[zy1233]

Summary

Two changes to the chat of the VS Code extension. Three commits:

Commit	Change
084001e	Markdown is drawn as Markdown (markdown-it as a parser only)
4376f93	A streamed answer is drawn at most once each 80 ms (perf; can be dropped alone)
48a6e09	The status row shows context use, turn time and folder (added in this update)

1. Markdown in the chat (commits 1 and 2)

The chat showed headings, tables and bold text as raw marks (##, |---|, **). The reader of 1.0.0 made only paragraphs, flat lists, code and links. Now markdown-it is a parser only: the webview still builds DOM nodes from plain data, and it never sets innerHTML (design rule 5).

agentText (host -> webview)
  setAgentText
    throttle: 80 ms, first piece at once, last piece always      (commit 2)
    drawText
      parseMarkdown(text)                    core/markdown.ts
        markdown-it + markdown-it-cjk-friendly: tokens only, html: false
        tokens -> Block[] / Inline[]         plain data, no HTML string
      renderBlocks(blocks)                   webview/dom.ts
        el() + textContent                   DOM nodes

 packages/pi-vscode/
   src/core/
-    markdown.ts          hand-written reader (paragraph, flat list, code, link)
+    markdown.ts          markdown-it tokens -> plain data
+    throttle.ts          draw throttle (commit 2)
   src/webview/
     dom.ts               Block[] -> DOM; now also headings, quote, rule, table, em, strong, del
     main.ts              (commit 2) draws a streamed answer at most each 80 ms
   media/chat.css         h1-h6, blockquote, hr, table
   scripts/notices.mjs    a bundled package inside another one (entities) is listed now

What does not change (rule 5, each one has a test):

• Raw HTML stays text (html: false).
• An image is never loaded. It shows as a link.
• A link opens through the host: http, https and mailto only. Another address shows as plain text with its address.
• Bare main.py, www.x.com and e-mail words are not linked. Only http(s):// and mailto: are.

2. Status row (commit 3)

The header of the chat and the VS Code status bar item show more. The rule: show only a number that someone measured (the agent, or the clock of the host). The extension counts no tokens and has no setting for the size of the window.

agent --usage_update {used, size, cost?}--> ChatController            core/chat-controller.ts
                                              readUsageUpdate()       core/status-row.ts  (finite numbers only)
send() / finishTurn() -- host clock ------> turn { elapsedMs | lastMs }
                                              stateMessage { usage?, turn?, cwd, ... }
  webview   renderHeader()                                            webview/main.ts
              ● zycode  Working · 1 min 12 s                           (the page ticks one element each 500 ms)
              agent 0.9.1 · deepseek-flash · folder pi-python          (full path in the tooltip)
              Context [bar] 42% · 54.2k of 128k tokens · $0.12         (only after a usage_update)
  status bar  StatusBar                                               ui/status-bar.ts
              zycode: Working · context 42%                            (tooltip: numbers, cost, last turn, folder)

Item	Source	With the bundled agent 0.9.1
Turn time (a clock, then last turn 8.4 s)	clock of the host	shows
Folder (name in the meta row, full path as tooltip)	cwd of the chat	shows
Context (bar, percent, tokens, cost)	ACP usage_update	does not show: 0.9.1 sends no usage_update

The bar turns to a warning colour from 80 % and to a danger colour from 90 %. A value over 100 % shows its real percent, and the bar stops at full.

 packages/pi-vscode/
   src/core/
+    status-row.ts        pure helpers: readUsageUpdate, describeContext, formatTokens/Cost/Duration
     chat-controller.ts   usage_update -> state.usage; host clock -> state.turn
     protocol.ts          StateMessage { usage?, turn? }
   src/webview/main.ts    clock, Context row (meter), folder in the meta row
   src/ui/status-bar.ts   `· context NN%` and the tooltip lines
   media/chat.css         .clock, .usage (meter, warn, danger); the title row wraps in a narrow view
+  test/status-row.test.ts, test/usage.test.ts
 docs/specs/
+  2026-11-13-agent-context-usage.md      investigation record (no agent change)
   2026-11-04-vscode-extension-design.md  section 7d, decisions D36 and D37

The agent is not changed. The request was to record the result only. docs/specs/2026-11-13-agent-context-usage.md has it: the agent has no usage_update; the window size is a 128,000 placeholder for every model (Model.context_window, no config key); Usage.cost stays 0; the harness already counts the tokens (estimate_context_tokens). The record also has a proposal (four options for the window size, a hook on turn_end, the release path) and the open decisions.

Evidence

1. Markdown

Before: the test of 1.0.0 said that these stay text.

parseMarkdown("# title\n**bold** <script>alert(1)</script>")
  -> one paragraph, text "# title\n**bold** <script>alert(1)</script>"

After: the same kind of input in headless Chrome (renders headings, emphasis, lists, a quote, a rule and a table as elements, with no raw marks left). The HTML part still stays text.

"## Title"            -> <h2>Title</h2>
"**b** ~~c~~"         -> <strong>b</strong> <del>c</del>
"| a | b |\n|:-|-:|"  -> div.table-wrap > table  (th.align-left, th.align-right)
"**重要：**文本"       -> <strong>重要：</strong>文本

The last line needs the plugin: plain markdown-it leaves ** raw in 3 of 10 common Chinese patterns that we tried.

