diff --git a/.github/dependabot.yml b/.github/dependabot.yml
index 30fdfc2..8c71c93 100644
--- a/.github/dependabot.yml
+++ b/.github/dependabot.yml
@@ -9,6 +9,15 @@ updates:
     commit-message:
       prefix: "deps(rust)"
     labels: [dependencies, rust]
+    # Tauri's plugins pin these to an older major, so bumping ours alone only adds a second copy.
+    # Drop each once Tauri moves.
+    ignore:
+      - dependency-name: zbus
+        update-types: [version-update:semver-major]
+      - dependency-name: dirs
+        update-types: [version-update:semver-major]
+      - dependency-name: sha2
+        update-types: [version-update:semver-minor]
     # Groups match in the order listed, so `tauri` claims its crates before `rust-minor`.
     groups:
       tauri:
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index c580a82..65cf761 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -76,6 +76,19 @@ jobs:
       - name: Test
         run: cargo test --locked --workspace --no-fail-fast
 
+      # Days and weeks follow the local zone, so run the tests again in the two zones furthest from UTC.
+      - name: Test in UTC+14
+        if: runner.os == 'Linux'
+        env:
+          TZ: Pacific/Kiritimati
+        run: cargo test --locked --workspace --no-fail-fast
+
+      - name: Test in UTC-11
+        if: runner.os == 'Linux'
+        env:
+          TZ: Pacific/Pago_Pago
+        run: cargo test --locked --workspace --no-fail-fast
+
   frontend:
     name: frontend
     runs-on: ubuntu-24.04
@@ -105,6 +118,17 @@ jobs:
       - name: Tests with coverage
         run: pnpm --dir app test:coverage
 
+      # app/vitest.config.ts honors TZ; the default run above is pinned to UTC.
+      - name: Tests in UTC+14
+        env:
+          TZ: Pacific/Kiritimati
+        run: pnpm --dir app exec vitest run
+
+      - name: Tests in UTC-11
+        env:
+          TZ: Pacific/Pago_Pago
+        run: pnpm --dir app exec vitest run
+
       - name: Build
         run: pnpm --dir app build
 
@@ -288,7 +312,7 @@ jobs:
         with:
           save-if: ${{ github.ref == 'refs/heads/main' }}
 
-      - uses: taiki-e/install-action@5488ac5a951645ce95062214902ac4b6bc1a1037 # v2
+      - uses: taiki-e/install-action@1919c4a95903d2c33e4a221102edca87309fc14b # v2
         with:
           tool: cargo-llvm-cov
 
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 0313cad..76d9178 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -4,6 +4,27 @@ All notable changes are listed here. Format follows [Keep a Changelog](https://k
 
 ## [Unreleased]
 
+## [0.5.7] - 2026-11-04
+
+### Changed
+
+- Your history is scored again once, under the fixed rules below.
+- Stored correction text is cleaned again once, with the wider redaction below.
+- Lessons: clearer hints and examples for Evidence and Target, and the Context lesson says the checker can't spot it.
+
+### Fixed
+
+- Messages from other agents and scheduled tasks in Claude Code no longer count as yours.
+- "No, that's all, thanks", "Nah, it's cool" and similar no longer count as corrections.
+- "You're right" after you ask a question is no longer a correction.
+- Adding context ("by the way…", "one other thing…") no longer counts against the AI.
+- Messages typed while Claude was busy are counted even when they include an image.
+- ChatGPT messages with an attachment are counted.
+- Codex handoff and voice chats are counted.
+- Large exports no longer use many times their size in memory.
+- Redaction catches more secrets, such as `*_KEY=` values, `pass:`, and Google and Slack tokens. Dates are no longer hidden.
+- Bedrock model names are recognized, and Opus 4 and 4.1 have prices.
+
 ## [0.5.6] - 2026-11-01
 
 ### Added
@@ -273,7 +294,8 @@ First release.
 - Per-source consent: nothing is read until you allow it, and Forget deletes a source's data
 - `noactually` command-line tool (`sources`, `allow`, `sync`, `stats`)
 
-[Unreleased]: https://github.com/HTMLGuyLLC/NoActually/compare/v0.5.6...HEAD
+[Unreleased]: https://github.com/HTMLGuyLLC/NoActually/compare/v0.5.7...HEAD
+[0.5.7]: https://github.com/HTMLGuyLLC/NoActually/compare/v0.5.6...v0.5.7
 [0.5.6]: https://github.com/HTMLGuyLLC/NoActually/compare/v0.5.5...v0.5.6
 [0.5.5]: https://github.com/HTMLGuyLLC/NoActually/compare/v0.5.4...v0.5.5
 [0.5.4]: https://github.com/HTMLGuyLLC/NoActually/compare/v0.5.3...v0.5.4
diff --git a/Cargo.lock b/Cargo.lock
index 80409b8..16a117f 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -90,9 +90,9 @@ dependencies = [
 
 [[package]]
 name = "async-executor"
-version = "1.14.0"
+version = "1.14.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "c96bf972d85afc50bf5ab8fe2d54d1586b4e0b46c97c50a0c9e71e2f7bcd812a"
+checksum = "f426a31126c735126297bc5ea7a521263c9faff4f63ca415267ee625651c4f37"
 dependencies = [
  "async-task",
  "concurrent-queue",
@@ -497,6 +497,17 @@ version = "0.2.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "f079e83a288787bcd14a6aea84cee5c87a67c5a3e660c30f557a3d24761b3527"
 
+[[package]]
+name = "chacha20"
+version = "0.10.2"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "65c35e4b699c7e15ccbe7ee35c005e4fc0a278d22238a2857e6ce2dadeda1b06"
+dependencies = [
+ "cfg-if",
+ "cpufeatures 0.3.1",
+ "rand_core",
+]
+
 [[package]]
 name = "chrono"
 version = "0.4.45"
@@ -597,6 +608,15 @@ dependencies = [
  "libc",
 ]
 
+[[package]]
+name = "cpufeatures"
+version = "0.3.1"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "5ca28b0ae3115b884660db4118d803791fd6756b6e88f39c0f3f7859060d7566"
+dependencies = [
+ "libc",
+]
+
 [[package]]
 name = "crc32fast"
 version = "1.5.3"
@@ -1379,6 +1399,7 @@ dependencies = [
  "js-sys",
  "libc",
  "r-efi 6.0.0",
+ "rand_core",
  "wasm-bindgen",
 ]
 
@@ -2233,9 +2254,9 @@ dependencies = [
 
 [[package]]
 name = "log"
-version = "0.4.34"
+version = "0.4.35"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f9f8bd3e56ce4dfc153cf470fffbfa98c7620958b312ca5c3a4b8d5181fd13c6"
+checksum = "5d15f86d12da44cd8d4ec3c73984537eb7f4cadcb9776a575a0a8a05ca11005c"
 
 [[package]]
 name = "mac-notification-sys"
@@ -2402,7 +2423,7 @@ dependencies = [
 
 [[package]]
 name = "noactually-app"
-version = "0.5.6"
+version = "0.5.7"
 dependencies = [
  "chrono",
  "dirs 6.0.0",
@@ -2430,7 +2451,7 @@ dependencies = [
 
 [[package]]
 name = "noactually-cli"
-version = "0.5.6"
+version = "0.5.7"
 dependencies = [
  "anyhow",
  "chrono",
@@ -2446,7 +2467,7 @@ dependencies = [
 
 [[package]]
 name = "noactually-engine"
-version = "0.5.6"
+version = "0.5.7"
 dependencies = [
  "regex",
  "serde",
@@ -2456,7 +2477,7 @@ dependencies = [
 
 [[package]]
 name = "noactually-sources"
-version = "0.5.6"
+version = "0.5.7"
 dependencies = [
  "chrono",
  "dirs 6.0.0",
@@ -2471,7 +2492,7 @@ dependencies = [
 
 [[package]]
 name = "noactually-store"
-version = "0.5.6"
+version = "0.5.7"
 dependencies = [
  "chrono",
  "dirs 6.0.0",
@@ -3092,15 +3113,6 @@ version = "0.2.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "4a6394b9e965e73d0a289ee54f589087e2c676aedf60885baf52c76b771e4958"
 
-[[package]]
-name = "ppv-lite86"
-version = "0.2.21"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "85eae3c4ed2f50dcfe72643da4befc30deadb458a9b590d720cde2f2b1e97da9"
-dependencies = [
- "zerocopy",
-]
-
 [[package]]
 name = "precomputed-hash"
 version = "0.1.1"
@@ -3207,32 +3219,20 @@ checksum = "f8dcc9c7d52a811697d2151c701e0d08956f92b0e24136cf4cf27b57a6a0d9bf"
 
 [[package]]
 name = "rand"
-version = "0.9.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b9ef1d0d795eb7d84685bca4f72f3649f064e6641543d3a8c415898726a57b41"
-dependencies = [
- "rand_chacha",
- "rand_core",
-]
-
-[[package]]
-name = "rand_chacha"
-version = "0.9.0"
+version = "0.10.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d3022b5f1df60f26e1ffddd6c66e8aa15de382ae63b3a0c1bfc0e4d3e3f325cb"
+checksum = "65c9fb96cbc91e3478eaae79a69fcd3f1ae4ad052e471fe6732fff548984b4af"
 dependencies = [
- "ppv-lite86",
+ "chacha20",
+ "getrandom 0.4.3",
  "rand_core",
 ]
 
 [[package]]
 name = "rand_core"
-version = "0.9.5"
+version = "0.10.1"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "76afc826de14238e6e8c374ddcc1fa19e374fd8dd986b0d2af0d02377261d83c"
-dependencies = [
- "getrandom 0.3.4",
-]
+checksum = "63b8176103e19a2643978565ca18b50549f6101881c443590420e4dc998a3c69"
 
 [[package]]
 name = "raw-window-handle"
@@ -3369,9 +3369,9 @@ dependencies = [
 
 [[package]]
 name = "rfd"
-version = "0.16.1"
+version = "0.17.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f4fe518b6db89699d59862e4d1e3a0226316090f4e4ec09f7dd686be25634c7c"
+checksum = "20dafead71c16a34e1ff357ddefc8afc11e7d51d6d2b9fbd07eaa48e3e540220"
 dependencies = [
  "block2",
  "dispatch2",
@@ -3379,16 +3379,18 @@ dependencies = [
  "gobject-sys",
  "gtk-sys",
  "js-sys",
+ "libc",
  "log",
  "objc2",
  "objc2-app-kit",
  "objc2-core-foundation",
  "objc2-foundation",
+ "percent-encoding",
  "raw-window-handle",
  "wasm-bindgen",
  "wasm-bindgen-futures",
  "web-sys",
- "windows-sys 0.60.2",
+ "windows-sys 0.61.2",
 ]
 
 [[package]]
@@ -3838,7 +3840,7 @@ source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "a7507d819769d01a365ab707794a4084392c824f54a7a6a7862f8c3d0892b283"
 dependencies = [
  "cfg-if",
- "cpufeatures",
+ "cpufeatures 0.2.17",
  "digest",
 ]
 
@@ -4339,9 +4341,9 @@ dependencies = [
 
 [[package]]
 name = "tauri-plugin-dialog"
-version = "2.8.3"
+version = "2.8.4"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b7461aaacf1901915ea7b9f5cd74f8224a351ed96af0a62c07fce934003cb34b"
+checksum = "ce0fb9c2b6d34a9720627ed930ebc8af26f61e648695aecaeb03d5c33c5896f5"
 dependencies = [
  "log",
  "raw-window-handle",
@@ -4402,9 +4404,9 @@ dependencies = [
 
 [[package]]
 name = "tauri-plugin-notification"
-version = "2.7.0"
+version = "2.8.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b939d7b19c8af1381a72ef8474b3aba2b56c1ddac2c23f0c8e96abbcf86b198b"
+checksum = "5bc6c194a95a535abdfbe3bc1af9a43935b8b9b29c216b8e71af8b59132703cf"
 dependencies = [
  "log",
  "notify-rust",
@@ -4459,9 +4461,9 @@ dependencies = [
 
 [[package]]
 name = "tauri-plugin-single-instance"
-version = "2.5.4"
+version = "2.5.5"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "561cd99484561bc39c954768e7b41a66cebbe1ac3d9d081c34361f740c489e2c"
+checksum = "726b6da0dd5bd1344f84f4108a622c07d01923baacb5c803a605189672c5352f"
 dependencies = [
  "serde",
  "serde_json",
@@ -5497,15 +5499,6 @@ dependencies = [
  "windows-targets 0.52.6",
 ]
 
-[[package]]
-name = "windows-sys"
-version = "0.60.2"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f2f500e4d28234f72040990ec9d39e3a6b950f9f22d3dba18416c35882612bcb"
-dependencies = [
- "windows-targets 0.53.5",
-]
-
 [[package]]
 name = "windows-sys"
 version = "0.61.2"
@@ -5539,30 +5532,13 @@ dependencies = [
  "windows_aarch64_gnullvm 0.52.6",
  "windows_aarch64_msvc 0.52.6",
  "windows_i686_gnu 0.52.6",
- "windows_i686_gnullvm 0.52.6",
+ "windows_i686_gnullvm",
  "windows_i686_msvc 0.52.6",
  "windows_x86_64_gnu 0.52.6",
  "windows_x86_64_gnullvm 0.52.6",
  "windows_x86_64_msvc 0.52.6",
 ]
 
-[[package]]
-name = "windows-targets"
-version = "0.53.5"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "4945f9f551b88e0d65f3db0bc25c33b8acea4d9e41163edf90dcd0b19f9069f3"
-dependencies = [
- "windows-link",
- "windows_aarch64_gnullvm 0.53.1",
- "windows_aarch64_msvc 0.53.1",
- "windows_i686_gnu 0.53.1",
- "windows_i686_gnullvm 0.53.1",
- "windows_i686_msvc 0.53.1",
- "windows_x86_64_gnu 0.53.1",
- "windows_x86_64_gnullvm 0.53.1",
- "windows_x86_64_msvc 0.53.1",
-]
-
 [[package]]
 name = "windows-threading"
 version = "0.2.1"
@@ -5593,12 +5569,6 @@ version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "32a4622180e7a0ec044bb555404c800bc9fd9ec262ec147edd5989ccd0c02cd3"
 
-[[package]]
-name = "windows_aarch64_gnullvm"
-version = "0.53.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "a9d8416fa8b42f5c947f8482c43e7d89e73a173cead56d044f6a56104a6d1b53"
-
 [[package]]
 name = "windows_aarch64_msvc"
 version = "0.42.2"
@@ -5611,12 +5581,6 @@ version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "09ec2a7bb152e2252b53fa7803150007879548bc709c039df7627cabbd05d469"
 
-[[package]]
-name = "windows_aarch64_msvc"
-version = "0.53.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b9d782e804c2f632e395708e99a94275910eb9100b2114651e04744e9b125006"
-
 [[package]]
 name = "windows_i686_gnu"
 version = "0.42.2"
@@ -5629,24 +5593,12 @@ version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "8e9b5ad5ab802e97eb8e295ac6720e509ee4c243f69d781394014ebfe8bbfa0b"
 
-[[package]]
-name = "windows_i686_gnu"
-version = "0.53.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "960e6da069d81e09becb0ca57a65220ddff016ff2d6af6a223cf372a506593a3"
-
 [[package]]
 name = "windows_i686_gnullvm"
 version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "0eee52d38c090b3caa76c563b86c3a4bd71ef1a819287c19d586d7334ae8ed66"
 
-[[package]]
-name = "windows_i686_gnullvm"
-version = "0.53.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fa7359d10048f68ab8b09fa71c3daccfb0e9b559aed648a8f95469c27057180c"
-
 [[package]]
 name = "windows_i686_msvc"
 version = "0.42.2"
@@ -5659,12 +5611,6 @@ version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "240948bc05c5e7c6dabba28bf89d89ffce3e303022809e73deaefe4f6ec56c66"
 
-[[package]]
-name = "windows_i686_msvc"
-version = "0.53.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1e7ac75179f18232fe9c285163565a57ef8d3c89254a30685b57d83a38d326c2"
-
 [[package]]
 name = "windows_x86_64_gnu"
 version = "0.42.2"
@@ -5677,12 +5623,6 @@ version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "147a5c80aabfbf0c7d901cb5895d1de30ef2907eb21fbbab29ca94c5b08b1a78"
 
-[[package]]
-name = "windows_x86_64_gnu"
-version = "0.53.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "9c3842cdd74a865a8066ab39c8a7a473c0778a3f29370b5fd6b4b9aa7df4a499"
-
 [[package]]
 name = "windows_x86_64_gnullvm"
 version = "0.42.2"
@@ -5695,12 +5635,6 @@ version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "24d5b23dc417412679681396f2b49f3de8c1473deb516bd34410872eff51ed0d"
 
