[BunsDev]

Now also carries #302. Val merged #302 (compare a turn's patch with the working tree, file by file) into this branch as eda8445. This PR's tree now matches #302's last head 4fd4af8, which passed every CI check and the local runs listed on #302. Merging this PR lands both CF-5 changes. Details are in #302.

Part of #278 (CF-5), tracker #273. This takes one open item from the CF-5 first-part checkpoint: a Codex memory citation can now be read. The cf5-evidence worktree was released after #287; the claim is here.

What changes

A reply's "From Codex memory" citation was display-only. Each one now has Show lines and Copy. Both call a new main-process command, evidence:memory.

main/engine/memory-citations.ts reads only inside Codex's memories folder: $CODEX_HOME/memories, else ~/.codex/memories. It refuses:

• an absolute path, or any ..;
• a symlink that resolves outside the folder (a symlink that stays inside is followed);
• a directory;
• a file over 8 MiB;
• a range that starts past the end of the file.

It returns only the cited lines, 1-based and inclusive, cut to 64 KiB, and says when the range ran past the end or the excerpt was cut.

It opens nothing in another app, executes nothing and fetches nothing.

Mac only. The command is registered local-only. A paired desktop never reads this Mac's memory files, and its card says so instead of showing an unknown-command error. The offline Store demo refuses it like its other system commands.

Streamer Mode masks the shown lines; Copy puts the lines themselves on the clipboard, as with other copies.

Native evidence

From local codex-cli 0.163.0 sessions on this Mac:

• agentMessage.memoryCitation.entries[] carry path, lineStart, lineEnd and note.
• The paths are relative to the memories folder, for example MEMORY.md.
• For a one-line citation, 3 of the note's 5 words appear on the cited line counted from 1, and none on it counted from 0, so lines are 1-based.

No memory text was copied into the repository; fixtures are synthetic.

Tests, red first

• test/memory-citations.test.ts, 7 tests:

  • the folder resolution;
  • 1-based inclusive ranges;
  • a range past the end, cut and flagged;
  • absolute paths, .., symlink escapes, a directory and an empty path refused;
  • an inside symlink followed;
  • bad ranges, a missing file, an oversized file and a missing folder refused;
  • a long excerpt bounded and flagged.

  These failed to compile before the module existed.

• e2e/conversation-evidence.spec.ts, 2 new tests. The app runs with a test-only CODEX_HOME, never the developer's.

  • Show lines gives exactly lines 3–5 with "Lines 3–5 of 7". Copy puts exactly those lines on the clipboard, no tool item is added, and the thread stays idle.
  • A citation to ../secret.txt shows "does not point inside Codex's memory folder" and nothing from that file.

  Both failed before the UI existed: 2 failed, 4 passed. All 6 pass now.

Checks

The change is one signed commit, 3165795 (3165795508646a0f2b33faf57c20449b5bb15dba). The PR head merges main at db0bbfd (#296, docs only, no app files) on top of it. Full local run on 3165795 on macOS (Apple Silicon), Node 24.18.1, pnpm 10.34.5:

Command	Result
pnpm install --frozen-lockfile	passed
pnpm build	passed
pnpm typecheck	passed
pnpm test	passed: desktop 1,530 passed and 16 skipped of 1,546 (existing skips); every other package fully green
pnpm test:e2e	desktop 333 passed, 3 skipped, 1 failed; site 7/7

The one failure was owner-removal.spec.ts:75: "Test timeout of 90000ms exceeded while running afterEach". That hook closes the app and deletes its temp folder; the test body itself passed. The machine's disk dropped to 1.9 GB free during the run. This PR does not touch that spec or what it covers. Run alone afterwards, owner-removal.spec.ts passed 6/6. CI decides.

Still open on #278

• The retained full log, up to 10 MiB per tool item. It needs raw output deltas through the observation contract, touching runner.ts, codex.ts and claude.ts, which Make Claude Stop truthful on claude -p: native interrupt, bounded exit, withdrawn prompts (CF-2) #286 and Answer each approval and question once, and show Codex questions as cards (CF-3) #293 are changing now.
• Structured web-search results. Codex's results are opaque JSON, with no local sample to build on.
• Comparing a turn's patch with the working tree beyond the label.
• New owned-artifact actions.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
modex	Ready	Preview	Oct 19, 2026 3:30am UTC

[BunsDev]

Coordination note from the #273 tracker session: #297 and #299 compose cleanly in either order. Both edit e2e/conversation-evidence.spec.ts. #299 adds a test-only CODEX_HOME to the setup, one more cited reply, and two memory-citation tests.

I trial-merged cf5-evidence-race into #299's head 3111e72; there were no conflicts. The combined conversation-evidence.spec.ts passed 6 of 6, with #297's per-test beforeEach. I then aborted the trial merge; nothing was pushed to either branch.

[github-code-quality]

Code coverage overview

Languages: TypeScript

TypeScript / code-coverage-agent

main 396b17c	cf5-memory-citations 9321b3b	+/-
63%	63%	0%

View coverage of the 7 most impacted files.

File	main 396b17c	cf5-memory-citations 9321b3b	+/-
apps/desktop/sr...ckends/codex.ts	96%	96%	0%
apps/store-desk...rc/main/demo.ts	95%	95%	0%
apps/desktop/sr...atchCompare.tsx	0%	0%	0%
apps/desktop/sr...red/evidence.ts	100%	100%	0%
apps/desktop/sr...shared/types.ts	100%	100%	0%
apps/desktop/sr.../patch-state.ts	0%	98%	+98%
apps/desktop/sr...ry-citations.ts	0%	100%	+100%

ℹ️ Calculations represent line coverage. Results may change as calculations are updated.

---

_{Updated October 19, 2026 04:07 UTC}