How many answers need this? In 47 session files of the author (131 answer texts, a small sample), 41 answers (31 %) had a heading, a table, bold text, a quote, a rule or a nested list.

	1.0.0	commits 1 and 2
dist/webview.js (minified)	11.1 KB	118.0 KB
dist/extension.js	-	no Markdown code in it (0 inputs in the esbuild metafile)
one draw, text 3,000 characters long	5 ms	7 ms
one draw, text 8,000 characters long	12 ms	16 ms
one draw, text 20,000 characters long	26 ms	37 ms (the reason for commit 2)

How I measured a draw: headless Chrome, both readers built in the same way (minified). The text is already N characters long. I add a piece of 4 characters and time the draw. The number is the median of 25 pieces, and the table shows the mean of two runs. The new reader is about 1.4 times slower for each draw. The reader of 1.0.0 also needs more than 20 ms at 20,000 characters. If a model sends 50 pieces each second (an assumption: I did not measure a real model), there are 20 ms between two pieces. Commit 2 limits the draws to 12 each second.

Other checks of commits 1 and 2:

• The Chrome tests also pass on the minified production bundle (run with a temporary config without the global dev build).
• Mutation check: without entry.draw.cancel() in resetTranscript, the test does not draw an old answer after the chat was reset fails.
• Hostile input: 12 slow or deep strings parse in less than 80 ms each. 3000 random mixes of marks (fixed seed) always give a valid shape (a table row has as many cells as its head).

2. Status row

Before (1.0.0 header, and this branch before commit 3): the state and the meta line only.

● zycode  Ready
agent 0.9.1 · deepseek-flash

After, with the bundled agent 0.9.1 (no usage_update):

● zycode  Ready · last turn 8.4 s
agent 0.9.1 · deepseek-flash · folder pi-python

After, with an agent that sends usage_update (a test message; see "Not verified"):

● zycode  Working · 1 min 12 s
agent 0.9.1 · deepseek-flash · folder pi-python
Context [####------] 42% · 54.2k of 128k tokens · $0.12

I looked at both in headless Chrome at a 380 px wide view (an iframe, because headless Chrome has a minimum window width). The longest case (the Always approve label, a clock and a danger bar at 96 %) fits: only the label moves to its own line, and nothing is cut off.

The tests (headless Chrome, status row; unit; controller with a fake agent):

state({ usage: { used: 54200, size: 128000, cost: { amount: 0.12, currency: "USD" } } })
  -> row "usage ok", text "42% · 54.2k of 128k tokens", cost "$0.12"
     meter min 0, max 128000, value 54200, aria-label "Context used"
usage of 79k, 80k, 89k, 90k, 112.5k (of 100k), and no size
  -> ok, warn, warn, danger, danger, plain
     over full: "113% · 113k of 100k tokens", bar value 100000
cost currency "<img src=x onerror=alert(1)>", cwd "/home/u/<b>x</b>"
  -> 0 <img>, 0 <b>                                  (text, never HTML)
busy, turn { elapsedMs: 5000 }, 2.2 s later
  -> "5 s", then "7 s"; the same <select> element     (an open menu stays open)
turn ends with { lastMs: 8432 } -> "last turn 8.4 s"; next state without turn -> no clock
controller: a usage_update with no `used` -> the SDK drops it, the numbers stay, the turn ends
controller: new chat, restart and stop clear the usage and the clocks; a clock that goes back gives 0

Tests at the head (all three commits)

• Windows, Node 24: npm run check = typecheck + eslint + 333 tests passed (255 before commit 3; +62 status-row, +11 usage, +5 Chrome). 18 are skipped (16 opt-in tests of the real agent, 2 for another OS).
• npm run notices (production build) passes: dist/webview.js 120.8 KB, dist/extension.js 216.0 KB. The list of third-party packages did not change.
• CI at 48a6e09 (Linux, Node 22.23.4): all checks pass. 335 tests passed and 16 skipped. The webview file ran its 26 tests in 24.6 s, with no unhandled error. The VSIX job built dist/webview.js at 120.77 KB. The smoke test started the agent from the VSIX. Ruff and pytest (3.11, 3.12, 3.13) pass; no Python file changed.

Not verified

• A real VS Code window and real themes. I used a headless Chrome page with the theme variables mocked. The status bar item (ui/status-bar.ts) needs a VS Code host, so no test draws it; its two helpers (contextBadge, statusRowLines) are tested.
• A real usage_update from a real agent. No agent sends one today. I used a fake agent and test messages. The shape is the one of ACP v1 (used and size required, cost optional); the SDK checks it with its own schema.
• The Context row with true numbers. It needs the agent change in the record. The size of the window is the open question there.

Merge Danger

Door: two-way. A revert brings back the old reader and the old header. There is no change of the wire, of the agent, of the pins (agent-assets.json) or of stored data. The extension only reads a standard message that an agent may send.

Blast Radius: the look of every answer in the chat, and the header (a clock and a folder in the meta row; the title row can wrap in a narrow view). The webview bundle grows by about 110 KB (11.1 KB in 1.0.0, now 120.8 KB). THIRD-PARTY-NOTICES lists 8 more packages (MIT, and BSD-2-Clause for entities). Users get it with the next VSIX release. This PR does not make one: the version stays 1.0.0 and the CHANGELOG has an Unreleased entry.

Note for the user of the bundled agent: the Context row stays hidden until an agent release sends usage_update. Then the extension shows it with no change.