-[[package]]
-name = "windows_x86_64_gnullvm"
-version = "0.53.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0ffa179e2d07eee8ad8f57493436566c7cc30ac536a3379fdf008f47f6bb7ae1"
-
 [[package]]
 name = "windows_x86_64_msvc"
 version = "0.42.2"
@@ -5713,12 +5647,6 @@ version = "0.52.6"
 source = "registry+https://github.com/rust-lang/crates.io-index"
 checksum = "589f6da84c646204747d1270a2a5661ea66ed1cced2631d546fdfb155959f9ec"
 
-[[package]]
-name = "windows_x86_64_msvc"
-version = "0.53.1"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d6bbff5f0aada427a1e5a6da5f1f98158182f26556f345ac9e04d36d0ebed650"
-
 [[package]]
 name = "winnow"
 version = "0.5.40"
@@ -5926,26 +5854,6 @@ dependencies = [
  "serde",
 ]
 
-[[package]]
-name = "zerocopy"
-version = "0.8.63"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "e5fe1f8f1b06191a00962174c61aa5005e0bb391a6d80d07e24d115c01a92ed8"
-dependencies = [
- "zerocopy-derive",
-]
-
-[[package]]
-name = "zerocopy-derive"
-version = "0.8.63"
-source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "863ad3ac83293fb4d740aedbfdc9240dd8d1a50c1099acd76ce80ce7c7230c7f"
-dependencies = [
- "proc-macro2",
- "quote",
- "syn 2.0.119",
-]
-
 [[package]]
 name = "zerofrom"
 version = "0.1.8"
diff --git a/Cargo.toml b/Cargo.toml
index 9e248d6..6a6e983 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -3,7 +3,7 @@ resolver = "2"
 members = ["crates/engine", "crates/sources", "crates/store", "crates/cli", "app/src-tauri"]
 
 [workspace.package]
-version = "0.5.6"
+version = "0.5.7"
 edition = "2021"
 # Oldest Rust that builds the whole workspace, app included. CI's `msrv` job checks it.
 rust-version = "1.90"
diff --git a/app/package.json b/app/package.json
index 4967abc..a247f1e 100644
--- a/app/package.json
+++ b/app/package.json
@@ -1,7 +1,7 @@
 {
   "name": "noactually-app",
   "private": true,
-  "version": "0.5.6",
+  "version": "0.5.7",
   "type": "module",
   "packageManager": "pnpm@10.32.1",
   "engines": {
@@ -27,10 +27,10 @@
   "devDependencies": {
     "@biomejs/biome": "2.6.4",
     "@tauri-apps/cli": "2.12.4",
-    "@vitest/coverage-v8": "5.0.5",
+    "@vitest/coverage-v8": "5.1.0",
     "happy-dom": "21.0.4",
     "typescript": "7.0.2",
     "vite": "8.3.5",
-    "vitest": "^5.0.5"
+    "vitest": "5.1.0"
   }
 }
diff --git a/app/pnpm-lock.yaml b/app/pnpm-lock.yaml
index f5cfb93..2c7f714 100644
--- a/app/pnpm-lock.yaml
+++ b/app/pnpm-lock.yaml
@@ -31,8 +31,8 @@ importers:
         specifier: 2.12.4
         version: 2.12.4
       '@vitest/coverage-v8':
-        specifier: 5.0.5
-        version: 5.0.5(vitest@5.0.5)
+        specifier: 5.1.0
+        version: 5.1.0(vitest@5.1.0)
       happy-dom:
         specifier: 21.0.4
         version: 21.0.4
@@ -43,8 +43,8 @@ importers:
         specifier: 8.3.5
         version: 8.3.5(@types/node@26.6.5)(esbuild@0.28.2)
       vitest:
-        specifier: ^5.0.5
-        version: 5.0.5(@types/node@26.6.5)(@vitest/coverage-v8@5.0.5)(happy-dom@21.0.4)(vite@8.3.5(@types/node@26.6.5)(esbuild@0.28.2))
+        specifier: 5.1.0
+        version: 5.1.0(@types/node@26.6.5)(@vitest/coverage-v8@5.1.0)(happy-dom@21.0.4)(vite@8.3.5(@types/node@26.6.5)(esbuild@0.28.2))
 
 packages:
 
@@ -635,25 +635,25 @@ packages:
     cpu: [x64]
     os: [win32]
 
-  '@vitest/coverage-v8@5.0.5':
-    resolution: {integrity: sha512-4KpkGdIrSqBihx/JyP/e+S+IKB98etg1XDQiLAxksLfcShnwi1VReGeOT9NI43hLHbknosKp/1CCLn9Zgcs8mg==}
+  '@vitest/coverage-v8@5.1.0':
+    resolution: {integrity: sha512-OY9aShBQG5KOIw2zCo9fDoCHI9RHliNHJ//nM7m6jGg41Hyke9JOGWmeMlK74xYErsxMoK8mIB3gAeEyFVA7LA==}
     peerDependencies:
-      '@vitest/browser': 5.0.5
-      vitest: 5.0.5
+      '@vitest/browser': 5.1.0
+      vitest: 5.1.0
     peerDependenciesMeta:
       '@vitest/browser':
         optional: true
 
-  '@vitest/istanbul-lib-coverage@1.0.2':
-    resolution: {integrity: sha512-9J/JMwOf9AoJhAywhrn7ScKTL38hsWQP/qPG60OtaAFcQ5OXPwKsxZFlbnuCKmZ61m8/lGgHYnFpdyQZUvG/iA==}
+  '@vitest/istanbul-lib-coverage@2.1.0':
+    resolution: {integrity: sha512-WOjuS+xcjSIy7ruYBQ+9BUcV7XZQ0FyVcbWbWxW9TRx/kLGx5SQaeFRIFDJxjNplb6YsT7SpKo2zuXXUo91OIQ==}
     engines: {node: '>=22'}
 
-  '@vitest/istanbul-lib-report@1.0.2':
-    resolution: {integrity: sha512-gUsfXZJbzPamoIY5TvHFiMMoXESBrUMo+xqaj+rYrWI69+EnvRlYBlP96ZnHPY4vX8kyUpgAnFUCg5wZG/HkDQ==}
+  '@vitest/istanbul-lib-report@2.1.0':
+    resolution: {integrity: sha512-+ZFta25nWtGCOBG9/IET8w1RO5Oawsd6H8Uf6AJ9EA/StFuKVtKAEJEJ6uRt84OH+T4YkI0ASmIHAhRA+0KVxw==}
     engines: {node: '>=22'}
 
-  '@vitest/mocker@5.0.5':
-    resolution: {integrity: sha512-xLL3lkDegJZw+MGoGaCCsFayDZthGgpAtr3Vziyfc4p7xH89K7KRPcRBBh/kEYm0Aqkt5DtTmkYpBtRV/FrHZg==}
+  '@vitest/mocker@5.1.0':
+    resolution: {integrity: sha512-5umB9/jlWcqVS6fsrFmv5Gdnpu6wAUio/lEOwAmvB7v+lFo/IW2B9FhAD0p19GCULTgMzLCdUmuCDCnAzWgppQ==}
     peerDependencies:
       msw: ^2.4.9 || ^3.0.0
       vite: ^6.0.0 || ^7.0.0 || ^8.0.0
@@ -663,8 +663,8 @@ packages:
       vite:
         optional: true
 
-  '@vitest/spy@5.0.5':
-    resolution: {integrity: sha512-lwJRExuH417257Q5wv9QYugN9IMJkoxcrKrYxGPQkBzl7wlNpTUxD/5yNX4ESKVUIShKPgkGchYPFMWOOk0wzw==}
+  '@vitest/spy@5.1.0':
+    resolution: {integrity: sha512-2DPO5OMHWKrBngC+fCmTQqemNF0gidRhp3lxbBmRhW+M1XJrfgPpqNqWVevXtn2f2ufD5ZaR1B6B641aAlcX5A==}
 
   assertion-error@2.0.1:
     resolution: {integrity: sha512-Izi8RQcffqCeNVgFigKli1ssklIbpHnCYc6AknXGYoB6grJqyeby7jv12JUQgmTAnIDnbck1uxksT4dzN3PWBA==}
@@ -904,20 +904,20 @@ packages:
       yaml:
         optional: true
 
-  vitest@5.0.5:
-    resolution: {integrity: sha512-cPODOMF8YnkiiYtDdHo6n4wWmBrOBBbOIYH2d33M4l3CdNn72h+qMmO6uXffUN6A5AZFXNND/d+krS9mgkXRfg==}
+  vitest@5.1.0:
+    resolution: {integrity: sha512-MzmNFnrwa8v8lF1bLxw3UJ4EREwVOVb75VqJZmUDRjJ6EluK/x9th7qTbaeDk41WxE1GxWqmTXcdsfobfH5sOA==}
     engines: {node: ^22.12.0 || ^24.0.0 || >=26.0.0}
     hasBin: true
     peerDependencies:
       '@edge-runtime/vm': '*'
       '@opentelemetry/api': ^1.9.0
       '@types/node': ^22.0.0 || >=24.0.0
-      '@vitest/browser-playwright': 5.0.5
-      '@vitest/browser-preview': 5.0.5
+      '@vitest/browser-playwright': 5.1.0
+      '@vitest/browser-preview': 5.1.0
       '@vitest/browser-webdriverio': ^5.0.0-beta.5 || >=5.0.0
-      '@vitest/coverage-istanbul': 5.0.5
-      '@vitest/coverage-v8': 5.0.5
-      '@vitest/ui': 5.0.5
+      '@vitest/coverage-istanbul': 5.1.0
+      '@vitest/coverage-v8': 5.1.0
+      '@vitest/ui': 5.1.0
       happy-dom: '*'
       jsdom: '*'
       vite: ^6.4.0 || ^7.0.0 || ^8.0.0
@@ -1298,34 +1298,34 @@ snapshots:
   '@typescript/typescript-win32-x64@7.0.2':
     optional: true
 
-  '@vitest/coverage-v8@5.0.5(vitest@5.0.5)':
+  '@vitest/coverage-v8@5.1.0(vitest@5.1.0)':
     dependencies:
       '@bcoe/v8-coverage': 1.0.2
-      '@vitest/istanbul-lib-coverage': 1.0.2
-      '@vitest/istanbul-lib-report': 1.0.2
+      '@vitest/istanbul-lib-coverage': 2.1.0
+      '@vitest/istanbul-lib-report': 2.1.0
       ast-v8-to-istanbul: 1.0.7
       magicast: 0.5.5
       obug: 3.0.0
       std-env: 4.3.0
       tinyrainbow: 3.2.0
-      vitest: 5.0.5(@types/node@26.6.5)(@vitest/coverage-v8@5.0.5)(happy-dom@21.0.4)(vite@8.3.5(@types/node@26.6.5)(esbuild@0.28.2))
+      vitest: 5.1.0(@types/node@26.6.5)(@vitest/coverage-v8@5.1.0)(happy-dom@21.0.4)(vite@8.3.5(@types/node@26.6.5)(esbuild@0.28.2))
 
-  '@vitest/istanbul-lib-coverage@1.0.2': {}
+  '@vitest/istanbul-lib-coverage@2.1.0': {}
 
-  '@vitest/istanbul-lib-report@1.0.2':
+  '@vitest/istanbul-lib-report@2.1.0':
     dependencies:
-      '@vitest/istanbul-lib-coverage': 1.0.2
+      '@vitest/istanbul-lib-coverage': 2.1.0
 
-  '@vitest/mocker@5.0.5(vite@8.3.5(@types/node@26.6.5)(esbuild@0.28.2))':
+  '@vitest/mocker@5.1.0(vite@8.3.5(@types/node@26.6.5)(esbuild@0.28.2))':
     dependencies:
       '@jridgewell/trace-mapping': 0.3.31
-      '@vitest/spy': 5.0.5
+      '@vitest/spy': 5.1.0
       estree-walker: 3.0.3
       magic-string: 1.4.5
     optionalDependencies:
       vite: 8.3.5(@types/node@26.6.5)(esbuild@0.28.2)
 
-  '@vitest/spy@5.0.5': {}
+  '@vitest/spy@5.1.0': {}
 
   assertion-error@2.0.1: {}
 
@@ -1551,10 +1551,10 @@ snapshots:
       esbuild: 0.28.2
       fsevents: 2.3.3
 
-  vitest@5.0.5(@types/node@26.6.5)(@vitest/coverage-v8@5.0.5)(happy-dom@21.0.4)(vite@8.3.5(@types/node@26.6.5)(esbuild@0.28.2)):
+  vitest@5.1.0(@types/node@26.6.5)(@vitest/coverage-v8@5.1.0)(happy-dom@21.0.4)(vite@8.3.5(@types/node@26.6.5)(esbuild@0.28.2)):
     dependencies:
       '@types/chai': 5.2.3
-      '@vitest/mocker': 5.0.5(vite@8.3.5(@types/node@26.6.5)(esbuild@0.28.2))
+      '@vitest/mocker': 5.1.0(vite@8.3.5(@types/node@26.6.5)(esbuild@0.28.2))
       chai: 6.3.0
       es-module-lexer: 2.3.2
       expect-type: 1.4.0
@@ -1569,7 +1569,7 @@ snapshots:
       why-is-node-running: 3.2.1
     optionalDependencies:
       '@types/node': 26.6.5
-      '@vitest/coverage-v8': 5.0.5(vitest@5.0.5)
+      '@vitest/coverage-v8': 5.1.0(vitest@5.1.0)
       happy-dom: 21.0.4
     transitivePeerDependencies:
       - msw
diff --git a/app/src-tauri/src/report_cards.rs b/app/src-tauri/src/report_cards.rs
index c36c15f..b37f922 100644
--- a/app/src-tauri/src/report_cards.rs
+++ b/app/src-tauri/src/report_cards.rs
@@ -111,6 +111,9 @@ pub struct Env {
     /// The `git` command, if there is one. Only used to look that a release tag is there before the
     /// plugin is moved to it; without git that look is skipped.
     pub git: Option<PathBuf>,
+    /// `$CODEX_HOME`, as the process had it (read once, in [`Env::folders`]). Tests leave it `None`
+    /// so they never reach the real Codex folder.
+    pub codex_home: Option<std::ffi::OsString>,
 }
 
 impl Env {
@@ -138,7 +141,17 @@ impl Env {
         let windows = cfg!(windows);
         let git_bash = if windows { platform::git_bash_here() } else { None };
         let exe = std::env::current_exe().ok();
-        Ok(Env { home, data, claude: None, codex: None, exe, windows, git_bash, git: None })
+        Ok(Env {
+            home,
+            data,
+            claude: None,
+            codex: None,
+            exe,
+            windows,
+            git_bash,
+            git: None,
+            codex_home: std::env::var_os("CODEX_HOME"),
+        })
     }
 
     /// The file whose presence turns tips on (the session-start hook looks for it).
diff --git a/app/src-tauri/src/report_cards/codex.rs b/app/src-tauri/src/report_cards/codex.rs
index 2395380..da21714 100644
--- a/app/src-tauri/src/report_cards/codex.rs
+++ b/app/src-tauri/src/report_cards/codex.rs
@@ -51,7 +51,7 @@ fn shim_file(env: &Env) -> PathBuf {
 
