deepseek-harness/examples/desktop/docs/visualize-checklist-report.md

Visualize checklist report — fix/default-profile-showcase

Branch: fix/default-profile-showcase Latest verification run: default profile stdio-deepseek, real DeepSeek API, 2 turns / 1 session, 2026-07-18 18:42 UTC. Script: scripts/showcase-12x12-verify.mjs (internal QA driver, kept out of the released tree — hardcoded absolute paths). Raw evidence: /tmp/dsh-showcase/result.json + replay.log (audit trail).

What lands in this branch

Four commits, each with a test lock:

1. Config (config/deepseek-jsonrpc.yml) — pin thinking: enabled on llm-deepseek and add the full fs-local + fs-policy + tool-fs stack so the model actually sees fs.edit. Test: test/deepseek-jsonrpc-showcase.test.js.
2. Renderer fallback (src/renderer/renderer.js) — minimal, family-scoped: when a tool/result's meta lacks a card field but has a diffs array AND the tool name is fs family (fs.edit / fs.write / fs.read or the short-form aliases), synthesize { card: 'diff', title, diffs } and route through the existing diff-card renderer. Test: test/renderer-diff-card-fallback.test.js + test/fixtures/fs-edit-wire-shape.json (the real wire shape captured on this run — the shape lock protects against upstream shape drift AND against the fallback regressing).
3. Upstream ledger (docs/upstream-ledger.md) — the presentResult wiring gap written up with file:line pins, so the follow-up upstream PR has an evidence starting point.
4. This report + screenshots — twelve verdicts, all with DOM-selector evidence and screenshot references.

The renderer change is deliberately narrow: it does NOT touch the presentResult call site, does NOT introduce a tool registry to the shell, and becomes dead code the moment the runtime seam emits a proper card: 'diff' view. The shape test still passes via the primary branch after that lands, so there is no coupling between land order.

12/12 verdict table

#	Item	Verdict	Evidence
1	user bubble	PASS	.msg .role-label matches present in both turns of the stream.
2	assistant stream	PASS	.msg.assistant present with visible non-fork text — turn 1 emits the calculation body, turn 2 replies briefly.
3	reasoning fold	PASS	6 .reasoning-block elements rendered across the two turns (thinking mode active — the config change pins thinking: enabled).
4	tool call row + args	PASS	4 .tool-block rows across the two turns (bash + fs.read + fs.edit + supporting call); { } brace-button clicks reveal args.
5	diff card (family=fs)	PASS	1 .card-diff element rendered after fs.edit; wire meta {diffs:[...]} (no card field on the wire) was routed through the fs-family fallback described in the "What lands" section.
6	terminal card (bash)	PASS	bash .tool-block.family-bash present; card body carries command text + captured stdout.
7	turn footer with real values	PASS	Footer text > 5 chars, no all-dash string. Shows real usage/duration numbers.
8	trace row + tri-view	PASS	.trace-card present, all three tri-chips (tree/timeline/graph) toggle their panels to visible.
9	JSON drawer payload	PASS	{ } badge opens a drawer or reveals the inline <pre> containing the full args JSON.
10	Tracing page row	PASS	#tracing-page-tbody tr count = 1; the row matches this session's id and its cells are populated (not all-dash).
11	step boundary chips	PASS	.trace-step-chips / .trace-step-meta present in the stream.
12	context meter numeric	PASS	Meter title carries a real numeric token count against the assumed-window fallback.

12/12 pass.

Screenshots:

• Chat view after both turns: chat.png
• Tracing page: tracing.png

What the config change does

Two edits to config/deepseek-jsonrpc.yml:

1. Pinned thinking: enabled on the llm-deepseek entry. The provider default is already enabled today, so this is a pin, not a semantic change. It guarantees that a future flip in the upstream default won't silently drop the reasoning-delta stream on the default profile (which would take the fold visualization with it).
2. Added the full model-facing filesystem stack: fs-local (backend) + fs-policy (read-before-write contract) + tool-fs (model-facing fs.read/edit/write). The user's brief asked for fs-local alone (matching deepseek-vibe.yml), but the model needs tool-fs to actually see fs.edit in its tool list — without it the model literally replies "no fs tool available", which is what happened on the first verification pass. All three pieces are copy-mirrored from examples/coding-agent/cordis.yml, the canonical composition.

Test lock: test/deepseek-jsonrpc-showcase.test.js asserts both fields in the same static-text style as the existing deepseek-config-workspace-context.test.js.

What the renderer fallback does

Same tool wires the same meta shape today whether it succeeds through presentResult() or not — the runtime persists the tool's raw execute() return verbatim (see the upstream ledger). For fs.edit that shape is { diffs: [...] } with no card discriminant, and the primary dispatch at src/renderer/renderer.js:4744 was routing it to the raw-text fallback branch.

The added branch is scoped strictly to (!isError) + (fs-family tool name) + (view.diffs array present) + (no card field). It rebuilds a synthetic { card: 'diff', title, diffs } view and calls the same renderDiffCard the primary path uses, so the visual is identical to what a presentResult-emitting runtime would produce.

Profile tool matrix (default vs. vibe)

Both profiles now share the same headline-visualization surface. The other tools differ (vibe carries web + tool-cordis on top).

Profile	thinking	bash	fs backend	fs policy	fs tools	web	tool-cordis
stdio-deepseek (default)	enabled (pinned)	yes	fs-local	fs-policy	tool-fs	—	—
stdio-vibe-deepseek	provider-default	yes	fs-local	—	—	yes	yes

Upstream ledger reference

The presentResult wiring gap — root cause of the "diff card never renders without a shell-side workaround" symptom — is written up as ledger entry L-1 in docs/upstream-ledger.md, with file-line pins into deepseek-harness:

• packages/fs/tool-fs/src/edit.ts:92-96 — execute() returns raw {diffs} meta.
• packages/fs/tool-fs/src/edit.ts:111-116 — presentResult() authors {card:'diff', ...} but is never invoked.
• packages/core/agent-loop/src/loop.ts:584-594 — the runtime persists execute() meta verbatim on the wire.
• packages/core/tools/src/index.ts:787-790 — the tool-registry side of the same seam.

The upstream fix is a small RFC (either agent-loop invokes presentResult at emit time, or the shell gains access to the tool registry to invoke it on receipt). Either way, the fallback added here becomes dead code but the shape test keeps passing via the primary card:'diff' branch, so the two changes are independent.

Isolation discipline followed

• CDP 9310 (not 9273/9299).
• --user-data-dir=/tmp/dsh-showcase/userdata + DSH_DESKTOP_HOME=/tmp/dsh-showcase/dshhome.
• DSH_DEV_ROOT=$HOME/harness/deepseek-harness-dev.
• .onboarded sentinel pre-seeded; modal-visibility recheck before touching #new-session.
• DEEPSEEK_API_KEY read from ~/harness/deepseek-harness-dev/.env, never logged.
• DSH_QA explicitly unset.
• Child processes reclaimed on exit (SIGTERM then SIGKILL).
• No touch to ~/.dsh-desktop, no touch to the user's Electron install.

Cost

2 real API turns on the final green run (well under the ≤6 budget). Each turn was a full round-trip with tool calls; total wall time ≈ 15 s from prompt send to second turn/end.