Implement the continuable background subagents RFC: a durable child
session with a series of Task-backed activations, each disposing its
run before the Task settles.
- dsh-subagent: rename SubagentRun.sendMessage to strict steer, drop
run-level resume, add SubagentProvider.resume dispatch via
SubagentService.resume, the continuation start field, and the
versioned model-hidden subagent/descriptor session event.
- dsh-subagent-inprocess/-spawn/-fork: publish the control-allocated
child id, append the descriptor inside the initial turn, implement
cold resume from the child's own transcript under the live parent
scope, and strict running-only steer.
- dsh-subagent-control (new): SubagentControlService owning stable
child ids, descriptor snapshot/fold/authorization, Task-backed
activation with settle-then-dispose ordering, the process-local
active-run association, and steer-or-resume sendMessage routing.
- dsh-tool-subagent: background route branches on the provider's
resume capability (continuable via the control service; one-shot
task for ACP), returning both child and task ids.
- dsh-tool-subagent-control (new): the globally named send_message
tool rendering steered/started routes.
Keyless coverage spans Task ownership and disposal ordering, running
delivery, cold follow-up, descriptor rejection and rollback, known-id
reconstruction, kill during lookup, admission races, and a new
subagent-continuable ACP snapshot scenario.
Windows-native execution foundation: PwshLocalExecutor implements the bash
executor seam over ctx.subprocess (pwsh -NoLogo -NoProfile -NonInteractive
-Command, one argv element, no quoting layer; resolvePwshPath probes
PowerShell 7 / PATH / Windows PowerShell 5.1 as a pure function), and
tool-pwsh is the minimal PowerShell-dialect model-facing tool over ctx.bash
(foreground only, managed DSH_* env, timeout/signal/exit markers, terminal
and generic presenters). Both packages carry full suites (real pwsh,
self-skipping without it) at per-file 100% coverage; vitest's Windows
exclusion narrows from packages/bash/* to the bash-requiring packages so the
pwsh suites run natively on Windows too. The CLI gains the workspace deps
and tsconfig projects without mounting either plugin; the Windows-default
roadmap is recorded as a proposed Agent Note.
Scope decision: previews for a browser that is not on the Host machine are
not supported. With that settled, host.openPath answers the supported case
completely — a file:// document in a real browser has full page capabilities
and no reach into /api — and the HTTP serving this branch had built answered
only the unsupported one.
Removed: the /f route and its listener, the workspace-file URL shape,
ApiProxy.workspaceRootOf, ConnectionHandle.fileUrl, and the port published
into the index page.
Kept, and finished:
- the produced-files row a turn ends with, derived from mutation locations;
- the path link now reads as a link at rest, not only on hover — the reported
"I can't open what it made" was this, sitting on a working capability;
- the Host opener prefers the default BROWSER for .html/.htm/.xhtml/.svg, so
a developer who binds .html to an editor still gets a rendered page
(macOS via the LaunchServices https handler, Linux via $BROWSER, every
failure falling back to the default application).
The retired designs and their measurements stay in the Agent Note, including
why same-origin serving was unsafe and why the sandbox that fixed it broke
the pages invisibly.
Serve one file at a time out of a Session's workspace under /f on the web
transport, and point the conversation's existing file-open affordance at it.
Clicking a write/edit/read row's path now opens that file in a browser tab —
including from a LAN client, where the Host's system opener is fenced to
loopback and answered nothing.
- /f/<sessionId>/<segments> in client-connection, behind the same
browser-trust fence as /api; realpath confinement, streamed reads,
GET/HEAD only, nosniff + no-store.
- Script-capable documents carry CSP sandbox: model-authored markup must not
be same-origin with /api, where events.mux is a readable GET stream.
- ApiProxy.workspaceRootOf answers where a Session's files live without
resuming an agent; the client program cannot reach the core services.
- The /f URL shape lives in dsh-host-apiproxy/api so both ends share one
encoding (client bundles may not value-import another plugin).
Holds the session.history response at the browser's network boundary so
the auto-selected open is observable, then asserts the visible frame and
the recorded phase timeline. Registered host-plane like the other
scaffold-booting e2e files (host include + client project exclude).