 /// Codex's own folder: `$CODEX_HOME`, else `~/.codex`. A variable that isn't an absolute path counts as unset.
 pub(super) fn codex_dir(env: &Env) -> PathBuf {
-    codex_dir_in(std::env::var_os("CODEX_HOME"), &env.home)
+    codex_dir_in(env.codex_home.clone(), &env.home)
 }
 
 fn codex_dir_in(var: Option<std::ffi::OsString>, home: &Path) -> PathBuf {
diff --git a/app/src-tauri/src/report_cards/tests.rs b/app/src-tauri/src/report_cards/tests.rs
index 0be7729..8087f95 100644
--- a/app/src-tauri/src/report_cards/tests.rs
+++ b/app/src-tauri/src/report_cards/tests.rs
@@ -85,7 +85,17 @@ impl Lab {
         fs::create_dir_all(home.join(".codex")).unwrap();
         fs::create_dir_all(&data).unwrap();
         Lab {
-            env: Env { home, data, claude: None, codex: None, exe: None, windows: false, git_bash: None, git: None },
+            env: Env {
+                home,
+                data,
+                claude: None,
+                codex: None,
+                exe: None,
+                windows: false,
+                git_bash: None,
+                git: None,
+                codex_home: None,
+            },
             dir,
         }
     }
@@ -154,6 +164,7 @@ impl Lab {
             windows: e.windows,
             git_bash: e.git_bash.clone(),
             git: e.git.clone(),
+            codex_home: e.codex_home.clone(),
         }
     }
 
diff --git a/app/src-tauri/tauri.conf.json b/app/src-tauri/tauri.conf.json
index 35deb2e..b4911d4 100644
--- a/app/src-tauri/tauri.conf.json
+++ b/app/src-tauri/tauri.conf.json
@@ -1,7 +1,7 @@
 {
   "$schema": "https://schema.tauri.app/config/2",
   "productName": "No, Actually...",
-  "version": "0.5.6",
+  "version": "0.5.7",
   "identifier": "app.noactually",
   "build": {
     "beforeDevCommand": "pnpm dev",
diff --git a/app/src/learn/lessons.json b/app/src/learn/lessons.json
index c5c9d75..78bc57d 100644
--- a/app/src/learn/lessons.json
+++ b/app/src/learn/lessons.json
@@ -1,7 +1,7 @@
 {
   "signal_hints": {
     "spot": "Target: name the file, function or page. 'In checkout.ts, the total is wrong.'",
-    "proof": "Evidence: paste the error, the output or an example. 'It says: Card declined.'",
+    "proof": "Evidence: paste the error, the output or an example. 'It says: TypeError: total is undefined.'",
     "done": "Finish line: say what done looks like. 'Done when the tests pass.'",
     "limits": "Guardrails: say what to leave alone. 'Don't change the API.'",
     "why": "Purpose: say who it's for or what it's part of. 'Our customers need this before tax season.'"
@@ -30,7 +30,7 @@
           "tips": [
             "Use the real name of the file, the function or the doc.",
             "Wrap exact names in backticks so they stand out.",
-            "If it's on a screen, say which page and which button."
+            "If it's on a screen, paste the page's link and name the button."
           ],
           "quiz": [
             {
@@ -64,7 +64,7 @@
             "text": "Checkout fails when I click Pay. The console says: TypeError: total is undefined. Please fix it in checkout.ts.",
             "signals": ["spot", "proof"]
           },
-          "template": "When I <do this>, I get <the exact error or output>. Here it is: <paste it>.",
+          "template": "When I <do this>, I get this: \"<paste the exact error or output>\".",
           "tips": [
             "Paste the exact error. Don't retell it from memory.",
             "Show what you got and what you expected.",
@@ -215,8 +215,8 @@
           "template": "<the task>. I already know <what you found>. I already tried <what didn't work>.",
           "tips": [
             "Say what you already tried and ruled out, and what happened, so nobody checks it twice.",
-            "Share facts the code can't show: what changed, and when it started.",
-            "Say facts about your setup and taste: “it's a Squarespace site”, “we use Tailwind”."
+            "The checker can't spot this one, so it won't light up. Say it anyway.",
+            "Say facts about your setup and taste: 'it's a small online shop', 'we use plain CSS'."
           ],
           "quiz": [
             {
@@ -321,7 +321,7 @@
           "title": "Plan big tasks first",
           "reviewed": "2026-10",
           "teaches": { "steer": "redo" },
-          "why": "A plan you can read in a minute corrections a wrong turn before hours of work get thrown away.",
+          "why": "A plan you can read in a minute catches a wrong turn before hours of work get thrown away.",
           "before": {
             "text": "Move the app from the old payment system to the new one.",
             "signals": []
diff --git a/app/src/test/fixtures.ts b/app/src/test/fixtures.ts
index 09798f1..0ca5da5 100644
--- a/app/src/test/fixtures.ts
+++ b/app/src/test/fixtures.ts
@@ -1,4 +1,5 @@
 // Shared test data, shaped like the Rust side's output.
+import { expect } from "vitest";
 import type { HallEntry, Meta, ModelRow, Overall, Progress, Prompts, RecentReportCard, ReportCardWeek, SourceInfo, Stats } from "../api";
 
 export const meta: Meta = {
@@ -87,6 +88,13 @@ export const prompts = (over: Partial<Prompts> = {}): Prompts => ({
 
 export const HOSTILE = `<img src=x onerror="alert(1)">&'"`;
 
+/** Hostile markup must come out escaped: not live, and not silently dropped either. */
+export const safe = (markup: string) => {
+  expect(markup).not.toContain("<img");
+  expect(markup).not.toContain('onerror="');
+  expect(markup).toContain("&lt;img");
+};
+
 export const row = (provider: string, family: string, version: string, corrections: number, messages = 100, tasks = 20): ModelRow => ({
   provider,
   family,
diff --git a/app/src/views/charts.test.ts b/app/src/views/charts.test.ts
index b3d2d0c..995d3eb 100644
--- a/app/src/views/charts.test.ts
+++ b/app/src/views/charts.test.ts
@@ -1,11 +1,7 @@
 import { describe, expect, it } from "vitest";
-import { HOSTILE } from "../test/fixtures";
+import { HOSTILE, safe } from "../test/fixtures";
 import { chartBox, lineChart, type Point, sparkline, stackChart } from "./charts";
 
-const safe = (markup: string) => {
-  expect(markup).not.toContain("<img");
-  expect(markup).not.toContain('onerror="');
-};
 const count = (m: string, needle: string) => m.split(needle).length - 1;
 
 const line = (points: Point[], over: Partial<Parameters<typeof lineChart>[0]> = {}) =>
diff --git a/app/src/views/improve.test.ts b/app/src/views/improve.test.ts
index 6e6a618..4c77e23 100644
--- a/app/src/views/improve.test.ts
+++ b/app/src/views/improve.test.ts
@@ -1,13 +1,9 @@
 import { describe, expect, it } from "vitest";
 import type { Goal, GoalStatus, GoalWeek, HistoryWeek, Progress, ReportCardWeek, SourceInfo } from "../api";
 import { fmtDay } from "../format";
-import { cards, HOSTILE, meta, noCards, progress, recentCard, source, stats, week } from "../test/fixtures";
+import { cards, HOSTILE, meta, noCards, progress, recentCard, safe, source, stats, week } from "../test/fixtures";
 import { buckets, canGrade, goalTitle, improveView } from "./improve";
 
-const safe = (markup: string) => {
-  expect(markup).not.toContain("<img");
-  expect(markup).not.toContain('onerror="');
-};
 const count = (m: string, needle: string) => m.split(needle).length - 1;
 
 /** `n` Mondays starting 2026-08-03, each with `tasks` tasks, and the report cards `report(i)` gives week `i`. */
diff --git a/app/src/views/learn.test.ts b/app/src/views/learn.test.ts
index 0e791e2..183f53a 100644
--- a/app/src/views/learn.test.ts
+++ b/app/src/views/learn.test.ts
@@ -3,18 +3,13 @@ import { beforeEach, describe, expect, it } from "vitest";
 import type { Check } from "../api";
 import { last } from "../format";
 import { LESSONS, TRACKS } from "../learn/lessons";
-import { HOSTILE, meta, prompts } from "../test/fixtures";
+import { HOSTILE, meta, prompts, safe } from "../test/fixtures";
 import { checkResultView, judgeQuiz, type LearnContext, learnView } from "./learn";
 
 const ctx = (over: Partial<LearnContext> = {}): LearnContext => ({ meta, prompts: null, done: [], lesson: null, draft: "", demo: false, ...over });
 const home = (over: Partial<LearnContext> = {}) => learnView(ctx(over)).value;
 const lesson = (id = "point-at-it", over: Partial<LearnContext> = {}) => learnView(ctx({ lesson: id, ...over })).value;
 
-const safe = (markup: string) => {
-  expect(markup).not.toContain("<img");
-  expect(markup).not.toContain('onerror="');
-};
-
 describe("learnView home", () => {
   it("lists every track with its lessons and the lesson count", () => {
     const m = home();
@@ -113,7 +108,7 @@ describe("learnView lesson", () => {
     const m = lesson("say-what-you-know");
     expect(m).toContain('<h1 class="q">Context</h1>');
     expect(m).toContain("I already tried");
-    expect(m).toContain("Say facts about your setup and taste: “it&#39;s a Squarespace site”, “we use Tailwind”.");
+    expect(m).toContain("Say facts about your setup and taste: &#39;it&#39;s a small online shop&#39;, &#39;we use plain CSS&#39;.");
     expect(m).toContain('<span class="chip sig-on">Target</span>'); // only the signals the checker can see are chipped
     expect(m.split("<fieldset").length - 1).toBe(2);
   });
diff --git a/app/src/views/views.test.ts b/app/src/views/views.test.ts
index 1c96619..7e08dc3 100644
--- a/app/src/views/views.test.ts
+++ b/app/src/views/views.test.ts
@@ -1,6 +1,6 @@
 import { describe, expect, it } from "vitest";
 import { emptyReportCards } from "../state";
-import { HOSTILE, hallEntry, meta, overall, row, source, stats } from "../test/fixtures";
+import { HOSTILE, hallEntry, meta, overall, row, safe, source, stats } from "../test/fixtures";
 import { headerView, statusPillView } from "./chrome";
 import { emptyGlanceView, glanceErrorView, glanceRebuildingView, glanceView } from "./glance";
 import { chartSummary, chartView, hallView, noChatsView, onboardingView, rebuildingView, scorecardView } from "./scorecard";
@@ -8,11 +8,6 @@ import { goneByCount, parseRetention, settingsView, tightens } from "./settings"
 import { allowArgs, sourcesView, watchingNames } from "./sources";
 
 const ctx = { meta, filter: {}, still: false };
-/** No hostile markup made it through as markup. */
-const safe = (markup: string) => {
-  expect(markup).not.toContain("<img");
-  expect(markup).not.toContain('onerror="');
-};
 
 describe("verdict text for screen readers", () => {
   it("matches the Hmm. on screen", () => {
diff --git a/app/vitest.config.ts b/app/vitest.config.ts
index ab2229e..f5bedba 100644
--- a/app/vitest.config.ts
+++ b/app/vitest.config.ts
@@ -6,7 +6,7 @@ export default defineConfig({
     include: ["src/**/*.test.ts"],
     environment: "node",
     // Dates and numbers format by the machine's locale and timezone; pin both so a run is the same anywhere.
-    env: { TZ: "UTC", LC_ALL: "en_US.UTF-8", LANG: "en_US.UTF-8" },
+    env: { TZ: process.env.TZ || "UTC", LC_ALL: "en_US.UTF-8", LANG: "en_US.UTF-8" },
     // Tests that read the stylesheet's own text (its colors) get it; every other CSS file is left empty.
     css: { include: [/src\/style\.css/] },
     restoreMocks: true,
diff --git a/crates/cli/src/helper/tests.rs b/crates/cli/src/helper/tests.rs
index 4f19808..6745adc 100644
--- a/crates/cli/src/helper/tests.rs
+++ b/crates/cli/src/helper/tests.rs
@@ -3,7 +3,7 @@
 use super::chat::Look;
 use super::*;
 use serde_json::{json, Value};
-use sources::testing::line;
+use sources::testing::{cc_prompt, cc_reply_with_usage, line};
 use std::time::Instant;
 
 const MODEL: &str = "claude-opus-5-5";
@@ -14,8 +14,7 @@ fn at(minute: u32) -> String {
 
 /// A message the person typed, as Claude Code records it. `n` names it: uuid `u-n`, prompt `p-n`.
 fn user(n: &str, minute: u32, text: &str) -> String {
-    line(&json!({"type":"user","uuid":format!("u-{n}"),"promptId":format!("p-{n}"),"timestamp":at(minute),
-        "message":{"content":text}}))
+    cc_prompt(&format!("u-{n}"), &format!("p-{n}"), &at(minute), text)
 }
 
 /// Tokens as (new input, cache write, cache read, output).
@@ -29,10 +28,14 @@ fn reply_from(
     text: &str,
     (input, cache_write, cache_read, output): Tokens,
 ) -> String {
-    line(&json!({"type":"assistant","uuid":format!("a-{n}"),"timestamp":at(minute),"message":{
-        "id":format!("m-{n}"),"model":model,"content":[{"type":"text","text":text}],
-        "usage":{"input_tokens":input,"cache_creation_input_tokens":cache_write,"cache_read_input_tokens":cache_read,
-            "output_tokens":output}}}))
+    cc_reply_with_usage(
+        model,
+        &format!("a-{n}"),
+        &format!("m-{n}"),
+        &at(minute),
+        text,
+        (input, cache_write, cache_read, output),
+    )
 }
 
 fn reply(n: &str, minute: u32, text: &str, tokens: Tokens) -> String {
@@ -789,7 +792,10 @@ fn the_codex_prompt_hook_reads_the_end_of_a_long_rollout_not_all_of_it() {
     let t = Instant::now();
     let out = handle(&Request::Hook(Hook::CodexPrompt), &fx.env, || input.to_string()).unwrap();
     let took = t.elapsed();
-    assert!(took.as_millis() < 2000, "a digest must come inside the hook's budget: {took:?}");
+    assert!(
+        took.as_millis() < 20_000,
+        "a digest must come inside the hook's budget (10x the real 2s, for debug builds and busy machines): {took:?}"
+    );
     let out: Value = serde_json::from_str(&out).unwrap();
     let digest = out["hookSpecificOutput"]["additionalContext"].as_str().unwrap();
     assert!(digest.contains("\nshowing the last 2 prompts\n"), "{digest}");
@@ -1207,14 +1213,14 @@ fn a_huge_transcript_is_not_read_to_say_nothing() {
     let message = tail::last_claude_message(&found, Some("p-9")).unwrap();
     let looked_at = t.elapsed();
     assert_eq!(message, "yes");
-    assert!(looked_at.as_millis() < 500, "the end of {size} bytes, not all of them: {looked_at:?}");
+    assert!(looked_at.as_millis() < 5_000, "the end of {size} bytes, not all of them: {looked_at:?}");
 
     let t = Instant::now();
     assert_eq!(fx.stop(&path, "9"), None);
     let whole = t.elapsed();
-    // The bound is loose on purpose: a debug build compiles the rules' patterns here, which a
-    // release build does in milliseconds. Reading all 50 MB would take far longer.
-    assert!(whole.as_secs() < 5, "{whole:?}");
+    // The bound is loose on purpose (10x): a debug build compiles the rules' patterns here, which a
+    // release build does in milliseconds. Reading all 50 MB would still take far longer.
+    assert!(whole.as_secs() < 50, "{whole:?}");
 }
 
 // ---- long chats ----
@@ -1259,10 +1265,10 @@ fn a_slash_command_is_ruled_out_in_one_look_at_the_end_of_a_big_file() {
     let message = tail::last_claude_message(&found, Some("p-c"));
     let took = t.elapsed();
     assert_eq!(message, None, "a command is not the person's message");
-    assert!(took.as_millis() < 500, "one look at the first window: {took:?}");
+    assert!(took.as_millis() < 5_000, "one look at the first window: {took:?}");
     let t = Instant::now();
     assert_eq!(fx.stop(&path, "c"), None);
-    assert!(t.elapsed().as_secs() < 2, "{:?}", t.elapsed());
+    assert!(t.elapsed().as_secs() < 20, "{:?}", t.elapsed());
 
     // The same when a long reply pushes the command out of the first window: wider, but only to find it.
     let path = write_long(
@@ -1327,7 +1333,7 @@ fn a_sign_off_in_a_100_mb_transcript_is_decided_from_the_end() {
     let t = Instant::now();
     assert_eq!(fx.stop(&path, "7"), Some(note("Want a report card on that? Type /grade:me")));
     let took = t.elapsed();
-    assert!(took.as_millis() < 2000, "a sign-off must decide inside the hook's budget: {took:?}");
+    assert!(took.as_millis() < 20_000, "a sign-off must decide inside the hook's budget: {took:?}");
 }
 
 #[test]
@@ -1362,7 +1368,7 @@ fn a_task_longer_than_the_last_8_mb_is_left_alone_and_nothing_more_is_read() {
     let path = write_long(&fx, &[(&first(), 20), (&begin, 12), (&end, 0)]);
     let t = Instant::now();
     assert_eq!(fx.stop(&path, "7"), None, "it can't tell what the task was without more than the limit");
-    assert!(t.elapsed().as_secs() < 5, "{:?}", t.elapsed());
+    assert!(t.elapsed().as_secs() < 50, "{:?}", t.elapsed());
 
     // What it could read is the last 8 MB: whole lines, the end of the chat, no task start.
     let from = tail_start(&path, 8 * 1024 * 1024).unwrap();
diff --git a/crates/cli/tests/helper.rs b/crates/cli/tests/helper.rs
index d6f697d..a796496 100644
--- a/crates/cli/tests/helper.rs
+++ b/crates/cli/tests/helper.rs
@@ -2,7 +2,7 @@
 //! transcript, an empty home folder and a throwaway database. Nothing real is read or written.
 
 use serde_json::{json, Value};
-use sources::testing::line;
+use sources::testing::{cc_prompt, cc_reply_with_usage};
 use std::io::Write;
 use std::path::{Path, PathBuf};
 use std::process::{Command, Output, Stdio};
@@ -12,14 +12,18 @@ fn at(minute: u32) -> String {
 }
 
 fn user(n: u32, minute: u32, text: &str) -> String {
-    line(&json!({"type":"user","uuid":format!("u-{n}"),"promptId":format!("p-{n}"),"timestamp":at(minute),
-        "message":{"content":text}}))
+    cc_prompt(&format!("u-{n}"), &format!("p-{n}"), &at(minute), text)
 }
 
 fn reply(n: u32, minute: u32, text: &str) -> String {
-    line(&json!({"type":"assistant","uuid":format!("a-{n}"),"timestamp":at(minute),"message":{
-        "id":format!("m-{n}"),"model":"claude-opus-5-5","content":[{"type":"text","text":text}],
-        "usage":{"input_tokens":10_000,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":5_000}}}))
+    cc_reply_with_usage(
+        "claude-opus-5-5",
+        &format!("a-{n}"),
+        &format!("m-{n}"),
+        &at(minute),
+        text,
+        (10_000, 0, 0, 5_000),
+    )
 }
 
