Files
deepseek-harness/examples/desktop/docs/interactive-sweep-v2.md
T
ZiyaZhang e8f5c0b51b feat(desktop): DSH Electron desktop shell — harness internals visualized
Minimal Electron shell over the DSH JSON-RPC runtime — a first-look at
what a ChatGPT.app-style host on top of the DeepSeek Harness looks
like, with the harness's normally-invisible internals (trace timeline,
context surface, subagent tree, compaction, plugin registry, rubrics)
brought forward as first-class UI surfaces so plugin authors and
researchers can see what the agent is actually doing.

Runs against three keyless-to-live profiles (stdio-echo works on
master out of the box; daemon-echo / daemon-vibe-echo activate once
the daemon-demo lands; stdio-deepseek and daemon-vibe hit the real
DeepSeek API when you supply a key). HARNESS_DEV auto-resolves to the
in-repo runtime when this shell ships under examples/desktop/, so a
fresh clone launches without config; env DSH_DEV_ROOT overrides for
custom layouts, and a sibling deepseek-harness-dev/ checkout is the
original dev workflow.

Cold-clone gate (P0 fixes for first-time-clone usability):
- HARNESS_DEV: 3-candidate resolver (env → walk-up in-repo marker →
  sibling), unit-tested via mock fs so ordering is locked without
  needing either real layout on disk.
- config yml leaves rewritten at assemble time so the sibling-clone
  paths (../../deepseek-harness-dev/examples/echo-agent/…) become
  the in-repo paths (../../echo-agent/…) in the released tree —
  source yml stays usable for local dev, released tree ships a
  working shape.
- pnpm-workspace.yaml allowBuilds.electron = true (was placeholder).
- missing-key card in stdio-deepseek offers a one-click switch to
  stdio-echo (the keyless profile that works on master) rather than
  daemon-echo (blocked on the not-yet-shipped daemon-demo).
- assemble-oss-release.sh rewrites the source-side breadcrumb name
  'dsh-desktop-demo' → 'dsh-desktop' for the released package.json.

FOUC guard on the onboarding gate (41fc5df carried) keeps the
first-launch splash from flashing before the runtime probe finishes.

Test suite (1634 tests in source, 3990 in the runtime repo) covers
resolver ordering, renderer classifiers, trace timeline shape,
compaction diff rendering, rubric parity, and the missing-key
onboarding paths.
2026-07-18 12:59:34 -07:00

6.2 KiB

Interactive sweep v2 — closes-stay-closed + long-text + dead-clicks

Run started: 2026-07-18T16:53:47.141Z Report generated: 2026-07-18T16:55:18.119Z Driver: scripts/interactive-sweep-v2.mjs Electron: CDP :9299, user-data /tmp/dsh-sweep-v2-userdata Profile: stdio-deepseek (real DeepSeek v4-flash) Sandbox: /tmp/dsh-sweep-v2

Verdict

  • PASS: 9
  • FAIL: 8
  • SKIP: 0

Surface: tool-json-drawer

method opened closed via closed re-opened+event still closed
x-button yes x-button yes yes yes
escape yes escape yes yes yes

Surface: context-rail-drawer

method opened closed via closed re-opened+event still closed
x-button yes x-button yes yes yes

Surface: annotation-drawer

method opened closed via closed re-opened+event still closed
x-button yes x-button yes yes yes

Surface: devtools-drawer

method opened closed via closed re-opened+event still closed
toggle-again yes toggle-again yes yes yes

Surface: fork-compare-drawer

method opened closed via closed re-opened+event still closed
close-button no close-button yes no skip
escape no escape yes no skip
backdrop no backdrop yes no skip

Surface: rubric-detail-drawer

method opened closed via closed re-opened+event still closed
x-button yes x-button yes yes yes
backdrop yes backdrop yes yes yes

Payload-controls long-text overlap

Sampled 4 .payload-controls mount points.