The two shipped surfaces offered different tools for no recorded reason:
session checkpoints, tool-result pruning, the goal tools, and Ralph were in
`tui.cordis.yml`; `tool-todo` and web search were in `web.cordis.yml`. Neither
offered session search, a string-replacement editor, or a repeat-tool guard,
though none of the three is surface-specific.
Move the rows that are not surface-specific into `base.cordis.yml` and add
those three. Web search moves there too — the TUI decision the change that
made it a Web default deferred. Both surfaces now assemble the same 27 tools.
This adds only. No row is removed from either surface and no existing row's
configuration is edited: executors, sandbox composition, access defaults,
`tools.mode`, and the workflow tool are exactly what they were. Two rows stay
surface-specific: `tmux-context` (no terminal multiplexer in a browser) and
`session-reference` (its index has one writer owner).
Ship `dsh-mcp-client` as a dependency without a row: the plugin mounts one
server per instance and `command` is required, so a default would name a
third-party server and spawn it outside `ctx.bash` on every launch. The CLI
README carries the YAML for mounting one from a personal config.
Review flagged that "equal by construction" rested on an engine behaviour I
had not measured: `scrollbar-gutter: stable` only equalizes the layers if the
engine applies it to `overflow: hidden` the way it does to `overflow-y: auto`.
Measured it on the running app across the three engines Playwright ships,
and the property does not hold up.
engine .input / .backdrop / .mirror wrap width
chromium 776 / 776 / 776 (768 / 768 / 768 with the declaration)
firefox 776 / 776 / 776 (unchanged by it — overlay scrollbar)
WebKit 768 / 776 / 776 (unchanged by it)
WebKit reserves for `overflow-y: auto` and not for `overflow: hidden`, so the
declaration left .input at 768 against 776 — exactly the gap it was meant to
close — on the one engine where that gap is observable at all, while costing
every chromium user 8px of text column unconditionally. Reverted: the
composer's metrics are now the same as before this PR.
The WebKit gap predates this change and is not closed here. It is recorded in
the Agent Note with the numbers, and the browser scenario asserts the equality
on the lane's engine so a regression into that state fails loudly. The mirror
is unaffected on WebKit for the drafts measured — the extents still agree — but
a draft whose wrapping turns on those 8px would clamp it.
The review's monotonicity concern resolves the same way: the declaration was
never worse than master, because WebKit already measured 768 against 776
without it. It simply was not better.
Also from this round: the wrap-width assertion now covers .mirror as well as
the two glyph layers — it is the height authority, so a mirror alone wrapping
wider would measure the box short and clip content below the 14-line cap with
every other assertion green. Plus `renderGeometry`'s missing `@param
trailingNewline`, and both e2e tsconfig lists restored to alphabetical order.
A composer draft past the 14-line cap could not be scrolled: the caret and
the selection moved, but the words stayed frozen at line 1, so the tail of
anything longer than the cap was unreachable while writing it.
The composer paints its text in two stacked layers. The textarea owns the
value, the selection and the caret but renders its own glyphs transparent;
every visible character is painted by the decoration backdrop beneath it,
which also carries the claim-token highlight, the chips and the ghost hint.
The backdrop is `inset: 0; overflow: hidden` — clipped, not scrolled — and
nothing linked its offset to the textarea's. Below the cap both layers rest
at 0, which is why the defect hid behind every short-draft screenshot and
fixture.
InputBar now mirrors the textarea's scrollTop onto the backdrop, from a
`scroll` listener (every gesture and every caret-driven scroll) and from a
layout effect keyed on the committed draft (an edit reflows both layers
without necessarily firing a scroll event).
Scrolling is layout, so jsdom cannot show this: the unit spec stubs both
offsets and proves the mirroring paths run, while a new browser scenario
measures the user-visible fact against the built client with a DOM Range
over the backdrop's own text — after a wheel gesture over a 40-line draft
the last line is on screen and the first has scrolled out. Confirmed both
directions: with the mirroring reverted and the packages rebuilt, the
golden reads `last draft line is on screen: false` while `textarea moved:
true`.
ZiyaZhang