 /// Two rounds, then thanks.
@@ -52,20 +56,34 @@ impl Home {
         p
     }
 
+    /// The binary with an empty environment but for this home, so no real Claude Code, Codex or
+    /// database folder (`CLAUDE_CONFIG_DIR`, `CODEX_HOME`, `XDG_*`, `APPDATA`...) is ever reached.
+    fn command(&self) -> Command {
+        let mut cmd = Command::new(env!("CARGO_BIN_EXE_noactually"));
+        cmd.env_clear()
+            .env("HOME", self.dir.path())
+            .env("USERPROFILE", self.dir.path())
+            .env("NOACTUALLY_DB", self.dir.path().join("noactually.db"));
+        // Windows can't start a program without PATH and SystemRoot, and TZ lets CI run these in other zones.
+        // The rest of the real environment stays out.
+        for keep in ["PATH", "SystemRoot", "TZ"] {
+            if let Some(v) = std::env::var_os(keep) {
+                cmd.env(keep, v);
+            }
+        }
+        cmd
+    }
+
     fn run(&self, args: &[&str], stdin: &str) -> Output {
         self.run_with(args, stdin, &[])
     }
 
     fn run_with(&self, args: &[&str], stdin: &str, env: &[(&str, &str)]) -> Output {
-        let mut child = Command::new(env!("CARGO_BIN_EXE_noactually"))
+        let mut child = self
+            .command()
             .args(args)
-            .env("HOME", self.dir.path())
-            .env("USERPROFILE", self.dir.path())
-            .env("NOACTUALLY_DB", self.dir.path().join("noactually.db"))
             .env("NOACTUALLY_CLAUDE_PROJECTS", self.dir.path().join("projects"))
             .env("NOACTUALLY_CODEX_SESSIONS", self.dir.path().join("sessions"))
-            .env_remove("CLAUDE_CODE_ENTRYPOINT")
-            .env_remove("NOACTUALLY_DEBUG")
             .envs(env.iter().copied())
             .stdin(Stdio::piped())
             .stdout(Stdio::piped())
@@ -129,15 +147,7 @@ fn an_argument_that_is_not_text_is_an_error_or_nothing_never_a_panic() {
     use std::os::unix::ffi::OsStringExt;
     let home = Home::new();
     let bad = std::ffi::OsString::from_vec(vec![b'x', 0xff, 0xfe]);
-    let run = |args: &[&std::ffi::OsStr]| {
-        Command::new(env!("CARGO_BIN_EXE_noactually"))
-            .args(args)
-            .env("HOME", home.dir.path())
-            .env("NOACTUALLY_DB", home.dir.path().join("noactually.db"))
-            .stdin(Stdio::null())
-            .output()
-            .unwrap()
-    };
+    let run = |args: &[&std::ffi::OsStr]| home.command().args(args).stdin(Stdio::null()).output().unwrap();
     // A helper command prints nothing and succeeds.
     let out = run(&["digest".as_ref(), "--transcript".as_ref(), bad.as_ref()]);
     assert!(out.status.success());
@@ -170,10 +180,9 @@ fn the_stop_hook_answers_in_json_and_only_when_it_has_something_to_say() {
     assert!(quiet.status.success() && quiet.stdout.is_empty() && quiet.stderr.is_empty());
 
     // `claude -p` has nobody to ask.
-    let mut headless = Command::new(env!("CARGO_BIN_EXE_noactually"));
+    let mut headless = home.command();
     headless
         .args(["hook", "stop"])
-        .env("NOACTUALLY_DB", home.dir.path().join("noactually.db"))
         .env("CLAUDE_CODE_ENTRYPOINT", "sdk-cli")
         .stdin(Stdio::piped())
         .stdout(Stdio::piped());
@@ -187,9 +196,8 @@ fn the_stop_hook_answers_in_json_and_only_when_it_has_something_to_say() {
 fn the_session_start_hook_gives_a_chat_in_an_app_the_rules_when_tips_are_on() {
     let home = Home::new();
     let run = |entrypoint: &str| {
-        let mut hook = Command::new(env!("CARGO_BIN_EXE_noactually"));
+        let mut hook = home.command();
         hook.args(["hook", "session-start"])
-            .env("NOACTUALLY_DB", home.dir.path().join("noactually.db"))
             .env("CLAUDE_CODE_ENTRYPOINT", entrypoint)
             .stdin(Stdio::piped())
             .stdout(Stdio::piped())
@@ -347,10 +355,12 @@ fn stats_report_cards_reports_a_card_from_a_synced_chat_as_text_and_json() {
     assert!(text.find("Extra minutes") < text.find("Extra cost"), "{text}");
     assert!(!text.contains("token"), "no counts of tokens in the text: {text}");
     assert!(text.contains("  Most missed     Finish line (1 of 1 task)\n"), "{text}");
+    // The day is shown in local time, so it moves with the zone (the chat is at 10:00 UTC).
+    let day = chrono::DateTime::parse_from_rfc3339(&at(0)).unwrap().with_timezone(&chrono::Local).date_naive();
     assert!(
-        text.contains(
-            "\n  Latest\n    2026-10-01  claude-code   B  left out: finish line  about 2 extra minutes · about $0.07 at pay-as-you-go prices"
-        ),
+        text.contains(&format!(
+            "\n  Latest\n    {day}  claude-code   B  left out: finish line  about 2 extra minutes · about $0.07 at pay-as-you-go prices"
+        )),
         "{text}"
     );
     assert_eq!(String::from_utf8_lossy(&out.stderr), "");
@@ -391,17 +401,15 @@ fn stats_report_cards_reports_a_card_from_a_synced_chat_as_text_and_json() {
 }
 
 /// A hook whose input never ends (the pipe stays open, nothing is written) must still give up
-/// on its own: exit 0 with nothing printed, inside the 2 second limit plus a little room.
+/// on its own: exit 0 with nothing printed. The hook's own limit is 2 seconds; the wait here is 10x
+/// that so a slow, busy machine doesn't fail it, while a hook that never gave up still does.
 #[test]
 fn a_hook_with_stdin_left_open_exits_quietly_in_time() {
     let home = Home::new();
-    let mut child = Command::new(env!("CARGO_BIN_EXE_noactually"))
+    let mut child = home
+        .command()
         .args(["hook", "stop"])
-        .env("HOME", home.dir.path())
-        .env("USERPROFILE", home.dir.path())
-        .env("NOACTUALLY_DB", home.dir.path().join("noactually.db"))
         .env("NOACTUALLY_CLAUDE_PROJECTS", home.dir.path().join("projects"))
-        .env_remove("CLAUDE_CODE_ENTRYPOINT")
         .stdin(Stdio::piped())
         .stdout(Stdio::piped())
         .stderr(Stdio::piped())
@@ -414,9 +422,9 @@ fn a_hook_with_stdin_left_open_exits_quietly_in_time() {
         if let Some(status) = child.try_wait().unwrap() {
             break status;
         }
-        if start.elapsed() > std::time::Duration::from_secs(3) {
+        if start.elapsed() > std::time::Duration::from_secs(30) {
             let _ = child.kill();
-            panic!("the hook was still running after 3 seconds");
+            panic!("the hook was still running after 30 seconds");
         }
         std::thread::sleep(std::time::Duration::from_millis(20));
     };
diff --git a/crates/engine/src/lib.rs b/crates/engine/src/lib.rs
index 4050214..fe58605 100644
--- a/crates/engine/src/lib.rs
+++ b/crates/engine/src/lib.rs
@@ -42,6 +42,7 @@ pub use tier::{
 pub use tip::{is_tip_rules, is_tip_stop, parse_tip, strip_tip, Tip, HAD_TO_ASK, TIP_RULES, TIP_STOP};
 
 use serde::{Deserialize, Serialize};
+use std::sync::LazyLock;
 
 /// One AI message: its text and the raw model id that wrote it, when the source records one.
 #[derive(Debug, Clone, Default, Serialize, Deserialize)]
@@ -213,7 +214,12 @@ impl Verdict {
 /// 11: a chat's first message follows no AI reply, so it is never a correction and keeps its task.
 /// 12: "I didn't catch the bug" admits again, but "I didn't catch that" (didn't hear you) doesn't;
 ///     "I should have enough context now" and "my error handling" aren't admissions.
-pub const RULES_VERSION: u32 = 12;
+/// 13: polite closers ("No, that's all for now", "That's not needed") and "same thing for…" aren't
+///     pushback; "you're right" after a plain question isn't an admission; "by the way" and the
+///     prompt grader's other openers add context; a '?' in code isn't the AI asking; the AI's own
+///     "I said" or "there is no" no longer picks the category; "because", "currently", "such as"
+///     and "Next.js" no longer tick a signal on their own.
+pub const RULES_VERSION: u32 = 13;
 
 /// Tools write placeholder ids for replies no model wrote (Claude Code's "<synthetic>" for errors,
 /// interruptions and "No response requested").
@@ -246,20 +252,24 @@ pub fn classify(turn: &Turn) -> Verdict {
     let ai_admit = rules::first(&rules::AI_ADMIT, next);
     let later = if ai_admit.is_none() { said_later(turn) } else { None };
     let (ai_admit, admit_text) = match &later {
-        Some((_, text, m)) => (Some(*m), text.as_str()),
+        Some((_, text, m)) => (Some(*m), head_chars(text, AI_HEAD)),
         None => (ai_admit, next),
     };
     let user_open = if pasted { None } else { rules::first(&rules::USER_OPEN, head) };
     let user_any = if pasted { None } else { rules::first(&rules::USER_ANY, body) };
     let context = rules::first(&rules::CONTEXT, head);
-    let answering = tail_chars(prev, ANSWER_TAIL).contains('?');
+    let answering = asks_you(prev);
+    // "Is the default port 5432?" then "You're correct": it confirmed what you asked, it didn't
+    // own a mistake. "Are you sure?", "Can you fix it?" and a statement before the question still count.
+    let only_asked = rules::plain_question(t);
+    let agreed = ai_admit.is_some_and(|m| rules::YOU_WERE_RIGHT.contains(&m.rule));
 
     // A chat's first message follows no AI reply, so there is nothing for it to correct: "still
     // broken: fix the login page" is a task, and keeps it.
     let label = match (ai_admit, user_open.or(user_any)) {
         _ if turn.prev_ai.is_none() => None,
         (Some(_), Some(_)) => Some(Label::Confirmed),
-        (Some(_), None) if context.is_none() => Some(Label::Admitted),
+        (Some(_), None) if context.is_none() && !(only_asked && agreed) => Some(Label::Admitted),
         (None, _) if user_open.is_some() && context.is_none() && !answering => Some(Label::Claimed),
         _ => None,
     };
@@ -294,6 +304,23 @@ pub fn classify(turn: &Turn) -> Verdict {
     Verdict::default()
 }
 
+/// Code in a reply: fenced blocks (an unclosed fence runs to the end) and inline spans.
+static CODE: LazyLock<regex::Regex> =
+    LazyLock::new(|| regex::Regex::new(r"(?s)```.*?(?:```|$)|`[^`\n]*`").expect("code"));
+
+/// A '?' that ends a sentence: after a word, before a space or the end (closing quotes, brackets
+/// and emoji allowed). Not `read()?;`, `user?.name`, `?page=2` or `a ? b : c`.
+static ASKS: LazyLock<regex::Regex> = LazyLock::new(|| {
+    regex::Regex::new(r#"[^\s?]\?+(?:["')\]*_]|\p{So}|\p{Sk}|\x{FE0F}|\x{200D})*(?:\s|$)"#).expect("asks")
+});
+
+/// True when the AI's reply ends by asking you something, so your "no" answers it. Code becomes a
+/// word, so "Want me to run `cargo test`?" still asks.
+fn asks_you(prev: &str) -> bool {
+    let prose = CODE.replace_all(prev, "code");
+    ASKS.is_match(tail_chars(&prose, ANSWER_TAIL))
+}
+
 /// The first later message that says the person was right: its index, its normalized text, and
 /// the match. "My mistake" doesn't count here: that far in, the AI may have found it on its own.
 fn said_later(turn: &Turn) -> Option<(usize, String, rules::Match)> {
diff --git a/crates/engine/src/model.rs b/crates/engine/src/model.rs
index 188a516..2858cca 100644
--- a/crates/engine/src/model.rs
+++ b/crates/engine/src/model.rs
@@ -29,6 +29,8 @@ impl Model {
 
 static DATE_SUFFIX: LazyLock<Regex> =
     LazyLock::new(|| Regex::new(r"[-@](20\d{6}|\d{4}-\d{2}-\d{2}|latest|preview.*|exp.*)$").unwrap());
+static BEDROCK_PREFIX: LazyLock<Regex> = LazyLock::new(|| Regex::new(r"^(?:[a-z]{2,6}\.)?anthropic\.").unwrap());
+static BEDROCK_REVISION: LazyLock<Regex> = LazyLock::new(|| Regex::new(r"-v\d+(?::\d+)?$").unwrap());
 static CLAUDE: LazyLock<Regex> =
     LazyLock::new(|| Regex::new(r"^claude-(?:(\d+)(?:[-.](\d))?-)?([a-z]+)(?:-(\d+)(?:[-.](\d+))?)?").unwrap());
 static GPT: LazyLock<Regex> =
@@ -49,7 +51,10 @@ pub fn normalize_model(raw: &str) -> Option<Model> {
     // Local runners prefix ids with an org or path ("lmstudio-community/Meta-Llama-3-8B-Instruct-GGUF").
     let base = raw.rsplit('/').next().unwrap_or(raw).to_ascii_lowercase();
     let base = base.trim_start_matches("meta-");
-    let id = DATE_SUFFIX.replace(base, "");
+    // Bedrock ids: "us.anthropic.claude-sonnet-4-5-20250929-v1:0" is the plain id plus a region/vendor prefix and a revision.
+    let base = BEDROCK_PREFIX.replace(base, "");
+    let base = BEDROCK_REVISION.replace(&base, "");
+    let id = DATE_SUFFIX.replace(&base, "");
     let id = id.as_ref();
     let m =
         |provider: &str, family: String, version: String| Some(Model { provider: provider.into(), family, version });
@@ -63,6 +68,7 @@ pub fn normalize_model(raw: &str) -> Option<Model> {
             (None, None) => return m("Anthropic", format!("Claude {tier}"), String::new()),
         };
         let version = match minor {
+            Some("0") => major.to_string(), // "claude-opus-4-0" is "claude-opus-4-20250514": version 4
             Some(min) if min.len() < 3 => format!("{major}.{min}"),
             _ => major.to_string(),
         };
@@ -135,6 +141,20 @@ mod tests {
         assert_eq!(n("claude-3-opus-20240229"), t("Anthropic", "Claude Opus", "3"));
     }
 
+    #[test]
+    fn bedrock_ids_match_their_plain_forms() {
+        assert_eq!(n("us.anthropic.claude-sonnet-4-5-20250929-v1:0"), t("Anthropic", "Claude Sonnet", "4.5"));
+        assert_eq!(n("anthropic.claude-opus-5-5"), t("Anthropic", "Claude Opus", "5.5"));
+        assert_eq!(n("global.anthropic.claude-opus-4-6-v1"), t("Anthropic", "Claude Opus", "4.6"));
+        assert_eq!(n("eu.anthropic.claude-3-5-sonnet-20241022-v2:0"), t("Anthropic", "Claude Sonnet", "3.5"));
+    }
+
+    #[test]
+    fn opus_4_has_one_version_name() {
+        assert_eq!(n("claude-opus-4-0"), t("Anthropic", "Claude Opus", "4"));
+        assert_eq!(n("claude-opus-4-20250514"), t("Anthropic", "Claude Opus", "4"));
+    }
+
     #[test]
     fn openai_ids() {
         assert_eq!(n("gpt-5.1"), t("OpenAI", "GPT", "5.1"));
diff --git a/crates/engine/src/pricing.rs b/crates/engine/src/pricing.rs
index 9f9e612..ccb593b 100644
--- a/crates/engine/src/pricing.rs
+++ b/crates/engine/src/pricing.rs
@@ -104,7 +104,10 @@ pub fn price(model: &Model) -> Option<Price> {
         ("Claude Opus", "5" | "4.8" | "4.7" | "4.6" | "4.5") => (5.0, 25.0, 0.50),
         ("Claude Sonnet", "5.5" | "5") => (2.0, 10.0, 0.20),
         ("Claude Sonnet", "4.6" | "4.5" | "4") => (3.0, 15.0, 0.30),
-        // The rate for prompts up to 100K tokens; longer prompts cost 5x, so this can undercount.
+        ("Claude Opus", "4.1" | "4") => (15.0, 75.0, 1.50),
+        // The rate for requests up to 100K input tokens; larger ones cost 5x ($0.50 / $2.50). `Usage` is
+        // summed over many requests (a task, a session), so the size of each request is unknown here:
+        // we keep the low rate, and this can undercount.
         ("Claude Haiku", "5.5") => (0.10, 0.50, 0.01),
         ("Claude Haiku", "4.5") => (1.0, 5.0, 0.10),
         _ => return None,
@@ -142,6 +145,16 @@ mod tests {
         assert!((usage().cost(&p) - 0.094).abs() < 1e-9);
     }
 
+    #[test]
+    fn opus_4_and_4_1_cost_fifteen_and_seventy_five() {
+        for id in ["claude-opus-4-20250514", "claude-opus-4-0", "claude-opus-4-1-20250805", "claude-opus-4-1"] {
+            let p = price(&normalize_model(id).unwrap()).unwrap_or_else(|| panic!("{id} has no price"));
+            assert_eq!(p, Price { input: 15.0, cache_write: 18.75, cache_read: 1.50, output: 75.0 }, "{id}");
+        }
+        let bedrock = price(&normalize_model("us.anthropic.claude-opus-4-1-20250805-v1:0").unwrap());
+        assert_eq!(bedrock.map(|p| p.input), Some(15.0));
+    }
+
     #[test]
     fn minus_is_the_step_from_an_earlier_total_and_starts_over_when_it_drops() {
         let a = Usage { input: 10, cache_write: 20, cache_read: 300, output: 40 };
diff --git a/crates/engine/src/prompt.rs b/crates/engine/src/prompt.rs
index f3435cf..6c68034 100644
--- a/crates/engine/src/prompt.rs
+++ b/crates/engine/src/prompt.rs
@@ -5,7 +5,7 @@
 //! A prompt's grade counts its [`Signal`]s, never its length, so padding a prompt earns nothing.
 //! Questions aren't prompts: asking how something works is not handing over a task.
 
