deputy. The loopback-socket check is dropped — binding policy
expresses reachability, and the fence is not an auth layer. The Agent
Note records the full threat model and the alternatives.
Remove the duplicate model-facing list tool from this branch; directory orientation remains available through bash ls. Keep the glob sampling fix, add a real ACP composition snapshot, and narrow the decision record to the shipped bug fix.
Master had advanced 392 commits. Every conflict was a derived artifact both
sides had regenerated — `docs/tool-catalog.md`, four recorded headless
session logs, and one TUI terminal expectation — so each was resolved by
taking master's version and re-running the generator and the keyless
snapshot refresh on the merged tree. No source file conflicted.
The new `examples/jsonrpc-agent` fixtures keep master's scrubbed
`{{system}}` / `{{tools}}` headers: that suite scrubs request headers, so a
new tool changes nothing there and the refresh output was discarded.
Asked what a workspace contained, an agent described one subfolder as the
whole project. `glob {"pattern": "*"}` matched 10030 paths across 22
top-level entries and the inline page was the first 100 of them, all under a
single unpacked archive.
Three properties compose into that page: a pattern with no `/` matches
basenames at any depth, so `*` means the whole tree rather than its top
level; `--sort=modified` orders oldest first, and unpacking an archive
restores timestamps that predate everything the user wrote; and the page was
the head of that order. Each is defensible alone, and together they make the
most ordinary request an agent receives produce a confident wrong answer.
A result within `globMaxResults` is unchanged — shown whole, in
modification-time order. Beyond it the page is filled round-robin across the
complete result's top-level entries, so one subtree cannot own every slot,
and the footer states that the page was sampled rather than taken in
modification-time order. Measured on a 24-entry, 716-file reproduction, the
head of 100 reaches 7 top-level names and the sampled page reaches 21. The
spill artifact still holds the complete sorted list.
The guidance and schema stop steering away from `ls`, state the any-depth
pattern rule, say results are files and never directories, and point at
`list` for a directory's contents.
`ctx.fs.listDir` has shipped since the filesystem seam gained it, with skill
discovery as its only consumer; the model-facing tool was deferred to a
separate decision. Nothing else could answer "what is in this directory":
`rg --files` backs glob and grep and never emits a directory entry, so an
empty directory is invisible, no output says which names are directories,
and no output gives an entry count.
`list` takes an optional `path`, defaulting to the session workspace so the
common question needs no argument, and returns the direct children of one
directory with their type. Two presentation rules carry it: directories sort
first, then files, then non-regular children, each alphabetically — so
truncation loses leaves rather than the tree — and the footer always states
the complete listing's size and composition, so a capped view can never read
as a whole directory.
It emits no `fs/observed`: seeing a filename is not reading a file, and a
listing must never satisfy the read-before-write gate.
The master merge left the generated catalogs/module graph stale, the
python runtime closure missing dsh-session-projection (now reached
through session-title and tool-todo), and apiproxy holding a dead
session-title dependency (the bespoke title frame is retired). The four
files the merge pushed under the per-file coverage floor (commands
executor + invariant, projection registry drive tails, TUI) join the
existing TODO(gui) deferral block per the GUI-lane policy; the remaining
coverage-run failures reproduce identically on pure origin/master
(environment-bound suites: sdk process exit, TUI PTY timing, workflow
worker timing, title loader slow-boot) and are not this branch's debt.
