Boot the glob sampling scenario from a standalone ACP composition that exposes only bash, glob, and grep. Regenerate the smaller header fixtures and trim implementation narration already owned by the Agent Note.
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.
The review made every name a JSON string with `<`, `>`, and `&` escaped. The
hazards behind that are real and each is now covered: a control character
splits one entry across lines, `</` closes the envelope, and a regular file
named `x@` reads as a socket named `x` under the non-regular marker.
Quote those, and only those. `list` is the tool an agent reaches for first
and its output sits in every transcript, so `"archive"/` on every ordinary
line is a permanent cost for a case that almost never occurs. A name is now
emitted verbatim unless it matches a control character, a leading quote, a
backslash, `</`, or a trailing `@`, and is otherwise a JSON string with `</`
neutralized — the delimiter treatment `dsh-workspace-context` already applies
to instruction text, extended to an interpolated path as its
`instruction-frame-paths` TODO asks.
Agent Note for the two changes: what the session log showed, what ordering
can and cannot fix (measured, correcting the first diagnosis), why the page
is sampled only past the cap, why `list` is needed alongside it, and the
alternatives each one beat.
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.
pi-ai is the library-backed twin adapter the tui-agent README already
points at ("swap one line to @deepseek-ai/dsh-llm-pi-ai"), and the
supported entry point for third-party providers (Anthropic, Google,
OpenRouter) mounted through the personal overlay under ~/.dsh. Making
it a declared workspace dep of the examples umbrella means
`pnpm install` resolves the symlink upstream so users configuring a
third-party provider via `~/.dsh/config.yaml` don't have to patch
`examples/package.json` locally (which their next git checkout would
wipe).
Placement matches the sibling llm-* cluster; workspace:* to match the
other adapters. No cordis.yml or README changes: mounting pi-ai
remains explicit and opt-in per the provider-routed-llm-adapters
Agent Note.
Rename the six model-facing tools pty_* -> terminal_* and align every
description, guidance section, ACP card title, and rendered result to
terminal terminology. Package and service internals keep their technical
PTY names (PtyService, "unknown PTY session", node-pty).
Harden the local backend teardown:
- a failed close is retryable: drop the memoized rejection so a later
terminal_close re-runs against the live process table
- service disposal clears the backend, reservation, and owner-cleanup
registries even when a close fails
- stop readiness polling before teardown so an in-flight send settles as
session_exit instead of a mis-inferred wait reason
- bound the sanitizer's pending buffer against unterminated escape runs
Update the tool catalog, package READMEs, the bilingual Agent Note, and the
acp/headless pty-tools snapshots to match.
Auto placement floated the new docks to the top of the main pane; every
dock now owns an explicit grid row (and the layout regression guards
them). Live tool rows use the streamed kind verb like persisted rows,
and the approval card's fallback reject is omitted when the request
already carries a reject option.
Prompts submitted while a turn runs queue above the composer (edit,
remove, auto-send on completion; interruption or a failure pauses the
queue with an explicit resume). Permission requests and elicitations
route from the ACP client in main to renderer cards with the request
options — pending cards replay when the renderer reloads, replacing
the blocking OS dialogs. The task plan docks above the composer instead
of scrolling inside the thread, multi-line user bubbles trim their
padding lines, and a turn that ends abnormally shows an inline error
line in the transcript.
todo/write events fold into plan targets: the chat transcript and
trajectory show a checklist card (done counts, active item), and live
ACP plan updates render the same card while the turn streams — the
task list was previously invisible on both paths. Tool rows label
themselves with the streamed ACP kind verb (read/edit/search/run…)
instead of a generic noun, and touched-file locations open in the OS
editor from the expanded row.
Turn completion no longer force-scrolls the conversation: only a
session switch snaps to the bottom, so finishing a turn does not yank
the view while the user reads. ACP streams richer tool titles (for
example a workflow run's description) than the persisted tool/call
name; remember them per call id and keep rendering them in chat and
trajectory after the persisted trace replaces the live turn.
Develop stays read-first: each artifact whose source is a repository
path gains an Open in editor action routed through a path-validated
main-process handler (shell.openPath), instead of an in-app write path.
Persona and prompt-assembly sources open the governing cordis.yml or
source file directly.
Every sidebar section now shares the same horizontal inset, so the
primary action, module nav, search field, and session rows start and
end on identical lines with one control height and one radius. Topbar
controls share one row height, and the round send-button style no
longer leaks onto the cancel button.
fixing-accessibility: expandable rows expose aria-expanded/controls, user
messages become keyboard-operable buttons, pane dividers and the composer
gain accessible names, the live status row announces politely, the busy
form sets aria-busy, feedback failures render inline next to the form,
and the muted/faint text tiers rise to AA-viable contrast.
fixing-motion-performance: composer autosize moves to CSS field-sizing
(dropping the per-keystroke measure/write cycle) and streaming scroll
writes coalesce per frame while direct user actions stay synchronous.
baseline-ui: data columns use tabular numerals, headings balance and
prose wraps pretty, and z-index joins the token scale.
improve-ui: straggler radii and control heights land on the declared
size scale (radius-xs joins the ramp).
fixing-metadata: the document language follows the active locale.
A finished turn whose persisted JSONL had not flushed yet left the live
turn on screen; the next prompt then patched its content into the stale
skeleton, so the new message rendered under the previous user bubble.