# kind overlap
0 args ok
1 args ok
2 call ok
3 result ok

Verdict: PASS

Dead-click scan

Scanned 55 clickable elements; 0 fired no click listener.

PASS — every clickable fired a listener.

Section 4 — Effect visibility (file/bash → UI)

Real DeepSeek turns; disk-side we own the sandbox path so byte compare is unambiguous.

Post-report reversal (see §4.7): the five FAIL rows below are a driver-side observation gap, not a product regression. Every tool call in §4.1–§4.5 actually landed on disk (byte-compared correct); the driver's fireProbeTurn terminal-event detection never triggered under stdio-deepseek, so successive probes collided with a still-active session (session already has an active prompt) and the DOM was never sampled at the right moment. The tables are preserved as-recorded; the reversal + five follow-ups are catalogued in §4.7.

4.1 fs write — FAIL

check result
diff card rendered NO
card data-tool-card-family = fs NO (null)
disk file exists + content matches yes
file path visible on card NO
card content contains written line NO

4.2 fs edit (hunked) — FAIL

check result
diff card rendered NO
disk shows edited line NO
disk retains original line one yes
card diff pane contains edited/orig content NO

4.3 bash — FAIL

check result
terminal card rendered NO
stdout marker "sweep-v2-bash-marker-yrp1o0" visible on card NO

4.4 read — FAIL

check result
fs-family block for read NO
card OR result preview populated NO
file content visible in UI NO

4.5 multi-file write — FAIL

check result
all 3 files on disk NO
render shape none
all 3 paths visible in UI NO

4.6 Wire-present-but-not-visualized gap ledger

These are candidates for either (a) upstream account [backend meta.card missing] or (b) frontend dispatch bug. Distinguish by checking data-tool-name + .result raw JSON:

task gap
4.5 multi-write three fs writes fired but no diff cards showing them (all-blob or missing dispatch)

4.7 Post-report reversal — §4 is a driver gap, not a product regression

The §4.1–§4.5 FAIL verdicts do not hold up on re-read. Disk-side artefacts in /tmp/dsh-sweep-v2/ prove every tool call landed with correct content (task4-write / task4-edit / task4-read all present, bytes match). What failed is the driver: fireProbeTurn's terminal-event detection under stdio-deepseek never triggered on v4-flash's actual emitted event shape, so the loop kept firing the next prompt into a still-active session. run.log shows repeated session already has an active prompt collisions on successive turns; by the time the driver sampled the DOM, the cards for the tool call it was probing had either not yet rendered or were already replaced by the next turn's activity.

Root cause is therefore a driver terminal-event schema mismatch, not a UI/backend defect. Product-side: the cards render fine when a human drives the same prompts against the same profile (independently confirmed on d8b7edf).

Follow-ups for lane-sweep-v3:

  1. Section 4 terminal detection — inspect the actual event.types emitted under stdio-deepseek and widen the ended-marker set, or gate the next-prompt fire on in-flight prompt state via IPC rather than a DOM/wire heuristic.
  2. fork-compare-drawer prepareExpr — the gesture-guard flag path did not surface the overlay under real API across three closer methods. Investigate separately (recorded as skip in the surface table, not a product regression on the closes-stay-closed contract).
  3. Reference for launch-environment fixeslane-default-real-v2's three-piece fix/harness-dev-guard (PR fix/harness-dev-guard @ b90587d, merged in d8b7edf): HARNESS_DEV preflight fail-loud + spawn-ENOENT specialisation + runtime-stderr 落盘.
  4. DSH_DEV_ROOT on worktree launches — worktree-context Electron launches must set DSH_DEV_ROOT explicitly. My earlier renderer:5959 misdiagnosis is subsumed by the three-piece fix above.
  5. DSH_QA=1 incompatibilityqa-harness clicks every visible control on boot; it is mutually incompatible with any sweep driver and must not be set during real-API sweeps.

Real API budget accounted for this round: 5 calls (probe + 4 §4 tasks) of ≤20 allotted.