fix(server): imported Codex and Claude history no longer shows internal records#25923

Problem

Onboarding import showed provider-internal records as if the user had typed them. Codex threads were titled # Files mentioned by the user: (or other injected markup), and their history showed user instructions, AGENTS.md, environment context, in-app browser context and the raw <send_user_message_question_reply> JSON as user messages. Claude threads showed <command-name>/compact… and <local-command-stdout>Compacted</local-command-stdout> bubbles. Reproduced on main (28286d523f) by importing synthetic Codex and Claude homes through /welcome in the web app.

Change

AgentSessionScanner now imports user text the way each provider shows it:

• Codex: when a rollout records user events (event_msg user_message in legacy rollouts, item_completed UserMessage in paginated ones), those are the only user messages; the response_item user copies carry injected context and are skipped, as in Codex's own thread history. This replaces the turn-id matching. Rollouts with no user events keep today's behaviour.
• Codex: a prompt that opens with generated context (IDE context, files mentioned, in-app browser context, referenced chats) is shown from after its last ## My request for Codex: line, as Codex's extract_prompt_request does. An attachment sent without text keeps its recorded text.
• Codex: a complete question-reply envelope is shown as > question and answer, as Codex's async_question_reply does; anything malformed stays verbatim.
• Claude: a local slash-command record is imported as typed (/compact, /model opus), and <local-command-stdout>/<local-command-stderr> records are omitted.

The welcome-wizard guide describes the new behaviour. Already-imported threads keep their stored messages.

Scope and approval

Fixes pingdotgg/t3code#24067

Covers the raw-record symptoms (4, 6, 8) and the injected-markup titles of #24067. Saved Codex thread names are #24982; missing compressed rollouts and record limits are #23173 and #23168. Image rendering and memory citations are not changed.

Verification

From apps/server: vp test run src/project/AgentSessionScanner.test.ts src/project/AgentSessionImporter.test.ts → 2 files, 69 tests passed. Each new test failed before its fix (Codex user events only, request after the files prefix, Claude commands as typed, paginated rollout, question replies across record kinds, marker in a code block, last marker after IDE context, referenced chats, attachment-only prompt). vp run check:changed passed (format, lint, knip, typecheck, tests).

Trade-offs: in a rollout that mixes response-only user turns with later user events, the response-only turns are not imported (Codex's own history drops them too); Claude local command output (/cost, /status, errors) is not imported.

Not checked: Windows and macOS (the parser is platform-independent; reproduced on Linux), desktop and mobile clients (they render the same server-imported messages), a real Codex desktop rollout (fixtures follow Codex's source), and the matrix rows below marked not captured.

Evidence

Captured 4/4 rows in the lead’s narrowed A–D matrix, at both widths and both themes: 56 stills, 28 before/after pairs, 56 diffs, 28 sheets. Before 28286d523f; after f651bd79c4. Every capture is the real web app, imported through onboarding into fresh isolated state.

Evidence index · reproduce.md · make-capture-fixtures.py · browser-driver.mjs · synthetic-fixtures.zip

T3 Browser panel; Linux fs-11; Chrome 156.0.0.0; no device or UA emulation; DPR 2; zoom/font scale 100%; root font 16px; frozen client Date 2026-10-27T12:00:00Z. Wide: 1440×1000 CSS px, full viewport exported as 1280×889 PNG by the tool. Narrow: 390×844 CSS px, full viewport exported as 780×1688 PNG. Dark and light on both sides. Reduced-motion override unavailable; actual prefers-reduced-motion=false.

Each row below states its CSS viewport and theme. These are full-viewport captures; wide PNGs are tool-downsampled, not cropped. Top pairs use scrollTop=0. Supplemental bottom pairs use the same end-of-history boundary on both sides; shorter after histories can have a different numeric offset. Exact offsets are in capture-manifest.json. Narrow sidebar title proof is shared across A–D.

Additional capture variation: the wide light C/D frames contain a small zero-threshold-only sidebar-border raster difference (C: 397 pixels, channel delta ≤8/255; D: 9 pixels, delta 1/255). It is disclosed below; no unrelated visible content or layout change was found. Before every accepted still: fonts ready, resource network quiet, pointer outside the content, same provider banner dismissed, thread details closed, then two screenshots ≥500ms apart with identical pixels. All 28 sheets were opened and visually inspected, including every full source frame and both diffs. Diff0 flags every nonzero RGB-channel change at full magenta intensity (binary ×255). DiffT16 flags max channel delta >16/255 in red on a faded grayscale after frame; no areas are masked.