-use crate::rules::{first, rule, rule_unless, Rule};
+use crate::rules::{first, rule, rule_not_after, rule_unless, Rule, ADDED_FORGOT, ADDED_OPENER};
 use crate::{head_chars, normalize_quotes, Label, Turn, Verdict};
 use serde::{Deserialize, Serialize};
 use std::sync::LazyLock;
@@ -368,9 +368,11 @@ static QUESTION: LazyLock<Vec<Rule>> = LazyLock::new(|| {
 });
 
 static EXPLAIN: LazyLock<Vec<Rule>> = LazyLock::new(|| {
-    vec![rule(
+    vec![rule_unless(
         "chat.explain",
         r"^(?:explain|tell me (?:about|what|why|how)|describe|define|summari[sz]e|tldr|eli5|(?:can|could|would) you (?:explain|tell me|help me understand|walk me through|describe|summari[sz]e|clarify))\b",
+        // "Summarize the changes into CHANGELOG.md" writes something: a task, not a question.
+        r"(?:(?:can|could|would) you )?summari[sz]e\b[^.?!\n]{0,80}\b(?:into|to|in|as)\s+(?:(?:a|an|the|my|our)\s+)?(?:[\w-]+\.[a-z]\w{0,4}\b|`|(?:new\s+)?(?:file|doc|document|changelog|readme|pr|commit|comment|issue|ticket|email|message|note|notes|table|list)\b)",
     )]
 });
 
@@ -409,15 +411,10 @@ static ADDED: LazyLock<Vec<Rule>> = LazyLock::new(|| {
     vec![
         rule_unless(
             "steer.added.also",
-            &format!(
-                r"{ACK}(?:also|oh,? and|ah,? and|and also|btw|by the way|fyi|one more thing|one other thing|another thing|additionally|in addition|plus,|p\.?s\.?)(?:\W|$)"
-            ),
+            &format!(r"{ACK}(?:{ADDED_OPENER})(?:\W|$)"),
             &format!(r"{ACK}(?:also|and also|plus,)\W*{WORKFLOW}"),
         ),
-        rule(
-            "steer.added.forgot",
-            r"\bi (?:forgot|neglected) to (?:mention|say|tell|add|include|note|specify)\b|\bforgot to mention\b|\bi should have (?:mentioned|said|told|added|specified)\b|\bi didn'?t mention\b",
-        ),
+        rule("steer.added.forgot", ADDED_FORGOT),
         rule_unless(
             "steer.added.make_sure",
             &format!(
@@ -457,9 +454,11 @@ static NEW_TASK: LazyLock<Vec<Rule>> = LazyLock::new(|| {
 /// Names the place.
 static SPOT: LazyLock<Vec<Rule>> = LazyLock::new(|| {
     vec![
-        rule(
+        rule_unless(
             "prompt.spot.file",
             r"\b[\w-]+\.(?:rs|ts|tsx|js|jsx|mjs|cjs|py|go|java|kt|kts|swift|rb|php|cs|cpp|cc|c|h|hpp|css|scss|sass|less|html?|md|mdx|json|jsonl|ya?ml|toml|ini|cfg|conf|sql|sh|bash|zsh|ps1|vue|svelte|astro|lock|env|txt|csv|tsv|xml|gradle|tf|proto|graphql|gql|ipynb|plist|xcconfig|dart|lua|ex|exs|scala)\b|(?:^|[\s(`'])\.(?:env|gitignore|npmrc|nvmrc|eslintrc|prettierrc|editorconfig|github|vscode)\b|(?-i:\b(?:Dockerfile|Makefile|Justfile|justfile|Gemfile|Procfile|README|CHANGELOG)\b)",
+            // A framework's name ("Next.js", "Node.js") is a word, not a file.
+            r"(?:next|node|nuxt|nest|vue|react|preact|angular|ember|backbone|express|three|chart|d3|alpine|solid|knockout|meteor|moment|p5|pixi|leaflet|electron)\.js\b",
         ),
         rule("prompt.spot.dir", r"(?:^|[\s(\[`'])(?:~|\.{1,2})?/?[A-Za-z_][\w.-]*/[\w.-]+/"),
         rule("prompt.spot.mention", r"(?:^|\s)@[\w-]*[/.][\w./-]+"),
@@ -499,7 +498,8 @@ static PROOF: LazyLock<Vec<Rule>> = LazyLock::new(|| {
         ),
         rule(
             "prompt.proof.example",
-            r"\be\.g\.|\b(?:for example|for instance|such as|like this|example:|sample:|here'?s (?:an|the|my) (?:example|output|error|log|screenshot|trace)|here is (?:an|the|my) (?:example|output|error|log|screenshot|trace))",
+            // "such as login" lists, it doesn't show anything.
+            r"\be\.g\.|\b(?:for example|for instance|like this|example:|sample:|here'?s (?:an|the|my) (?:example|output|error|log|screenshot|trace)|here is (?:an|the|my) (?:example|output|error|log|screenshot|trace))",
         ),
         rule("prompt.proof.quote", r#""[^"\n]{6,200}""#),
         rule(
@@ -575,7 +575,7 @@ static LIMITS: LazyLock<Vec<Rule>> = LazyLock::new(|| {
         ),
         rule(
             "prompt.limits.no",
-            r"\bno (?:new |extra |additional |external |third[- ]party )?(?:dependencies|deps|packages|libraries|libs|frameworks|breaking changes|migrations|comments|emojis?|placeholders|mocks|todos?)\b|\b(?:must not|mustn'?t|should not|shouldn'?t|cannot|can'?t|may not) (?:change|break|touch|use|add|remove|affect|modify|rely|depend|exceed|call|import|be)\b|\b(?:out of scope|scope (?:it|this) to|limit(?:ed)? (?:it |this |the change )?to|stick (?:to|with)|stay within|no more than)\b",
+            r"\bno (?:new |extra |additional |external |third[- ]party )?(?:dependencies|deps|packages|libraries|libs|frameworks|breaking changes|migrations|comments|emojis?|placeholders|mocks|todos?)\b|\b(?:must not|mustn'?t|should not|shouldn'?t|cannot|can'?t|may not) (?:change|break|touch|use|add|remove|affect|modify|rely|depend|exceed|call|import)\b|\b(?:out of scope|scope (?:it|this) to|limit(?:ed)? (?:it |this |the change )?to|stick (?:to|with)|stay within|no more than)\b",
         ),
     ]
 });
@@ -583,9 +583,11 @@ static LIMITS: LazyLock<Vec<Rule>> = LazyLock::new(|| {
 /// Gives the purpose.
 static WHY: LazyLock<Vec<Rule>> = LazyLock::new(|| {
     vec![
-        rule(
+        rule_not_after(
             "prompt.why.because",
             r"\b(?:because|since (?:we|i|they|users?|our \w+) (?:need|want|can'?t|cannot|don'?t|use|have|rely|depend|ship|support)|so that|in order to|that way)\b",
+            // "it fails because the cache is stale" says what's wrong, not what the work is for.
+            r"\b(?:fails?|failing|failed|breaks|broke|broken|crash(?:es|ed|ing)?|errors? out|hangs?|freezes|times? out|stuck|slow|empty|blank|stale|missing|wrong|(?:doesn'?t|does not|didn'?t|won'?t|can'?t) (?:work|load|show|run|build|start|open|compile)|(?:isn'?t|is not|not) (?:working|loading|showing|running|building|starting))\b(?:\s+(?:again|now|anymore|sometimes|often|randomly|here|there))?[\s,]*",
         ),
         rule(
             "prompt.why.goal",
@@ -595,7 +597,12 @@ static WHY: LazyLock<Vec<Rule>> = LazyLock::new(|| {
             "prompt.why.who",
             r"\b(?:our|my) (?:users?|customers?|clients?|team|boss|manager|designer|pm|stakeholders?|visitors|readers|audience)\b|\b(?:users?|customers?|people) (?:need|want|can'?t|cannot|keep|report|complain|expect)\b|\bfor (?:the|our|a|my) (?:demo|launch|release|client|customer|presentation|meeting|deadline|portfolio|interview)\b",
         ),
-        rule("prompt.why.now", r"\b(?:currently|at the moment|as of now)\b"),
+        // "Currently we're moving to Postgres" says what it's part of; "currently the button is
+        // red" only describes it.
+        rule(
+            "prompt.why.now",
+            r"\b(?:currently|at the moment|as of now|right now),? (?:i|we|they)(?:'?re|'?m| are| am)? (?:\w+ing|about to|in the middle of)\b",
+        ),
     ]
 });
 
diff --git a/crates/engine/src/prompt_tests.rs b/crates/engine/src/prompt_tests.rs
index a6ec6ae..77fbfed 100644
--- a/crates/engine/src/prompt_tests.rs
+++ b/crates/engine/src/prompt_tests.rs
@@ -125,11 +125,25 @@ const EXAMPLES: &[(&str, &[Signal])] = &[
     ("Don't work on main, use a branch", &[Limits]),
     // Purpose.
     ("cache the results because the API is rate limited", &[Why]),
+    ("cache the results because we're launching it on Friday", &[Why]),
     ("our users can't find the export button", &[Why]),
     ("cache it since we rely on a slow API", &[Why]),
     ("I'm trying to cut the bundle size for the launch", &[Why]),
     ("Goal: one place to see every invoice", &[Why]),
-    ("currently the report takes a minute to open", &[Why]),
+    ("currently we're moving the reports to Postgres", &[Why]),
+    // Saying what's wrong or how it is now isn't saying what it's for.
+    ("the login fails because the cache is stale", &[]),
+    ("the page is blank because the query times out", &[]),
+    ("it doesn't load, because the token expired", &[]),
+    ("Currently the button is red", &[]),
+    ("currently the report takes a minute to open", &[]),
+    // A framework's name isn't a file, a list isn't an example, and "can't be opened" isn't a rule.
+    ("move the site to Next.js", &[]),
+    ("the Node.js and Vue.js versions are old", &[]),
+    ("fix App.js", &[Spot]),
+    ("add social sign-in, such as login with email", &[]),
+    ("the PDF can't be opened on Windows", &[]),
+    ("this can't break the import", &[Limits]),
     // Specific: several at once.
     (
         "In src/checkout.ts the total should include tax. Don't touch the cart API, our customers depend on it.",
@@ -226,6 +240,8 @@ fn hellos_questions_and_payloads_are_not_prompts() {
         "is there a way to skip the hook?",
         "explain the retention levels",
         "can you explain what this regex matches",
+        "summarize the retention docs",
+        "summarize the changes in 2.0",
         "<command-name>/clear</command-name>",
         "/review",
         "   ",
@@ -240,6 +256,9 @@ fn hellos_questions_and_payloads_are_not_prompts() {
         "Can I get a script that renames files?",
         "Is it possible to add dark mode?",
         "/Users/me/app/src/a.ts is broken",
+        // Summing up into a file writes something.
+        "Summarize the changes into CHANGELOG.md",
+        "could you summarize the thread into a new issue",
     ] {
         assert!(matches!(first_message(p), Prompt::Task { .. }), "{p:?}");
     }
@@ -288,6 +307,10 @@ fn steering_that_is_on_you() {
         ("Done.", "let's try a different approach", Steer::Redo),
         ("Done.", "go back to the old version", Steer::Redo),
         ("Done.", "also make it work offline", Steer::Added),
+        ("Done.", "By the way, we deploy on Vercel", Steer::Added),
+        ("Done.", "oh, also the dates are in UTC", Steer::Added),
+        ("Done.", "p.s. it has to run on Node 18", Steer::Added),
+        ("Done.", "I should have mentioned it's for mobile", Steer::Added),
     ];
     for (prev, human, want) in cases {
         assert_eq!(steer(prev, human), Some(want), "{human:?}");
diff --git a/crates/engine/src/redact.rs b/crates/engine/src/redact.rs
index 20f3570..65ee1de 100644
--- a/crates/engine/src/redact.rs
+++ b/crates/engine/src/redact.rs
@@ -5,7 +5,7 @@ use regex::{Captures, Regex};
 use std::sync::LazyLock;
 
 /// Bump when [`redact`] catches more, so text stored under older rules is cleaned once.
-pub const REDACTION_VERSION: u32 = 2;
+pub const REDACTION_VERSION: u32 = 3;
 
 /// The longest snippet kept, in chars.
 pub const SNIPPET_CHARS: usize = 140;
@@ -13,6 +13,14 @@ pub const SNIPPET_CHARS: usize = 140;
 /// A value after a secret's name: escaped-quoted (in JSON inside a string), quoted (it may hold
 /// spaces), already redacted (so a second pass changes nothing), or bare.
 const VALUE: &str = r#"(?P<v>\\"[^"]*?\\"|"[^"]*"|'[^']*'|\[secret\]|\S+)"#;
+/// A value that only counts when [`looks_secret`] agrees: quoted, or long with a digit in it.
+const SECRET_LIKE: &str = r#"(?P<s>"[^"\s]{6,}"|'[^'\s]{6,}'|[A-Za-z0-9_+/=.-]+)"#;
+
+/// True for a quoted value, or a bare one of 8 or more characters with a digit (`abc123def456`). Plain
+/// words and short ids (`id`, `users.id`, `5`) are code or prose.
+fn looks_secret(value: &str) -> bool {
+    value.starts_with(['"', '\'']) || (value.len() >= 8 && value.contains(|c: char| c.is_ascii_digit()))
+}
 
