fix(clients): answer history no longer shows provider context from review comments #35897

Review item 1 was correct: the earlier helper removed every <t3_context version="1">…</t3_context> span, so XML a user typed into an answer was lost, and an answer that was only that XML displayed as empty. The helper now proves an envelope came from provider projection before removing it. projectPendingUserInputAnswer projects each chip separately with projectComposerContextForProvider, which writes [Kind: label; ref=<id>] + \n\n<t3_context version="1">\n<context kind="…" id="<id>"… … </t3_context>. composerContextProviderTextForDisplay walks the provider markers, and for each it checks that the envelope prefix starts right after the marker and that the entry's id attribute equals the marker's ref (escaped as escapeAttribute escapes it). Only then does it replace marker plus envelope with [Kind: label], cut at the first </t3_context>; payload text is escaped so that tag cannot appear inside an entry. Everything else, including typed XML and markers with no envelope, is returned unchanged, as on main. The reviewer's two examples now display in full, so hasQuestionAnswer stays true and the preview shows the answer. The existing repro test case gained an assertion that typed envelope XML stays unchanged. A throwaway test (not committed) ran real projectPendingUserInputAnswer output with two review-comment chips, one comment containing </t3_context>, through getQuestionAnswerText and got 'See [Review comment: notes.md L2] and [Review comment: notes.md L5] please'. Worst case on adversarial input: the marker regex /[([A-Z][^[]:;\n]): ([^[];\n]); ref=([^[]\s]+)]/g has character classes that cannot cross '[' or ']', so its scan is close to linear. Each marker then gets one fixed-prefix startsWith, one indexOf for the closing quote and one for the closing tag. Consumed spans advance the cursor and matches before it are skipped, so total work stays about linear in the answer's length.

Fixes pingdotgg/t3code#35843

Reproduction on main#

Commit 95cb068c10 (origin/main). Ran T3CODE_DEV_SEED=0 vp run dev with a fresh database and added a scratch git project with one uncommitted line in notes.md. Started a thread on the Replay (dev) provider's question scenario. That scenario's question is choice-only, so as a temporary, uncommitted test fixture I pointed it at a copy of its transcript with the enum removed, making the question free-text. Answer submission, storage and display ran as they are on main. In the web client (headless Chromium via Playwright), I typed 'See my note: ' into the answer, added a file comment 'Rename this to something clearer' on notes.md line 2 (the chip 'notes.md L2' appeared in the answer), and pressed Submit. The answered-question work-log row read 'See my note: [Review comment: notes.md L2; ref=review-comment_file-comment-…] <context kind="review-comment" …> file: notes.md range: L2 (1-1) section: File comment comment: Rename this to something clearer md: line changed '. Expanding the row showed the same block as multi-line raw text under the question. A plain answer is unaffected because it contains no envelope. The failing unit test on factory/repro-35843 shows the same thing at the shared helpers.

Evidence#

Real web client, Chromium, 1280×800, light theme, fresh synthetic databases. Start a free-text Replay question, type “See my note: ”, add a file comment “Rename this to something clearer” on notes.md L2, submit the answer with its chip, then expand answer history.

Proven: the scenario fails on base and passes on head. Base displays the internal context XML, comment text, and code; head retains the user's words and readable comment mention without the envelope.

	Before — 95cb068c10	After — bb43f3bb61
Result	Failed: context envelope found in answer history	Passed: no envelope; readable answer retained
Submitted chip
Summary
Expanded history
Recording	Before interaction	After interaction

Scenario source · Temporary startup fixture patch. Apply the patch before running the scenario; it removes the replay question's enum identically on both sides. Neither artifact was committed.

Not shown:

• Desktop and mobile were not run; this proof covers the real web client using the shared answer-history helper.
• Literal user-authored XML preservation and other context kinds were not exercised; this scenario targets the reported review-comment answer.

Tests#

vp test run packages/client-runtime/src/work-log/userInput.test.ts packages/shared/src/composerContextReferences.test.ts: 2 files, 40 tests passed. A throwaway test (deleted, not committed) ran real projectPendingUserInputAnswer output with two review-comment chips through getQuestionAnswerText: passed, output shown in summary. vp run check:changed: format, lint, knip, typecheck (35 packages) and tests (3 files) all pass.

Not in this change#

• Answers are still stored in provider form on the server. Storing the composer form plus records, as normal chat messages do, would let clients render real chips, but that is a contract and server change.
• An answer chip whose record is missing still shows its raw markdown t3-context link, which is unchanged from main.
• The user's own text is not escaped when an answer is projected for the provider, so a user who types an exact marker followed by a matching envelope is indistinguishable from projected output. That behaviour was already on main and is out of scope here.

Who did the work#

• Written by Claude Opus 5.5, in Claude Code
• Reviewed by GPT-6.1 Sol, in Codex, and by Claude Opus 5.5, each on its own
• Proof captured by GPT-6.1 Sol and checked by Claude Opus 5.5

Run by the t3code factory in T3 Code.