app: markdown projection re-lexes the whole accumulated message on every streaming delta #56202

Activity

1. opencode-agent commented on Oct 16, 2026

   

   opencode-agentBot

   on Oct 16, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the detailed report. I reproduced this on the latest dev (bc5ebf00a4c1).

   I streamed synthetic messages through project(previous, text, true) one appended delta at a time, counting the characters passed to marked's lexer:

   • One paragraph, 250 → 2000 words: one full lex per delta. Characters lexed grow 215K → 861K → 3.4M → 13.8M, about 4× each time the length doubles. Time goes from 15 ms to about 8.4 s.
   • Many short paragraphs: 221K → 14.2M characters lexed.
   • Growing list, 250 → 2000 items: 323K → 22.3M characters lexed, 63 ms → 5.1 s.
   • Open code fence: stays linear (a single lex in total), because only that tail grows in place.

   The 2000-word paragraph took far longer than its lexed-character count would suggest, so another superlinear cost may be on that path, possibly the heal/remend step over the tail. I haven't confirmed this.

   The script below reproduces this from packages/session-ui. Your open PR #55857 targets this path. #54305 and #54428 cover the separate store and DOM costs for each delta.

   Reproduction script (lex-count.ts)

   // Reproduces #56202: counts characters fed to marked's lexer while project() streams a message.
   // Run from /workspace/repo/packages/session-ui:  bun /workspace/repro/lex-count.ts
   const { Lexer } = await import(Bun.resolveSync("marked", process.cwd() + "/src/components"))
   let chars = 0, calls = 0
   const orig = Lexer.prototype.lex
   Lexer.prototype.lex = function (src: string) { chars += src.length; calls++; return orig.call(this, src) }
   const { project } = await import(process.cwd() + "/src/components/markdown-stream.ts")
   
   function run(label: string, deltas: string[]) {
     chars = 0; calls = 0
     let p: any, text = ""
     const t = performance.now()
     for (const d of deltas) { text += d; p = project(p, text, true) }
     console.log(`${label.padEnd(34)} len=${String(text.length).padStart(7)} deltas=${String(deltas.length).padStart(5)} lexCalls=${String(calls).padStart(5)} lexChars=${String(chars).padStart(10)} ms=${(performance.now() - t).toFixed(0)}`)
   }
   const word = (i: number) => `word${i % 97} `
   for (const n of [250, 500, 1000, 2000]) {
     run(`paragraph ${n} words`, Array.from({ length: n }, (_, i) => word(i)))
   }
   for (const n of [250, 500, 1000, 2000]) {
     run(`short paragraphs ${n} words`, Array.from({ length: n }, (_, i) => word(i) + (i % 10 === 9 ? "\n\n" : "")))
   }
   for (const n of [250, 500, 1000, 2000]) {
     run(`list ${n} items`, Array.from({ length: n }, (_, i) => `- item ${i}\n`))
   }
   // control: open code fence tail (has a fast path)
   for (const n of [250, 500, 1000, 2000]) {
     run(`open code fence ${n} lines`, ["```ts\n", ...Array.from({ length: n }, (_, i) => `const x${i} = ${i}\n`)])
   }

   Run it: bun lex-count.ts
2. opencode-agent

   added

   reproduced

   on Oct 16, 2026
3. dcerisano commented on Oct 26, 2026

   

   dcerisano

   on Oct 26, 2026

   Last edited by dcerisano

   Author

   More actions

   Fixed by #57016 (merged Oct 20), which replaced the marked + remend re-lex path with @opencode/patchmark. The projection cost was diagnosed and benchmarked here first; the prior implementation is #55857, and the benchmarks and tests that still apply are in #59391.

   For the record, the project's own bot reproduced this on Oct 16 — three days before #57016 — and noted "Your open PR #55857 targets this path."

   Please close this against #57016.
4. Tarquinen commented on Nov 2, 2026

   

   Tarquinen

   on Nov 2, 2026

   Member

   More actions

   Thanks for the report! As you noted, #57016 fixed this, so I'm closing this issue.

   — from 𝕺𝖕𝖊𝖓𝕮𝖔𝖉𝖊
5. Tarquinen

   closed this as completedon Nov 2, 2026