 /// What follows a home folder: more path, up to where prose or a list takes over again.
 /// A folder's name can hold a few spaces or commas when another folder follows it.
@@ -91,6 +99,7 @@ const NOT_A_VALUE: &[&str] = &[
     "please",
     "required",
     "reset",
+    "see",
     "should",
     "so",
     "some",
@@ -164,7 +173,7 @@ static REDACTIONS: LazyLock<Vec<Rule>> = LazyLock::new(|| {
         (r(r"[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Za-z]{2,}"), "[email]"),
         (
             r(
-                r"\b(sk-[A-Za-z0-9_-]{16,}|(?:sk|rk|pk)_(?:live|test)_[A-Za-z0-9]{8,}|github_pat_[A-Za-z0-9_]{20,}|gh[pousr]_[A-Za-z0-9]{20,}|glpat-[A-Za-z0-9_-]{10,}|npm_[A-Za-z0-9]{20,}|xox[abpr]-[A-Za-z0-9-]{10,}|(?:AKIA|ASIA)[0-9A-Z]{16}|hf_[A-Za-z0-9]{20,}|gsk_[A-Za-z0-9]{20,}|whsec_[A-Za-z0-9]{16,}|SG\.[A-Za-z0-9_-]{16,}\.[A-Za-z0-9_-]{16,}|AIza[0-9A-Za-z_-]{30,}|eyJ[A-Za-z0-9_-]{10,}\.[A-Za-z0-9_-]{10,}\.[A-Za-z0-9_-]{10,})",
+                r"\b(sk-[A-Za-z0-9_-]{16,}|(?:sk|rk|pk)_(?:live|test)_[A-Za-z0-9]{8,}|github_pat_[A-Za-z0-9_]{20,}|gh[pousr]_[A-Za-z0-9]{20,}|glpat-[A-Za-z0-9_-]{10,}|npm_[A-Za-z0-9]{20,}|xox[abprs]-[A-Za-z0-9-]{10,}|xapp-\d-[A-Za-z0-9-]{10,}|ya29\.[A-Za-z0-9_-]{20,}|GOCSPX-[A-Za-z0-9_-]{10,}|(?:AKIA|ASIA)[0-9A-Z]{16}|hf_[A-Za-z0-9]{20,}|gsk_[A-Za-z0-9]{20,}|whsec_[A-Za-z0-9]{16,}|SG\.[A-Za-z0-9_-]{16,}\.[A-Za-z0-9_-]{16,}|AIza[0-9A-Za-z_-]{30,}|eyJ[A-Za-z0-9_-]{10,}\.[A-Za-z0-9_-]{10,}\.[A-Za-z0-9_-]{10,})",
             ),
             "[secret]",
         ),
@@ -196,9 +205,14 @@ static REDACTIONS: LazyLock<Vec<Rule>> = LazyLock::new(|| {
             )),
             "$1=[secret]",
         ),
+        // Any other key name (OPENAI_KEY, license_key) only when the value looks like a key, so code such as
+        // `primary_key: id` is left alone.
+        (r(&format!(r#"(?i)\b([A-Za-z0-9_.-]*_key)(?:\\?["'])?\s*[:=]\s*{SECRET_LIKE}"#)), "$1=[secret]"),
         // Short names only count whole (`pw:`, `DB_PASS=`), so "bypass" and "compass" are left alone.
         (r(&format!(r#"(?i)\b((?:[A-Za-z0-9]+[_.-])*pw)(?:\\?["'])?\s*[:=]\s*{VALUE}"#)), "$1=[secret]"),
         (r(&format!(r#"(?i)\b((?:[A-Za-z0-9]+[_.-])*pass)(?:\\?["'])?\s*=\s*{VALUE}"#)), "$1=[secret]"),
+        // `pass:` is also a test count ("Pass: 5"), so it needs a value that looks like a password.
+        (r(&format!(r#"(?i)\b((?:[A-Za-z0-9]+[_.-])*pass)(?:\\?["'])?\s*:\s*{SECRET_LIKE}"#)), "$1=[secret]"),
         // "my password is hunter2" takes the next word, even when it is only "too" (better lost than leaked).
         (r(r"(?i)\b(passwd|pass(?:word|code|phrase))\s+(?:is|was)\s+\S+"), "$1 [secret]"),
         (r(&format!(r"(/Users|/home)/[^/\s]+{UNDER}")), "~/[path]"),
@@ -208,7 +222,9 @@ static REDACTIONS: LazyLock<Vec<Rule>> = LazyLock::new(|| {
         (r(r"(^|[^\w/.-])/root(?:/\S*|\b)"), "${1}~"),
         (r(r"/(private|var|tmp|Volumes|opt|mnt)/\S+"), "[path]"),
         (r(r"\b\d{1,3}(\.\d{1,3}){3}\b"), "[ip]"),
-        (r(r"\+?\d[\d\s().-]{8,}\d"), "[number]"),
+        // An ISO date ("2026-11-15") is not a phone number: the `keep` branch matches it first and leaves it be.
+        // It must end there, so a number that only starts like one ("0466-12-3456") is still caught.
+        (r(r"(?P<keep>\b\d{4}-\d{2}-\d{2})(?:[^\d-]|$)|\+?\d[\d\s().-]{8,}\d"), "[number]"),
     ];
     rules.into_iter().map(|(re, with)| Rule { re, with }).collect()
 });
@@ -218,7 +234,10 @@ pub fn redact(text: &str) -> String {
     REDACTIONS.iter().fold(text.to_string(), |acc, rule| {
         rule.re
             .replace_all(&acc, |c: &Captures<'_>| {
-                if c.name("v").is_some_and(|v| not_a_value(v.as_str())) {
+                if c.name("keep").is_some()
+                    || c.name("v").is_some_and(|v| not_a_value(v.as_str()))
+                    || c.name("s").is_some_and(|s| !looks_secret(s.as_str()))
+                {
                     return c[0].to_string();
                 }
                 let mut out = String::new();
@@ -459,6 +478,44 @@ mod tests {
         assert_eq!(redact(prose), prose);
     }
 
+    #[test]
+    fn any_key_name_pass_colon_and_more_tokens_are_caught() {
+        // Split so secret scanners don't take these made-up samples for real keys.
+        for (name, value) in
+            [("OPENAI", "abc123def4"), ("STRIPE", concat!("rk_", "live_abc123def456")), ("LICENSE", "ABCD-1234-IJKL")]
+        {
+            assert_eq!(redact(&format!("{name}_KEY={value}")), format!("{name}_KEY=[secret]"), "{name}");
+        }
+        assert_eq!(redact("license_key: 'ABCD-EFGH'"), "license_key=[secret]");
+        assert_eq!(redact("pass: swordfish99"), "pass=[secret]");
+        // Code and test output that only look like secrets.
+        for code in ["primary_key: id", "foreign_key: users.id", "sort_key=name", "Pass: 5", "pass: true, fail: 0"] {
+            assert_eq!(redact(code), code);
+        }
+        assert_eq!(redact("pass: 'hunter22'"), "pass=[secret]");
+        // Split or built at runtime so secret scanners don't take these made-up samples for real tokens.
+        let ya29 = format!("ya29.{}", "a0AfH6SMBxabcdefghij0123456789");
+        let gocspx = format!("GOCSPX-{}", "abcdefghijklmnop");
+        let xoxs = format!("xox{}", "s-123456789012-abcdefghijkl");
+        let xapp = format!("xapp-{}", "1-A0123456789-abcdefghijkl");
+        for key in [ya29, gocspx, xoxs, xapp] {
+            assert_eq!(redact(&format!("key {key} end")), "key [secret] end", "{key}");
+        }
+        // Near misses: prose after a key name, and a pass that is a verdict.
+        let prose = "the API_KEY: see docs, LICENSE_KEY is required, pass the test, a monkey=1";
+        assert_eq!(redact(prose), prose);
+    }
+
+    #[test]
+    fn iso_dates_stay_but_phone_numbers_go() {
+        assert_eq!(redact("due 2026-11-15 and 2026-11-16."), "due 2026-11-15 and 2026-11-16.");
+        assert_eq!(redact("at 2026-11-15T10:00:00Z"), "at 2026-11-15T10:00:00Z");
+        assert_eq!(redact("call 555-123-4567 on 2026-11-15"), "call [number] on 2026-11-15");
+        assert_eq!(redact("call 0466-12-3456 now"), "call [number] now");
+        assert_eq!(redact("ref 2026-11-15-1234"), "ref [number]");
+        assert_eq!(redact("card 4111 1111 1111 1111"), "card [number]");
+    }
+
     #[test]
     fn flags_with_a_space_before_the_value_are_caught() {
         assert_eq!(redact("run --password hunter2 now"), "run --password [secret] now");
diff --git a/crates/engine/src/rules.rs b/crates/engine/src/rules.rs
index bf94bce..27be4e7 100644
--- a/crates/engine/src/rules.rs
+++ b/crates/engine/src/rules.rs
@@ -51,14 +51,40 @@ pub(crate) fn rule_unless(id: &'static str, pattern: &str, unless: &str) -> Rule
     Rule { id, re: re(id, pattern), unless: Some(re(id, &format!("^(?:{unless})"))), not_after: None }
 }
 
+/// A rule that doesn't fire when the text before the match ends with `before`.
+pub(crate) fn rule_not_after(id: &'static str, pattern: &str, before: &str) -> Rule {
+    Rule { id, re: re(id, pattern), unless: None, not_after: Some(re(id, &format!("(?:{before})$"))) }
+}
+
 pub fn first(rules: &[Rule], text: &str) -> Option<Match> {
     rules.iter().find_map(|r| r.find(text))
 }
 
+/// "not necessary", "not needed": nothing left to do, not a complaint. A macro so `concat!` can
+/// build [`APPROVAL`] from it.
+macro_rules! not_needed {
+    () => {
+        r"not\s+(really\s+|strictly\s+)?(necessary|needed|required|important|urgent)\b"
+    };
+}
+
 /// "that's fine", "it works", "you're right": a subject plus an approval word. Appended to an
 /// opener like "no, " it marks approval ("No, that's fine, ship it"), not pushback. Present tense
-/// only: "No, it worked before you changed it" and "it was fine until…" are corrections.
-const APPROVAL: &str = r"(that|this|it|you)('s|'re|s|re|\s+is|\s+are|\s+looks?|\s+sounds?|\s+seems?)?(\s+(all|also|actually|just|totally|perfectly|pretty|really|now))?\s+(fine|good|great|ok|okay|perfect|right|correct|works?|working)\b";
+/// only: "No, it worked before you changed it" and "it was fine until…" are corrections. "That's
+/// not necessary" and "that's all for now" close a task politely; "that's all wrong" doesn't.
+const APPROVAL: &str = concat!(
+    r"(that|this|it|you)('s|'re|s|re|\s+is|\s+are|\s+looks?|\s+sounds?|\s+seems?)?(\s+(all|also|actually|just|totally|perfectly|pretty|really|now))?\s+((fine|good|great|ok|okay|perfect|right|correct|works?|working|alright|all right|cool)\b|",
+    not_needed!(),
+    r"|all(\s+for\s+now)?\s*([,.!;:—–-]|$|thanks\b|thank you\b))"
+);
+
+/// Openers that add what the AI couldn't have known: "also", "by the way", "p.s.". Not a bare
+/// "ps": "ps aux" and a "PS C:\>" prompt are pastes. Shared by
+/// [`CONTEXT`] and the prompt grader's "you added more", so the two never disagree.
+pub(crate) const ADDED_OPENER: &str = r"also|oh,? and|ah,? and|oh,? also|ah,? also|and also|btw|by the way|fyi|one more thing|one other thing|another thing|additionally|in addition|plus,|p\.s\.|ps[:,]";
+
+/// "I forgot to mention", "I should have said": a requirement added late, on you.
+pub(crate) const ADDED_FORGOT: &str = r"\bi (?:forgot|neglected) to (?:mention|say|tell|add|include|note|specify)\b|\bforgot to mention\b|\bi should have (?:mentioned|said|told|added|specified)\b|\bi didn'?t mention\b";
 
 /// Right before an admission, these leave it open: "let me check if you're right", "I don't think
 /// you're right", "that might be a good catch".
@@ -104,8 +130,9 @@ pub static AI_ADMIT: LazyLock<Vec<Rule>> = LazyLock::new(|| {
         rule_unless(
             "ai.should_have",
             r"\bi should have\b",
-            // What the AI expects to have, not what it failed to do: "I should have enough context now".
-            r"i should have\s+(enough|everything|all|what|which|the (context|info|information|data|access|files?|details?|answer|results?|rest)|access|a (good|clear|better|full|complete)|more (time|context|info|information|data|details?|access)|some|plenty|sufficient|both|time|permission|it (now|soon|ready|done|working|fixed)|that (now|soon))\b",
+            // What the AI expects to have, not what it failed to do: "I should have enough context
+            // now", "I should have a fix ready in a minute", "I should have this done shortly".
+            r"i should have\s+(enough|everything|all|what|which|the (context|info|information|data|access|files?|details?|answer|results?|rest)|access|a (good|clear|better|full|complete)|more (time|context|info|information|data|details?|access)|some|plenty|sufficient|both|time|permission|(it|this|that|these|those|them|(a|the|an) (fix|patch|draft|pr|change|update|version|build|plan|answer))\s+((all|fully|mostly)\s+)?(now|soon|shortly|ready|done|working|fixed|finished|sorted|up|back|in a (minute|moment|sec|second|bit|few)))\b",
         ),
         rule("ai.incorrectly", r"\bi incorrectly\b"),
         rule("ai.that_was_wrong", r"\bthat was (wrong|incorrect|a mistake)\b"),
@@ -123,7 +150,7 @@ pub static AI_ADMIT: LazyLock<Vec<Rule>> = LazyLock::new(|| {
 
 /// Credits the person, rather than only owning up. "Good catch" isn't here: that far into a reply
 /// it's mostly about a test or the AI's own check.
-const YOU_WERE_RIGHT: [&str; 2] = ["ai.youre_right", "ai.you_are_right"];
+pub(crate) const YOU_WERE_RIGHT: [&str; 2] = ["ai.youre_right", "ai.you_are_right"];
 
 /// The admissions that say the person was right, for a later message in a reply.
 pub fn you_were_right(text: &str) -> Option<Match> {
@@ -133,17 +160,26 @@ pub fn you_were_right(text: &str) -> Option<Match> {
 /// You open your message by pushing back.
 pub static USER_OPEN: LazyLock<Vec<Rule>> = LazyLock::new(|| {
     vec![
+        // "Not needed" is approval ("No, that's not necessary."), unless it says to undo the work.
+        rule(
+            "you.unneeded",
+            &format!(
+                r"^((no|nope|nah)\s*[,.!—–-]+\s*)?((that|this|it)('s|\s+is)?\s+)?{}[^\n]{{0,60}}?\b(remove|revert|undo|delete|take (it |that |this )?out|roll (it |that |this )?back|i (said|asked))\b",
+                not_needed!()
+            ),
+        ),
         rule_unless(
             "you.no_that",
             r"^(no|nope|nah)\s*[,.!—–-]+\s*(that|this|it|you|not|the|we|i said|still|wrong)\b",
-            &format!(r"(no|nope|nah)\s*[,.!—–-]+\s*{APPROVAL}"),
+            &format!(r"(no|nope|nah)\s*[,.!—–-]+\s*({APPROVAL}|{})", not_needed!()),
         ),
         rule("you.no", r"^(no|nope|nah|wrong)[.!]*$"),
         rule_unless(
             "you.wrong",
             r"^(wrong|incorrect|not quite|not true|that'?s not|that is not|that isn'?t|this is not|this isn'?t)\b",
-            // "that's not a problem", "this isn't bad": a negated complaint is approval.
-            r"(that'?s not|that is not|that isn'?t|this is not|this isn'?t)\s+(a (big )?deal|a problem|an issue|bad|wrong)(\s+at all)?\s*([,.!;:\n]|$)",
+            // "that's not a problem", "this isn't bad": a negated complaint is approval. "That's not
+            // needed" says there's nothing more to do.
+            r"(that'?s not|that is not|that isn'?t|this is not|this isn'?t)\s+((really|strictly)\s+)?(a (big )?deal|a problem|an issue|bad|wrong|necessary|needed|required|important|urgent)(\s+(at all|right now|now|yet|anymore|for now|today))?\s*([,.!;:\n]|$)",
         ),
         rule(
             "you.its_wrong",
@@ -157,8 +193,18 @@ pub static USER_OPEN: LazyLock<Vec<Rule>> = LazyLock::new(|| {
         rule("you.i_said", r"^(i said|i told you|like i said|as i said|i already (said|told|asked))\b"),
         rule("you.i_asked", r"^i asked (you )?(to|for|not)\b"),
         rule("you.why_did_you", r"^(why did you|why would you|what\??$|wtf|huh\??$)"),
-        rule("you.doesnt_work", r"^(doesn'?t|didn'?t|does not|did not|isn'?t|not) work"),
-        rule("you.same_error", r"^(same|still the same) (error|issue|problem|thing)"),
+        rule_unless(
+            "you.doesnt_work",
+            r"^(doesn'?t|didn'?t|does not|did not|isn'?t|not) work",
+            // "Not working on the API today": a plan, not a complaint. "Not working on Safari" still is one.
+            r"not working on\b[^.!?\n]*\b(today|tonight|this (week|morning|afternoon)|for now|focus)\b",
+        ),
+        rule_unless(
+            "you.same_error",
+            r"^(same|still the same) (error|issue|problem|thing)\b",
+            // "Same thing for the footer please" asks for more of the same.
+            r"(same|still the same) thing\s+(for|with|on|in|to|here|please)\b",
+        ),
         rule_unless(
             "you.actually",
             r"^actually,? (it|that|this|you|the|no)\b",
@@ -187,8 +233,12 @@ pub static USER_ANY: LazyLock<Vec<Rule>> = LazyLock::new(|| {
 /// You added information the AI couldn't have known, or you're just acknowledging. Never a correction.
 pub static CONTEXT: LazyLock<Vec<Rule>> = LazyLock::new(|| {
     vec![
-        rule("ctx.also", r"^(also|oh and|btw|fyi|one more thing|additionally|and also|plus)\b"),
-        rule("ctx.forgot_to_mention", r"\bi forgot to (mention|say|tell)\b"),
+        rule("ctx.also", &format!(r"^(?:{ADDED_OPENER})(?:\W|$)")),
+        // At the start only: "Why did you add a footer? I didn't mention a footer." is pushback.
+        rule(
+            "ctx.forgot_to_mention",
+            &format!(r"^(?:(?:oh|ah|oops|sorry|wait|also|and|ok|okay|btw)[,.!]*\s+)*(?:{ADDED_FORGOT})"),
+        ),
         rule("ctx.ack", r"^(ok|okay|great|thanks|perfect|cool|nice|yes|yep|sure)\b"),
     ]
 });
@@ -209,33 +259,69 @@ pub static AI_PUSHBACK: LazyLock<Vec<Rule>> = LazyLock::new(|| {
     ]
 });
 
