1 Commits

Author	SHA1	Message	Date
Dongsen	12ce5ad38b	fix(web): ignore <artifact> tags inside markdown code spans and fences (#1132) ... * test(web): add failing parser cases for <artifact> recitation in markdown code Cover the three real-world prose contexts where the model legitimately quotes the artifact tag without intending to emit one: - inside an inline backtick span - inside a fenced code block - spread across streaming chunks crossing the fence boundary Establishes the RED baseline before parser code-fence awareness lands. * fix(web): ignore <artifact> tags inside markdown code spans and fences The streaming artifact parser scanned the buffer with a raw indexOf, guarded only by 'next char must be whitespace'. That meant any literal <artifact ...> the model recited while documenting the protocol — even inside backticks or a ```html fence — flipped the parser into artifact mode, swallowed the rest of the reply from the chat UI, and (when a matching </artifact> appeared in the recitation) silently wrote a spurious file to disk via persistArtifact. Replace findOpenTag with a linear scan that tracks fenced code blocks (```) and inline code spans (`), skipping any <artifact prefix found inside either. If the buffer ends mid-fence, return a partial match anchored at the fence start so the next streaming chunk can resolve the boundary without losing fence context. Closes #1130. * fix(web): match renderer fence/inline-code rules in artifact parser Codex review on PR #1132 caught that the previous fix toggled inFence on any triple-backtick run anywhere in the buffer, including mid-line, while the chat renderer (apps/web/src/runtime/markdown.tsx) only treats ``` as a fence when it occupies a whole line matching /^[ ]{0,3}```(\w[\w+-]*)?\s*$/. That asymmetry would suppress a real <artifact> tag emitted after a prose sentence like "the opening marker is ```html and the response then writes:". Rework findOpenTag in three passes that mirror the renderer: 1. Walk \n-terminated lines; only a line that matches FENCE_LINE_RE toggles fence state. Open fences without a close (or with an unterminated tail line) return partial so the next chunk can resolve. 2. Collect inline code spans with /`[^`]+`/g — the same regex used by renderInline — so what the parser skips matches what the user sees as code. Unmatched trailing backticks after the last \n hold back. 3. Find the first <artifact …> outside any skip range; preserve the existing partial-prefix tail handling. Adds a regression test covering the exact case Codex reported. * test(web): pin parser behavior on double-backtick and in-fence string literal recitation Two cases raised in PR #1132 review: - a real artifact tag wrapped in '``<artifact …>``' (double-backtick inline code span) should not be treated as a real artifact - a fenced JS example whose body contains a string literal like 'const fence = "```";' should not pop fence state early and let a later literal <artifact> be parsed as real Both already pass on 96e88ca because the line-anchored fence regex and the renderer-aligned inline regex handle them correctly. Pinning the behavior so future regressions surface as test failures. * fix(web): make stripArtifact markdown-aware to stop truncating literal recitations The streaming artifact parser was hardened in 96e88ca to skip <artifact> recitations inside backticks and fences, but the post-stream stripper at AssistantMessage.tsx still ran a naive 'content.indexOf("<artifact")' over the same text events. As reported by lefarcen on PR #1132, that meant chat replies with literal protocol recitations could still get silently truncated mid-explanation — even though the parser preserved them in the text stream and the file panel was no longer polluted with ghost files. Extract the renderer-aligned classification (FENCE_LINE_RE, INLINE_CODE_RE, computeSkipRanges, rangeContains) into a single source of truth at apps/web/src/artifacts/markdown-context.ts so the parser and the stripper agree on what counts as code. Add apps/web/src/artifacts/strip.ts with a markdown-aware stripArtifact that: - ignores any <artifact open inside a fenced block or inline code span - looks for </artifact> with the same skip-range filter, so a real open paired with a literal close inside backticks does not strip a literal body that is meant to render - returns content unchanged when an open exists with no matching real close (the previous implementation sliced to end-of-string, which would nuke trailing prose on a malformed or still-streaming tag) Refactor parser.ts to import the shared helpers; behavior preserved (all seven existing parser tests still pass). New strip.test.ts covers six cases including the empirically-verified inline-backtick regression. * fix(web): align artifact stripper/parser fence rules with renderer exactly Two gaps surfaced in review at a0bf05f: - markdown-context.ts used a single FENCE_LINE_RE that allowed 0-3 leading spaces and reused the same pattern for opening and closing fences. The chat renderer (runtime/markdown.tsx:44 and :49) is asymmetric — opens with /^```(\w[\w+-]*)?