[only21mil]

Problem

The welcome wizard guide (docs/user/welcome-wizard.md) says conversation import skips files larger than 16 MiB and reads 64 MiB per project on each attempt. Both numbers are out of date. Someone with a large Codex or Claude session reads that it will be skipped and may not try, even though the app imports it.

#10430 raised the per-file limit from 16 MiB to 4 GiB and the per-attempt byte budget from 64 MiB to 4 GiB. It also added a separate 32 MiB memory limit for the history kept from one conversation. The guide was not updated.

Change

One paragraph in docs/user/welcome-wizard.md, 4 lines added and 2 removed:

• Files larger than 4 GiB are skipped, not 16 MiB.
• A conversation is also skipped when the message text and metadata it keeps need more than 32 MiB of memory. Tool output does not count toward that limit.
• Each import attempt reads up to 4 GiB per project, not 64 MiB. The 100-file and 100,000-record limits were already correct.

No code changes.

Scope and approval

There is no issue for this. It is a factual correction to user docs that contradict the code since #10430. Each new number is taken from a constant in apps/server/src/project/AgentSessionScanner.ts.

Verification

Checked each statement against the code on main:

• 4 GiB per file: MAX_IMPORTED_TRANSCRIPT_BYTES = 4 * 1024 * 1024 * 1024 (AgentSessionScanner.ts:64). The file is skipped before reading (:823) and in the import loop (:1403).
• 32 MiB kept history: MAX_IMPORT_HISTORY_BYTES = 32 * 1024 * 1024 (:66). The reader throws "Transcript selected history exceeds the 32 MiB memory budget" (:841-844). The read then returns null (:906-908) and the conversation is marked Skipped.
• Memory, not file bytes: the budget counts selected JSON values at 2 bytes per character plus 64 per token (AgentSessionJson.ts:67-69), and key names at 2 bytes per character (:164). Unselected values are never assembled, so the values of tool output and tool input do not count. The selected fields are the TranscriptRecord schema (AgentSessionScanner.ts:85-110): message text blocks and session metadata.
• 4 GiB per project per attempt: MAX_IMPORT_BYTES (:67) is set once per recentThreads call (:1367). The importer calls it once per project (AgentSessionImporter.ts:214).
• Unchanged sentences still hold: one conversation at a time (:1370 stream concurrency 1, :1462 single-permit lock), 100 files (:68) and 100,000 records (:69).

pnpm exec vp fmt --check docs/user/welcome-wizard.md passes. No code reads or generates this file, so there is no test to add.

Not checked:

• I did not import a real session over 16 MiB in the app. fix(server): import transcripts with oversized tool records #10430 added that behavior; this PR only changes the docs.
• Nothing on screen changes, so there are no screenshots.

Two open PRs, #22411 and #23122, append a paragraph after this one. Their doc hunks still apply cleanly on top of this change.

Prior PRs: @Yash-Singh1's #10430 changed the limits this paragraph describes. @t3dotgg's #5362 wrote the original paragraph.

Documentation change and PR description written with Claude Opus 5.5 in T3 Code Nightly.

🤖 Generated with Claude Opus 5.5 in T3 Code Nightly

[macroscopeapp]

Approvability

Verdict: Approved at 47dcf31

Macroscope's review found this PR approvable — This updates stale conversation import limits and clarifies the existing memory constraint in user documentation only. It does not modify application behavior, product defaults, or static-analysis configuration.

^{You can add or adjust custom eligibility rules. Learn more.}