+/// Doubt inside a question: "are you sure?", "isn't it 5433?". That's pushing back, so a "you're
+/// right" after it still admits a mistake; after a plain question it only confirms what you asked.
+pub static DOUBT: LazyLock<Regex> = LazyLock::new(|| {
+    re(
+        "doubt",
+        r"\breally\?|\b(are you (sure|certain|positive)|you sure|is (that|this|it) (really )?(right|correct|true|accurate)|did you (actually |really )?(check|test|verify|confirm|run|read|look)|(shouldn'?t|won'?t|can'?t|wouldn'?t|isn'?t|aren'?t|doesn'?t|don'?t|wasn'?t|didn'?t) \w+|what about|hold on|wait)\b",
+    )
+});
+
+/// A request inside a question: "can you fix that?" asks for work, not information.
+static REQUEST: LazyLock<Regex> = LazyLock::new(|| re("request", r"\b((can|could|would|will) you|please|pls|plz)\b"));
+
+/// Sentence ends before the last one.
+static SENTENCE_END: LazyLock<Regex> = LazyLock::new(|| re("sentence", r"[.!?]+\s"));
+
+/// Longer than this (in chars), a question carries more than a quick check.
+const PLAIN_QUESTION: usize = 120;
+
+/// One short question asking for information, and nothing else: "Is the default port 5432?". Not
+/// a statement then a question ("It crashes. Can you fix it?"), a request, or doubt.
+pub fn plain_question(t: &str) -> bool {
+    t.ends_with('?')
+        && t.chars().count() <= PLAIN_QUESTION
+        && !SENTENCE_END.is_match(t)
+        && !REQUEST.is_match(t)
+        && !DOUBT.is_match(t)
+}
+
 /// Short "are you done?" style questions: an AI "Not quite" there is a status, not a correction.
 pub static STATUS_Q: LazyLock<Regex> =
     LazyLock::new(|| Regex::new(r"(?i)\b(finish|done|left|still going|status|ready)\b").unwrap());
 
-static CATEGORIES: LazyLock<Vec<(Category, Regex)>> = LazyLock::new(|| {
-    let c = |cat, p: &str| (cat, Regex::new(&format!("(?i){p}")).unwrap());
+/// (category, words read in your message, words read in the AI's reply).
+type CategoryWords = (Category, Option<Regex>, Option<Regex>);
+
+/// Each category's words, per side. "I said", "you missed" and "there is no" tell your side; in the
+/// AI's reply they're about something else ("I already changed it", "there is no reason to keep
+/// it"). "I misread" is the AI's side.
+static CATEGORIES: LazyLock<Vec<CategoryWords>> = LazyLock::new(|| {
+    let re = |p: &str| (!p.is_empty()).then(|| Regex::new(&format!("(?i){p}")).unwrap());
+    let c = |cat, yours: &str, ais: &str| (cat, re(yours), re(ais));
+    let made_up = r"\b(doesn'?t|does not) exist\b|\bmade (that|it) up\b|\bhallucinat|\bno such\b|\bnot a real\b|\bthat'?s not a thing\b";
+    let missing = r"\bmissing\b|\bwhat about\b|\bleft out\b";
+    let still =
+        r"\bstill\b|\bsame (error|issue|problem)\b|\b(doesn'?t|didn'?t|not) work|\bbroken\b|\berror\b|\bcrash|\bfail";
     vec![
-        c(
-            Category::MadeItUp,
-            r"\b(doesn'?t|does not) exist\b|\bmade (that|it) up\b|\bhallucinat|\bno such\b|\bnot a real\b|\bthere is no\b|\bthat'?s not a thing\b",
-        ),
+        c(Category::MadeItUp, &format!(r"{made_up}|\bthere is no\b"), made_up),
         c(
             Category::IgnoredInstructions,
             r"\bi (said|told you|asked|already)\b|\blike i said\b|\bnot what i (asked|said)\b|\byou ignored\b|\bre-?read\b",
+            "",
         ),
         c(
             Category::Misread,
-            r"\bi (misread|misunderstood|misinterpreted)\b|\bi (read|took) .{0,40} as\b|\bnot what i (meant|wanted)\b|\bi meant\b",
+            r"\bnot what i (meant|wanted)\b|\bi meant\b",
+            r"\bi (misread|misunderstood|misinterpreted)\b|\bi (read|took) .{0,40} as\b",
         ),
         c(
             Category::MissedSomething,
-            r"\byou (missed|forgot|skipped|didn'?t|did not|never)\b|\bmissing\b|\bwhat about\b|\bleft out\b",
-        ),
-        c(
-            Category::StillBroken,
-            r"\bstill\b|\bsame (error|issue|problem)\b|\b(doesn'?t|didn'?t|not) work|\bbroken\b|\berror\b|\bcrash|\bfail",
+            &format!(r"\byou (missed|forgot|skipped|didn'?t|did not|never)\b|{missing}"),
+            missing,
         ),
+        c(Category::StillBroken, still, still),
     ]
 });
 
@@ -247,6 +333,10 @@ pub fn category(human: &str, ai: &str, admit: Option<(usize, usize)>) -> Categor
         Some((s, e)) => format!("{} {}", &ai[..s], &ai[e..]),
         None => ai.to_string(),
     };
-    let text = format!("{human}\n{ai}");
-    CATEGORIES.iter().find(|(_, re)| re.is_match(&text)).map_or(Category::JustWrong, |(c, _)| *c)
+    CATEGORIES
+        .iter()
+        .find(|(_, yours, ais)| {
+            yours.as_ref().is_some_and(|r| r.is_match(human)) || ais.as_ref().is_some_and(|r| r.is_match(&ai))
+        })
+        .map_or(Category::JustWrong, |(c, ..)| *c)
 }
diff --git a/crates/engine/src/tests.rs b/crates/engine/src/tests.rs
index 346e95f..c78f6ad 100644
--- a/crates/engine/src/tests.rs
+++ b/crates/engine/src/tests.rs
@@ -246,6 +246,7 @@ const EXAMPLES: &[(&str, &str, &str)] = &[
     ("ai.incorrectly", "I incorrectly assumed v2.", "It was incorrectly set upstream."),
     ("ai.that_was_wrong", "That was wrong of me.", "That was the wrong branch?"),
     ("ai.didnt_check", "I didn't check the logs.", "I didn't change the logs."),
+    ("you.unneeded", "No, that's not needed, remove it", "No, that's not needed, thanks"),
     ("you.no_that", "no, that's the wrong file", "no, that's fine, ship it"),
     ("you.no", "nope.", "no problem"),
     ("you.wrong", "not quite, the year is off", "that's not a problem"),
@@ -256,7 +257,7 @@ const EXAMPLES: &[(&str, &str, &str)] = &[
     ("you.i_asked", "I asked for a list", "I asked around and nobody knows"),
     ("you.why_did_you", "why did you delete that?", "why does this need a key?"),
     ("you.doesnt_work", "doesn't work on firefox", "doesn't matter, keep going"),
-    ("you.same_error", "same error as before", "same here"),
+    ("you.same_error", "same thing, it still crashes", "same thing for the footer please"),
     ("you.actually", "actually, that's the old API", "actually, that works now"),
     ("you.any.doesnt_exist", "the flag doesn't exist in v2", "the flag didn't exit cleanly"),
     ("you.any.not_working", "the login is not working", "I'm not worried about it"),
@@ -464,3 +465,163 @@ fn expecting_to_have_something_is_not_an_admission() {
         assert!(rules::first(&rules::AI_ADMIT, reply).is_some(), "{reply}");
     }
 }
+
+#[test]
+fn should_have_it_ready_is_a_promise_not_an_admission() {
+    for reply in
+        ["I should have a fix ready in a minute.", "I should have this done shortly.", "I should have that now."]
+    {
+        assert_eq!(rules::first(&rules::AI_ADMIT, reply).map(|m| m.rule), None, "{reply}");
+    }
+    for reply in ["I should have fixed this first.", "I should have tested that."] {
+        assert_eq!(rules::first(&rules::AI_ADMIT, reply).map(|m| m.rule), Some("ai.should_have"), "{reply}");
+    }
+}
+
+#[test]
+fn polite_closers_and_routine_requests_are_not_claims() {
+    for h in [
+        "No, that's not necessary.",
+        "No, that's all for now, thanks",
+        "No, that's all.",
+        "Nah, it's cool",
+        "No, that's alright",
+        "Nope, good to go",
+        "nope, not needed",
+        "That's not needed",
+        "that's not really necessary right now",
+        "Actually, that's not necessary",
+        "Same thing for the footer please",
+        "Not working on the API today, focus on the UI",
+    ] {
+        assert_eq!(label("Done. Anything else", h, "Sounds good."), None, "{h}");
+    }
+    // Real corrections still count.
+    for h in [
+        "No, that's wrong",
+        "No, that's not what I asked",
+        "No, that's all wrong",
+        "That's not right",
+        "same error",
+        "still the same issue",
+        "doesn't work",
+        "not working",
+        "not working at all",
+        "Not working on Safari",
+        "not working on my phone, the menu won't open",
+        "Not working on Safari anymore",
+        "Not working on mobile now",
+        "Not working on my machine right now either",
+        "Same thing",
+        "Same thing happens",
+        "Same thing, it still crashes",
+        // "Not needed" plus undoing the work is pushback.
+        "No, that's not needed, remove it",
+        "That's not needed, revert it",
+        "That's not necessary, I said only change the header",
+    ] {
+        assert_eq!(label("Done. Anything else", h, "Checking."), Some(Label::Claimed), "{h}");
+    }
+}
+
+#[test]
+fn agreeing_with_a_plain_question_is_not_an_admission() {
+    for (h, reply) in [
+        ("Is the default port 5432?", "You're correct, the default is 5432."),
+        ("does it run on node 20?", "You're right, it does."),
+        ("So I need to restart after editing .env?", "You're right, a restart is needed."),
+    ] {
+        assert_eq!(label("Set it to 5432.", h, reply), None, "{h}");
+    }
+    // Doubt and pushback still make it a correction.
+    assert_eq!(
+        label("Set it to 5433.", "No, that's wrong. Why did you use 5433?", "You're right, it's 5432."),
+        Some(Label::Confirmed)
+    );
+    assert_eq!(label("Set it to 5433.", "are you sure?", "You're right, it's 5432."), Some(Label::Admitted));
+    assert_eq!(label("Set it to 5433.", "isn't it 5432?", "You're correct, it's 5432."), Some(Label::Admitted));
+    // Owning a mistake after a question still counts: only agreement is ruled out.
+    assert_eq!(label("Set it to 5433.", "Is the default port 5432?", "My mistake, it is."), Some(Label::Admitted));
+    // A statement before the question, a request, or doubt: still an admission.
+    for h in [
+        "The function returns null when the list is empty. Can you fix it?",
+        "This crashes on empty input, can you fix that?",
+        "Shouldn't we handle errors here?",
+        "Shouldn't you check for null first?",
+        "Won't that break the build?",
+        "Can't you just use a map?",
+        "Isn't 5432 the default port?",
+        "Hold on, what about the empty case?",
+        "explain why you used 5432, it should be 5433",
+    ] {
+        assert_eq!(label("Set it to 5433.", h, "You're right, fixing it."), Some(Label::Admitted), "{h}");
+    }
+    // Later messages follow the same rule.
+    assert_eq!(after_checking("Is it Globex?", &["You're right, it's Globex."]).label, None);
+}
+
+#[test]
+fn added_context_never_counts_against_the_ai() {
+    for h in [
+        "By the way, we deploy on Vercel",
+        "oh, also we use pnpm",
+        "one other thing: it runs on Windows",
+        "Another thing, the API is v2",
+        "p.s. the dates are UTC",
+        "In addition, it must support Safari",
+        "I should have mentioned we use Postgres",
+        "I didn't mention it, but we deploy on Fly",
+    ] {
+        assert_eq!(label("Here's the config.", h, "Good catch! I'll add a vercel.json."), None, "{h}");
+    }
+    // Pastes and pushback that only look like context.
+    for h in [
+        "PS C:\\proj> npm run dev\nError: Cannot find module 'x'",
+        "ps aux shows it's still running",
+        "Why did you add a footer? I didn't mention a footer.",
+    ] {
+        assert!(rules::first(&rules::CONTEXT, h).is_none(), "{h}");
+    }
+    assert_eq!(
+        label("Added a footer.", "Why did you add a footer? I didn't mention a footer.", "Removing it."),
+        Some(Label::Claimed)
+    );
+}
+
+#[test]
+fn a_question_mark_in_code_is_not_the_ai_asking() {
+    for prev in [
+        "Done:\n```rust\nlet f = File::open(p)?;\n```",
+        "It reads `user?.name` now.",
+        "It now calls `read()?` before parsing.",
+        "See https://example.com/docs?page=2 for the flags.",
+        "It returns ok ? a : b now.",
+    ] {
+        assert_eq!(label(prev, "no, that's the wrong file", "Switching."), Some(Label::Claimed), "{prev}");
+    }
+    assert_eq!(label("Want me to run `cargo test`?", "No, the header is wrong", "Okay."), None);
+    for prev in [
+        "Want me to ship it?",
+        "Should I keep `read()?` here?",
+        "Use \"v2\"?\n\nOr v3?",
+        "Want me to run `cargo test`?",
+        "Should I proceed?🙂",
+        "Should I proceed? 🙂 ",
+    ] {
+        assert_eq!(label(prev, "no, that's the wrong file", "Switching."), None, "{prev}");
+    }
+}
+
+#[test]
+fn the_ais_incidental_words_do_not_pick_the_category() {
+    let cat = |h: &str, n: &str| classify(&turn("x", h, n)).category;
+    assert_eq!(cat("that's wrong", "You're right. There is no reason to keep that branch."), Some(Category::JustWrong));
+    assert_eq!(cat("that's wrong", "You're right, I already changed the wrong file."), Some(Category::JustWrong));
+    assert_eq!(cat("that's wrong", "You're right, let me re-read it."), Some(Category::JustWrong));
+    // Your side still picks it.
+    assert_eq!(cat("there is no such page", "You're right."), Some(Category::MadeItUp));
+    assert_eq!(cat("I already told you to keep it", "You're right."), Some(Category::IgnoredInstructions));
+    // Only the start of a later message is read for the category too.
+    let late = format!("You're right. {} I said it doesn't exist.", "x".repeat(300));
+    assert_eq!(after_checking("that's wrong", &[&late]).category, Some(Category::JustWrong));
+}
diff --git a/crates/sources/src/chatgpt.rs b/crates/sources/src/chatgpt.rs
index f10a5bb..d96695b 100644
--- a/crates/sources/src/chatgpt.rs
+++ b/crates/sources/src/chatgpt.rs
@@ -10,7 +10,8 @@ pub struct ChatGptExport;
 
 fn text_parts(msg: &Value) -> Option<String> {
     let content = &msg["content"];
-    if content["content_type"].as_str() != Some("text") {
+    // "multimodal_text" is a message with an attachment: its parts mix text with image records.
+    if !matches!(content["content_type"].as_str(), Some("text" | "multimodal_text")) {
         return None;
     }
     let parts: Vec<&str> = content["parts"].as_array()?.iter().filter_map(Value::as_str).collect();
@@ -128,6 +129,27 @@ mod tests {
         assert!(ChatGptExport.scan(&p, 0, false).err().unwrap().to_string().contains("not a valid export"));
     }
 
+    #[test]
+    fn messages_with_an_attachment_keep_their_text_parts() {
+        let multimodal = |role: &str, parts: Value| {
+            json!({"id":role,"author":{"role":role},"create_time":1.0,
+                "content":{"content_type":"multimodal_text","parts":parts},"metadata":{}})
+        };
+        let conv = json!({"conversation_id":"c1","current_node":"b","mapping":{
+            "a": {"parent": null, "message": multimodal("user", json!([
+                {"content_type":"image_asset_pointer","asset_pointer":"file-service://x","size_bytes":10},
+                "no, that's the wrong button"]))},
+            "b": {"parent": "a", "message": multimodal("assistant", json!(["You're right."]))}
+        }});
+        let ev = events(&conv);
+        assert_eq!(ev.len(), 2);
+        assert!(matches!(&ev[0], Event::Human { text, .. } if text == "no, that's the wrong button"));
+        let image_only =
+            json!({"content":{"content_type":"multimodal_text","parts":[{"content_type":"image_asset_pointer"}]}});
+        assert_eq!(text_parts(&image_only), None);
+        assert_eq!(text_parts(&json!({"content":{"content_type":"code","parts":["x"]}})), None);
+    }
+
     #[test]
     fn hidden_messages_are_skipped_and_cycles_end() {
         let conv = json!({"current_node":"b","mapping":{
diff --git a/crates/sources/src/claude_code.rs b/crates/sources/src/claude_code.rs
index 082f421..8bd8c7f 100644
--- a/crates/sources/src/claude_code.rs
+++ b/crates/sources/src/claude_code.rs
@@ -57,6 +57,11 @@ const NOISE: &[&str] = &[
     "<side-sessions-event",
     "<ci-monitor-event",
     "[Artifact comment",
+    // Other agents and scheduled tasks. Newer files say so in `origin`; older ones only in the text.
+    "<agent-message",
+    "<cross-session-message",
+    "<scheduled-task",
+    "Another Claude session sent a message",
 ];
 