What else changed

• A: title changes from the files-prefix line to “Hey, what's up?” in the sidebar and breadcrumb; six messages become four. Instructions/environment bubbles disappear; the generated request prefix disappears; the question-reply JSON becomes a blockquote question and answer. Both assistant replies remain.
• B: six messages become four; injected AGENTS instructions/environment disappear. Both paginated user turns and both assistants remain; title stays “Build a paginated transcript viewer.”
• C: eight messages become six; tag wrappers become /compact and /model opus, stdout/stderr bubbles disappear. Prompt, follow-up, assistants and title remain.
• Removing and shortening messages moves later messages, timestamps, settled footer, minimap marks and the scroll-end affordance. Text behind the fixed translucent composer changes where the before history overflows. These are consequences of transcript content/height.
• D: both narrow full frames have zero changed pixels in each theme. Wide frames have the shared sidebar’s changed A title (837 dark / 863 light changed pixels at zero; 671 each at threshold16); the control conversation itself is identical. The wide light C pairs additionally have 397 sidebar-border pixels with channel delta ≤8/255; the wide light D pair has 9 sidebar-border pixels with delta 1/255. No unrelated visible layout or styling change was found.

After onboarding flow recording

Steps/timestamps in the 36.07s uncut after clip: 00:00–00:05 Projects → Select none → Select all → Import 1 project; 00:06–00:07 imported/Settled list; 00:08–00:13 open A and dismiss the unauthenticated Codex banner; 00:14–00:17 open B; 00:18–00:22 open C and dismiss the same banner; 00:23–00:36 open and hold D. Thread details close during navigation. The clip includes real loading transitions; settled stills above are the comparison evidence. No before interaction recording was requested for these static rows.

after-import-flow.mp4

Recording environment: T3 Browser panel; Linux fs-11; Chrome 156.0.0.0; no device or UA emulation; DPR 2; zoom/font scale 100%; root font 16px; frozen client Date 2026-10-27T12:00:00Z. Wide: 1440×1000 CSS px, full viewport exported as 1280×889 PNG by the tool. Narrow: 390×844 CSS px, full viewport exported as 780×1688 PNG. Dark and light on both sides. Reduced-motion override unavailable; actual prefers-reduced-motion=false. Recording profile: 1440×1000 CSS px, dark; recorder output 2880×2000 video. This is responsive web, not native mobile.

A — Codex legacy desktop

Expected after: title “Hey, what's up?”; clean request, assistant, quoted “Which database?” / “Use PostgreSQL”, assistant. The question envelope is present in both a response item and canonical event_msg.user_message in the seed.

B — Codex paginated

Expected after: request, assistant, follow-up, assistant. There are no legacy user_message events; canonical turns are item_completed/UserMessage text items. Response records carry generated instructions/environment.

C — Claude commands

Expected after: plain prompt, assistant, /compact, /model opus, plain follow-up, assistant; no stdout/stderr bubbles.

D — Unchanged control

The ordinary prompt’s fenced literal “## My request for Codex:” and introduction remain intact.

Shared narrow sidebar proof — all four titles

Older matrix / unit coverage and limits

The supplied older 22-row audit targets an earlier head. The lead explicitly narrowed visual proof to A–D plus one after import-flow clip. A covers C1, C covers H1, D covers the C5 preservation control, and B adds current paginated Codex. The following older rows are covered by unit tests for parser behavior per that scope; they are not claimed as separate real-client captures:

Not captured, client render paths unchanged and parser outcome covered by unit tests: an assistant-first history (mixed rollout), a long cleaned message's expand/collapse, and the partial/all-skipped import warnings. The proof audit flagged the wide light C/D pairs for a sidebar-border raster difference of at most 8/255 (gone at threshold 16) that this change does not cause. An independent rebuild of row A at 390×844 dark matched the delivered capture with zero changed pixels. The captures were taken at f651bd79c4; the later commits produce identical A–D output (checked by running both parser revisions).

Production fix, tests, and PR description by Claude Opus 5.5 in Claude Code (T3 Code issue flow); reviews and evidence by GPT-6.1 Sol in Codex and Claude Opus 5.5.

🤖 Generated with Claude Code