\s*$/, closes with /^```\s*$/, and rejects any leading indentation on either side. Indented " ```html" was being treated as a code fence even though the renderer keeps it as a paragraph, and a literal "```html" line inside an open fenced example was closing the skip range early — both could expose a real or literal <artifact …> to the wrong handler. - stripArtifact discarded computeSkipRanges' unclosedFenceStart, so a fenced literal that ends at EOF without a trailing newline (very common for chat output) leaked the inner <artifact …> recitation to the stripper, reproducing the original #1130 truncation symptom on a narrower input shape. Split FENCE_LINE_RE into FENCE_OPEN_RE / FENCE_CLOSE_RE with no leading indentation, gate the fence state machine on the right side of the toggle, and have stripArtifact extend skip ranges to end-of-content when a fence is left open. Also tightened the parser's tail-line hold-back regex to match the renderer's no-leading-space rule. Added regression tests for the EOF-unclosed-fence case, the indented pseudo-fence (renderer treats as paragraph, stripper must strip the real artifact), and a "```html" line inside an open fence. Refs nexu-io/open-design#1130 * refactor(web): align streaming tail-line fence guard with FENCE_OPEN_RE The streaming parser's tail-line hold-back used a stricter local regex (/^```\w*$/) than the renderer's FENCE_OPEN_RE (/^```(\w[\w+-]*)?\s*$/), missing valid opener tails like ```c++, ```ts-, or ``` (trailing space). In practice these tails are still held back by the unmatched-backtick parity scan that runs immediately after — three backticks in a tail line are odd, so firstUnmatched stays set and the parser holds from that position. So this wasn't a runtime correctness bug, just a regex divergence that future readers could trip on. Drop the local regex and reuse FENCE_OPEN_RE so the tail check matches the same shape the rest of the pipeline already uses. Pinned the behavior with three new parser tests (`+`/`-` info-string suffix and trailing-space tails arriving as the first chunk) — they pass at HEAD, proving the parity scan was already covering these cases. Refs nexu-io/open-design#1132 (lefarcen polish P2) * fix(web): scope inline-code skip ranges per block and reject <artifact prefix-shared opens INLINE_CODE_RE previously ran over the whole buffer, so an unmatched backtick in one paragraph could pair with a backtick in a later paragraph and create a phantom inline span that swallowed any real <artifact …> between them. Mirror runtime/markdown.tsx by splitting the buffer on fence / blank / heading / list / hr boundaries and running INLINE_CODE_RE per block region instead. stripArtifact accepted any unskipped `<artifact` substring as a real open, while the streaming parser already required a following whitespace character — so prose like `<artifactual>demo</artifact>` was being truncated to `prefix suffix`. Extract the parser's real-open guard into isRealArtifactOpenAt and reuse it from both sides. While reordering findOpenTag for the shared guard, also fix the related hold-back ordering issue tracked at #1141: a stray tail-line backtick or fence-opener prefix used to suppress an artifact already complete earlier in the buffer. Scan for the earliest complete real open first, then pick the earliest hold-back position only when no complete tag was found. Regressions pinned in parser.test.ts and strip.test.ts for both new finding shapes. * fix(web): keep HR-shaped lines inside paragraph regions for inline-code scanning The previous walker closed inline-scan regions on lines matching the HR regex, but `parseBlocks()` in runtime/markdown.tsx does not break a paragraph on HR — its inner accumulation loop only breaks on blank / fence / heading / ul / ol (runtime/markdown.tsx:95-104). HR is only an HR block in the outer loop's first-look, never mid-paragraph. So inputs like `intro \`\n---\n<artifact …>…</artifact>\n---\nclosing \`` are one paragraph in the renderer, whose two stray backticks pair to cover the literal artifact recitation — but the walker was splitting on the `---` lines, leaving the recitation outside skip ranges, and the parser/stripper would treat it as a real tag. Drop HR from the paragraph-break list (HR-shaped lines carry no backticks of their own, so keeping them inside the surrounding region is benign either way) and document the renderer-mirror rationale. Regressions pinned on both sides.	2026-05-11 19:29:22 +08:00