Key each live turn and rebuild the skeleton on mismatch, and poll the
persisted trace after a lagging read so the conversation converges
without waiting for a user action. Module nav rows also collapse to one
line at the primary-action control height.
Replace the full-innerHTML render loop with a static shell plus
per-region updates, so composer drafts, fold state, focus, and scroll
survive streaming turns. Fold session events into one trace graph
consumed by Chat, Trajectory, Waterfall, and the shared inspector
drawer, with live ACP updates patched into a keyed live-turn region.
Align the visual system with a tokenized design spec: a 4px spacing
base with fixed control/row height steps, foreground-derived text
tiers and borders (color-mix), neutral interaction overlays, tiered
motion durations with a reduced-motion collapse, hover-revealed
scrollbars, and drawer-aware layout elasticity. Localize trajectory
role chips and row previews.
The renderer entry (app.ts) joins the coverage exclude list as a
self-executing DOM bootstrap: jsdom lifecycle specs exercise its
behavior, and extractable logic lives in covered modules
(trace-graph.ts, renderer-content.ts).
POST /api/sessions/:id/reveal resolves the file server-side (execFile,
argv array — no shell) and opens the platform file manager; the header
button shows the full path on hover.
The parent session log has no record of subagent spawns (the workflow
engine spawns them internally; the only durable linkage is the child
header's parentSession field), so the workbench reconstructs call ->
children attribution by creation-time window. Surface that panel in the
chat tool block too, not just the trajectory row.
A localhost viewer over persisted session JSONL for agent developers and
researchers. Three first-class views — Chat (markdown-rendered surface
conversation with a second-level inspector), Trajectory (turn/step tree
with scroll-spy over a step-grouped event table, inline annotations),
and Waterfall (timing summary + aligned time track that deep-links into
Trajectory). Subagent sessions group under their parentSession with
spawn links and breadcrumbs; failed tool calls are marked in every view;
?session=&view=&sel= makes any selection a shareable deep link.
Motion follows an audited restraint baseline (plans/ documents the
audit): one strong ease-out token, enter-only animations, press feedback
on pushbuttons, prefers-reduced-motion support.
AGENTS.md gains the research/ layout entry; the AGENTS.md word ceiling
rises 1370 -> 1375 to fit it (one line, relocation not applicable for a
top-level layout entry).
Introduce packages/loop as a capability seam covering goal-based (and
naturally proactive) loops around the existing agent-loop, deferring
time-based scheduling to a future dsh-schedule RFC. Four cordis service
seams — loop-as-session, pluggable Evaluator/Budget with EvaluatorSpec
tiers and protectedPaths, RoundHandoff, GoalReflector — plus a driver-
enforced Default-FAIL contract and reuse of packages/fs policy gate for
reward-hacking defense. Bilingual pair.
Replaces the 20-line stub with a bilingual README pair covering:
- getting started (`pnpm run demo:repl` + Node ^22.19 || ≥24 + pnpm 11.7)
- editor integration (Zed / VS Code / Cursor over ACP)
- programmatic embedding via `@deepseek-ai/dsh-app-boot` (with the
two failure paths of `boot()` vs `installFailLoud` spelled out)
- writing a plugin (`echo-tool` grounded in `examples/echo-agent`, with
the runnable command `pnpm run demo:echo`)
- Packages table grouped by family, including `session-query`,
`context`, and Support (with `invariants` called out as a runtime
diagnostic mounted by `dsh-agent-core`, not test-only)
- deep-dive links, community, license
Bilingual pairing: English and Chinese sides share byte-identical code
blocks and mirroring link ordinals; language-asymmetric community
channels (Discord/X in EN, WeCom in ZH) sit inside HTML `<a href>`
tags that the pairing gate's structural signature excludes by design.
ZH side follows `docs/i18n/terminology.md` — first `agent` occurrence
annotated as `agent（智能体）`, plain `agent` thereafter.
Ships an in-repo skill at `.agents/skills/dsh-arch-diagram/` that owns
the two architecture PNGs (`assets/arch-{en,zh}.png`). HTML templates
+ shared CSS + a `render.sh` that renders via Chrome headless. The
renderer is portable (auto-detects Chrome/Chromium on macOS + Linux,
uses `--no-sandbox` under root for container envs, discovers the local
http.server port from Python's own startup line so `lsof` isn't
required). The 12-card capability row uses `flex: 0 0 auto` so
adjacent card backgrounds cannot clip descriptions; layout tuned to
1560px page / 96% row-width / 1820×580 render viewport.
Consolidates the four review-response commits from the initial round
of ds-review-bot feedback; each review round is documented in a
top-level PR comment for traceability.
Entry point for the skill's operator and for any future handoff: what the
private periodic tool produces, where the saved candidate diff and SKILL.md
land on the operator's machine, and how the promote helper turns the
operator's decision into a draft PR from a clean master checkout. Companion
to the maintenance RFC.
A private out-of-repo maintainer tool turns adopted human review feedback
into candidate skill updates. Only the resulting SKILL.md diff enters this
repository through normal PR review; the tool source, adapter binaries,
provider credentials, and scheduler stay on the maintainer's machine to
keep this repository focused on shipped product code. Kept in proposed/
until end-to-end verification against master lands a candidate PR; the
acceptance criteria list the exact observations that unlock promotion.
NI0317