Viewer silently strips `<think>` tags from assistant text, hiding model-emitted reasoning

[alexandrasouly-aisi]

packages/inspect-components/src/chat/MessageContent.tsx's purgeInternalContainers regex-deletes <think>…</think> (alongside <internal> / <content-internal>) from ContentText before rendering:

const purgeInternalContainers = (text: string): string => {
  const internalTags = ["internal", "content-internal", "think"];
  internalTags.forEach((tag) => {
    const regex = new RegExp(`<${tag}[^>]*>[\\s\\S]*?<\\/${tag}>`, "gm");
    text = text.replace(regex, "");
  });
  return text.trim();
};

That's fine when <think> is a scaffold/bridge artifact and the same content is also present as a structured ContentReasoning block. But it isn't always:

Gemini via OpenRouter learns and emits the <think> format itself. inspect_ai's OpenRouter provider replays prior-turn reasoning back to Gemini as inline <think signature="reasoning-details://…">…</think> text (see _providers/openrouter.py — _strip_reasoning_details = "gemini" in family → reasoning_to_think_tag(content), because replaying reasoning_details triggers a Gemini-side thought-signature validation error). After a couple of turns Gemini pattern-matches the format and starts emitting its own reasoning the same way — as literal <think signature="…"><summary>…</summary>…encrypted…</think> in the assistant ContentText, with only a placeholder ContentReasoning(redacted=True, summary="") alongside. In one multi-turn agentic run I have on hand this happened on ~85% of turns from turn 3 onward.

The viewer then shows those turns as Reasoning encrypted by model provider. + an empty text block, when the log actually contains the model's full plaintext reasoning summary — including, in safety-eval contexts, the turn where the model explicitly states an intent it later acts on covertly. A reader using inspect view to audit that trajectory would not see it; they'd have to know to open the raw JSON or read a different agent's tool-result echo of the same content.

Potential fix: drop "think" from internalTags. The markdown pipeline already escapeHtmlCharacters-escapes all tags before rendering, so <think> will display as literal text — noisy, but visible. This could be toggled on and off via a button or a flag, to avoid the noise for users who don't want it.

Higher effort fix: hoist a <think signature="…">…</think> text block into a proper ContentReasoning at render time (mirroring inspect_ai.model._reasoning.parse_content_with_reasoning — parse signature/redacted attrs, extract nested <summary>, and if the signature is reasoning-details://… decode it via the same rules as openrouter_reasoning_details_to_reasoning). That would render it under the existing REASONING (SUMMARY) label instead of as raw escaped markup.

Separately, the provider-side path that produces this (_strip_reasoning_details → text replay → model mimics format) is worth revisiting — but the viewer should never silently discard content regardless of how it got there.

[WatchTree-19]

i think this is already fixed on main, though it's easy to miss because the code lives in the ts-mono submodule rather than this repo.

the purgeInternalContainers snippet quoted above was removed upstream by meridianlabs-ai/ts-mono#370 (2c7ad034, 2026-06-27), "chat: preserve literal reasoning-like tags". that commit dropped the tag stripping and added coverage that pins the new behaviour, including an e2e test named "renders reasoning-like tags as literal assistant text" asserting <think>reasoning evidence</think>, <internal> and <content-internal> are all visible, plus a new MessageContent.test.tsx. it also renamed the scout test from "strips internal tags from text content" to "renders internal tags as literal text content". so it landed as the "drop think from internalTags" option you suggested.

on this repo:

• the submodule pin is currently a016822e (2026-07-14). comparing 2c7ad034...a016822e gives ahead_by 29 / behind_by 0, so the pin includes the fix.
• looking at the submodule bumps, the last pin predating the fix was 0039d610 (06-24) and the first bump after it was 0965cfb6 (07-06), so main picked this up around 2026-07-06.
• the checked-in bundle (src/inspect_ai/_view/dist/assets/index.js) has no internalTags and no "think" string literal. the remaining content-internal hits are the separate tool-result envelope stripping (chat/tools/tool.ts), which is unrelated.

so a viewer built from current main should render model-emitted <think> as literal escaped text rather than dropping it. any chance the run you audited was on a release/build from before the 07-06 bump? if you can still reproduce on current main then i've misread something and it's worth digging further.

the wider points still stand regardless: the "higher effort" fix (hoisting <think signature=...> into a real ContentReasoning so it renders under REASONING (SUMMARY) instead of raw escaped markup) isn't done, and neither is revisiting the provider-side _strip_reasoning_details path that teaches gemini the format in the first place. happy to take a crack at the hoisting piece if that's wanted as a separate issue.

[alexandrasouly-aisi]

Thanks — you're right, we were on 0.3.241 (released 2026-06-22, pre-2c7ad034). I've verified the pristine 0.3.241 dist/assets/index.js still has ["internal", "content-internal", "think"] in purgeInternalContainers; ≥0.3.246 does not. So the base-case fix is in — this can close.

The two follow-ups you mention (hoisting <think signature="reasoning-details://…"> into a rendered ContentReasoning, and revisiting the openrouter provider's _strip_reasoning_details → text-replay path that induces the format-mimicking in the first place) are both worth tracking separately; happy to open those if useful.

[WatchTree-19]

thanks for confirming, and for pinning the exact boundary (0.3.241 vs >=0.3.246) - that's a cleaner answer than i had.

yes please on both follow-ups, and i'd be glad to take the hoisting one if you open it.

rough shape i'd propose for it, so the issue can be scoped tightly: in ts-mono's packages/inspect-components/src/chat/MessageContent.tsx, detect a leading <think ...>...</think> block in a ContentText and lift it into a rendered reasoning block rather than leaving it as escaped markup, mirroring inspect_ai.model._reasoning.parse_content_with_reasoning - parse the signature / redacted attrs, pull out a nested <summary>, and where the signature is a reasoning-details://... value decode it the same way openrouter_reasoning_details_to_reasoning does. render the result under the existing REASONING (SUMMARY) label, and leave the remaining text as the visible content. the existing e2e case from #370 ("renders reasoning-like tags as literal assistant text") would need to stay green for the plain <think> / <internal> / <content-internal> cases, so the lift should only trigger on the signed form rather than any <think>.

the provider-side one (_strip_reasoning_details -> text replay -> gemini mimics the format) feels like the more architectural of the two and probably wants a maintainer's view on intended behaviour before anyone touches it, so i'd leave that one to you to frame.

happy to work to whatever scope you land on - just tag me on the issue.