 #[derive(Deserialize)]
@@ -71,6 +76,19 @@ struct Line {
     meta: bool,
     message: Option<Message>,
     attachment: Option<Attachment>,
+    origin: Option<Origin>,
+}
+
+/// Who a record came from: `human`, `peer` (another agent), `task-notification`. Older files have none.
+#[derive(Deserialize)]
+struct Origin {
+    kind: Option<String>,
+}
+
+/// True when the record says another agent, a scheduled task or a coordinator wrote it. Kinds not
+/// listed are kept: a new way for the person to type (voice, a phone) mustn't vanish from their history.
+fn not_human(origin: Option<&Origin>) -> bool {
+    origin.and_then(|o| o.kind.as_deref()).is_some_and(|k| matches!(k, "peer" | "task-notification" | "coordinator"))
 }
 
 #[derive(Deserialize)]
@@ -111,6 +129,7 @@ struct Attachment {
     kind: Option<String>,
     prompt: Option<serde_json::Value>,
     content: Option<serde_json::Value>,
+    origin: Option<Origin>,
 }
 
 fn text_of(content: &Content) -> Option<String> {
@@ -137,6 +156,25 @@ fn is_noise(s: &str) -> bool {
     NOISE.iter().any(|n| t.starts_with(n))
 }
 
+/// Removes `<pasted_content …>…</pasted_content …>` blocks, keeping what was typed around them.
+/// A block with no closing tag is left alone.
+fn strip_pasted(text: &str) -> String {
+    const OPEN: &str = "<pasted_content";
+    let mut out = String::with_capacity(text.len());
+    let mut rest = text;
+    while let Some(i) = rest.find(OPEN) {
+        let after = &rest[i..];
+        // The block ends at the closing tag with the same id, so a pasted tag inside it can't end it early.
+        let Some(open_end) = after.find('>') else { break };
+        let close = format!("</{}", &after[1..=open_end]);
+        let Some(end) = after[open_end..].find(&close).map(|c| open_end + c + close.len()) else { break };
+        out.push_str(&rest[..i]);
+        rest = &after[end..];
+    }
+    out.push_str(rest);
+    out.trim().to_string()
+}
+
 impl Source for ClaudeCode {
     fn id(&self) -> &'static str {
         self.id
@@ -238,6 +276,9 @@ struct Read {
 const SKILL: &str = "grade:me";
 
 fn read_line(buf: &[u8], pos: u64, usage: bool) -> Read {
+    if buf.is_empty() {
+        return Read { skipped: true, ..Read::default() }; // a line over the cap (see `for_each_line`)
+    }
     // Cheap filters before any JSON work. Tool results are user-role records too.
     let is_user = contains(buf, br#""type":"user""#);
     let is_ai = contains(buf, br#""type":"assistant""#);
@@ -333,6 +374,15 @@ fn sample(line: &Line, pos: u64) -> Option<(Option<String>, Sample)> {
     Some((m.id.clone(), Sample { pos, ts, model: m.model.clone(), usage }))
 }
 
+/// A person's message, unless it is injected context or only a pasted block.
+fn human(text: String, ts: DateTime<Utc>, line: &Line, pos: u64) -> Option<Event> {
+    if is_noise(&text) {
+        return None;
+    }
+    let text = if text.contains("<pasted_content") { strip_pasted(&text) } else { text };
+    (!text.trim().is_empty() && !is_noise(&text)).then(|| Event::Human { text, ts, key: line.uuid.clone(), pos })
+}
+
 /// The event one transcript line holds, if it's a message a person typed or a model wrote.
 fn event(line: &Line, pos: u64) -> Option<Event> {
     if line.sidechain || line.meta {
@@ -341,17 +391,24 @@ fn event(line: &Line, pos: u64) -> Option<Event> {
     let ts = line.timestamp.as_deref().and_then(parse_ts);
     match line.kind.as_deref()? {
         "user" => {
+            if not_human(line.origin.as_ref()) {
+                return None;
+            }
             let text = line.message.as_ref().and_then(|m| m.content.as_ref()).and_then(text_of)?;
-            (!is_noise(&text)).then_some(Event::Human { text, ts: ts?, key: line.uuid.clone(), pos })
+            human(text, ts?, line, pos)
         }
-        // A message typed while Claude was busy, delivered later as an attachment.
+        // A message typed while Claude was busy, delivered later as an attachment. The prompt is
+        // text, or content blocks when an image came with it.
         "attachment" => {
             let a = line.attachment.as_ref()?;
-            let Some(serde_json::Value::String(text)) = &a.prompt else { return None };
-            if a.kind.as_deref() != Some("queued_command") || is_noise(text) {
+            if a.kind.as_deref() != Some("queued_command") || not_human(line.origin.as_ref().or(a.origin.as_ref())) {
                 return None;
             }
-            Some(Event::Human { text: text.clone(), ts: ts?, key: line.uuid.clone(), pos })
+            let text = match a.prompt.as_ref()? {
+                serde_json::Value::String(t) => t.clone(),
+                other => Content::deserialize(other).ok().as_ref().and_then(text_of)?,
+            };
+            human(text, ts?, line, pos)
         }
         "assistant" => {
             let m = line.message.as_ref()?;
@@ -412,6 +469,12 @@ mod tests {
         assert_eq!((scan.turns.len(), scan.skipped_lines), (1, 2), "noise that no rule wants isn't counted");
     }
 
+    #[test]
+    fn a_line_over_the_cap_is_counted_as_skipped() {
+        assert!(read_line(b"", 0, false).skipped);
+        assert!(!read_line(b"unrelated\n", 0, false).skipped);
+    }
+
     #[test]
     fn notes_when_a_chat_was_given_the_tip_rules() {
         let scan = scan_str(&hook_context("hook_additional_context", engine::TIP_RULES, false), false);
@@ -507,6 +570,97 @@ mod tests {
         assert_eq!(scan.turns[0].key.as_deref(), Some("q1"));
     }
 
+    #[test]
+    fn queued_commands_with_an_image_are_read_by_their_text() {
+        let queued = line(&json!({"type":"attachment","uuid":"q1","timestamp":TS,"attachment":{
+            "type":"queued_command","prompt":[{"type":"text","text":"this looks off"},
+            {"type":"image","source":{"type":"base64","media_type":"image/png","data":"AAAA"}}]}}));
+        let image_only = line(&json!({"type":"attachment","uuid":"q2","timestamp":TS,"attachment":{
+            "type":"queued_command","prompt":[{"type":"image","source":{"type":"base64","data":"AAAA"}}]}}));
+        let body =
+            cc_ai("claude-opus-5-5", TS, "Done.") + &image_only + &queued + &cc_ai("claude-opus-5-5", TS, "Fixing it.");
+        let scan = scan_str(&body, true);
+        assert_eq!(scan.turns.len(), 1);
+        assert_eq!(scan.turns[0].turn.human, "this looks off");
+        assert_eq!(scan.turns[0].key.as_deref(), Some("q1"));
+    }
+
+    #[test]
+    fn messages_from_other_agents_and_tasks_are_not_the_persons() {
+        let peer_text = "Another Claude session sent a message:\nYou're right, it broke.";
+        let peer_queued = line(&json!({"type":"attachment","uuid":"q1","timestamp":TS,"origin":{"kind":"peer"},
+            "attachment":{"type":"queued_command","prompt":"<agent-message from=\"x\">hi</agent-message>"}}));
+        let nested = line(&json!({"type":"attachment","uuid":"q2","timestamp":TS,"attachment":{
+            "type":"queued_command","origin":{"kind":"task-notification"},"prompt":"anything at all"}}));
+        let tagged = line(&json!({"type":"user","uuid":"p1","timestamp":TS,"origin":{"kind":"peer"},
+            "message":{"content":"plain text, tagged by origin"}}));
+        let human_tagged = line(&json!({"type":"user","uuid":"h1","timestamp":TS,"origin":{"kind":"human"},
+            "message":{"content":"no, that's wrong"}}));
+        for body in [
+            cc_user("p", TS, peer_text),
+            cc_user("p", TS, "<scheduled-task name=\"x\">run</scheduled-task>"),
+            line(&json!({"type":"attachment","uuid":"q","timestamp":TS,"attachment":{
+                "type":"queued_command","prompt":"<cross-session-message>hi</cross-session-message>"}})),
+            peer_queued,
+            nested,
+            tagged,
+        ] {
+            let scan = scan_str(&(body.clone() + &cc_ai("claude-opus-5-5", TS, "ok")), true);
+            assert!(scan.turns.is_empty(), "{body}");
+        }
+        let scan = scan_str(&(human_tagged + &cc_ai("claude-opus-5-5", TS, "You're right.")), true);
+        assert_eq!(scan.turns.len(), 1, "origin human is kept");
+        let new_kind = line(&json!({"type":"user","uuid":"v1","timestamp":TS,"origin":{"kind":"voice"},
+            "message":{"content":"no, that's wrong"}}));
+        let scan = scan_str(&(new_kind + &cc_ai("claude-opus-5-5", TS, "You're right.")), true);
+        assert_eq!(scan.turns.len(), 1, "a kind not known to be another writer is kept");
+    }
+
+    #[test]
+    fn a_chat_with_no_turns_is_still_listed_as_read() {
+        let body = cc_user("p", TS, "Another Claude session sent a message:\nhi") + &cc_ai("claude-opus-5-5", TS, "ok");
+        let scan = scan_str(&body, true);
+        assert!(scan.turns.is_empty());
+        assert_eq!(scan.sessions.len(), 1, "a rebuild must know this chat was read");
+    }
+
+    #[test]
+    fn a_peer_message_does_not_cut_the_replies_short() {
+        let peer = line(&json!({"type":"attachment","uuid":"q1","timestamp":TS,"origin":{"kind":"peer"},
+            "attachment":{"type":"queued_command","prompt":"<agent-message from=\"x\">status</agent-message>"}}));
+        let body = cc_ai("claude-opus-5-5", TS, "I moved the header.")
+            + &cc_user("u1", TS, "the header still overlaps the menu")
+            + &cc_ai("claude-opus-5-5", TS, "Let me look.")
+            + &peer
+            + &cc_ai("claude-opus-5-5", TS, "You're right, my fix was wrong. Sorry about that.");
+        let scan = scan_str(&body, true);
+        assert_eq!(scan.turns.len(), 1);
+        let t = &scan.turns[0].turn;
+        assert_eq!(t.human, "the header still overlaps the menu");
+        assert!(t.later_ai.iter().any(|r| r.contains("my fix was wrong")), "{:?}", t.later_ai);
+        assert_eq!(engine::classify(t).label, Some(engine::Label::Admitted));
+    }
+
+    #[test]
+    fn pasted_blocks_are_dropped_and_the_typed_text_kept() {
+        let pasted = |body: &str| format!("<pasted_content id=\"1\">{body}</pasted_content id=\"1\">");
+        let humans = |text: String| {
+            let body =
+                cc_ai("claude-opus-5-5", TS, "Done.") + &cc_user("u1", TS, &text) + &cc_ai("claude-opus-5-5", TS, "ok");
+            scan_str(&body, true).turns.iter().map(|t| t.turn.human.clone()).collect::<Vec<_>>()
+        };
+        assert_eq!(humans(pasted("error: x") + "\nno, that's wrong"), ["no, that's wrong"]);
+        assert_eq!(humans(format!("{}\n{}\nwhy?", pasted("a"), pasted("b"))), ["why?"]);
+        assert_eq!(humans(pasted("only a paste")).len(), 0);
+        let open = "<pasted_content id=\"1\">never closed no, wrong";
+        assert_eq!(humans(open.to_string()), [open]);
+        assert_eq!(strip_pasted("see <pasted_content>x</pasted_content> now"), "see  now");
+        // A paste that quotes another paste's closing tag ends only at its own.
+        let nested =
+            "<pasted_content id=\"2\">log </pasted_content id=\"9\"> more</pasted_content id=\"2\">still wrong";
+        assert_eq!(strip_pasted(nested), "still wrong");
+    }
+
     #[test]
     fn synthetic_replies_are_skipped() {
         let body = cc_ai("claude-opus-5-5", TS, "Use /admin.")
diff --git a/crates/sources/src/codex.rs b/crates/sources/src/codex.rs
index 07c11c6..165e637 100644
--- a/crates/sources/src/codex.rs
+++ b/crates/sources/src/codex.rs
@@ -24,6 +24,8 @@ const NOISE: &[&str] = &[
     "<user_instructions",
     "<permissions",
     "<turn_aborted",
+    "<skill",
+    "<hook_prompt",
 ];
 
 /// Codex's folder: `$CODEX_HOME`, else `~/.codex`. A variable that isn't an absolute path counts as
@@ -40,14 +42,20 @@ fn is_zst(p: &Path) -> bool {
     p.extension().is_some_and(|e| e == "zst")
 }
 
+/// How a message with attachments begins.
+const PREAMBLES: &[&str] = &["# Files mentioned by the user", "# Files pasted by the user", "Image attachment"];
+
 /// Drops the attachment preamble Codex puts before what you actually typed.
 fn strip_preamble(text: &str) -> &str {
     // "## My request for Codex:" in older builds, "## My request:" in newer ones.
     let mut t = text.trim_start();
-    if let Some(i) = t.find("## My request") {
-        let rest = &t[i..];
-        t = rest.find(':').map_or(rest, |c| &rest[c + 1..]).trim_start();
-        return t;
+    // Only when the message opens with an attachment preamble: a person can write that heading too.
+    if PREAMBLES.iter().any(|p| t.starts_with(p)) {
+        if let Some(i) = t.find("## My request") {
+            let rest = &t[i..];
+            t = rest.find(':').map_or(rest, |c| &rest[c + 1..]).trim_start();
+            return t;
+        }
     }
     t = t.strip_prefix("# Files mentioned by the user:").unwrap_or(t).trim_start();
     // Attachment lines look like "## Screenshot 1.png: /path" or "## repo: ~/code/repo/".
@@ -85,12 +93,13 @@ fn strip_images(text: &str) -> String {
     out.trim().to_string()
 }
 
-/// Only threads a person started. Guardian / subagent threads are model-to-model.
+/// Only threads a person is in. Guardian / subagent threads are model-to-model.
 fn human_thread(meta: &Value) -> bool {
     if meta.get("source").is_some_and(|s| s.get("subagent").is_some()) {
         return false;
     }
-    meta.get("thread_source").and_then(Value::as_str).is_none_or(|s| s == "user")
+    // Handoffs from ChatGPT and voice chats are a person talking; only model-to-model threads are skipped.
+    meta.get("thread_source").and_then(Value::as_str).is_none_or(|s| !matches!(s, "guardian_review" | "subagent"))
 }
 
 impl Source for Codex {
@@ -274,6 +283,10 @@ impl Rollout {
 
     fn line(&mut self, pos: u64, buf: &[u8]) {
         let fresh = pos >= self.from;
+        if buf.is_empty() {
+            self.skipped += usize::from(fresh); // a line over the cap (see `for_each_line`)
+            return;
+        }
         let meta = contains(buf, b"session_meta") || contains(buf, b"turn_context");
         let counts = self.usage && contains(buf, b"token_count");
         if !(meta || counts || (fresh && contains(buf, b"\"message\""))) {
@@ -358,13 +371,48 @@ mod tests {
         );
         assert_eq!(strip_images("poorly reviewed <image name=x path=\"/a.png\"> </image>"), "poorly reviewed");
         assert_eq!(strip_preamble("hello"), "hello");
+        assert_eq!(strip_preamble("# Files pasted by the user:\n- a.png\n\n## My request:\nfix it"), "fix it");
+    }
+
+    #[test]
+    fn a_heading_a_person_wrote_keeps_their_full_text() {
+        let text = "no, that's wrong. Put it under the ## My request: heading";
+        assert_eq!(strip_preamble(text), text);
+    }
+
+    #[test]
+    fn skill_and_hook_bodies_are_not_the_persons_messages() {
+        let body = opening()
+            + &msg("assistant", "output_text", "Done.", "a2")
+            + &msg("user", "input_text", "<skill>\n<name>x</name>\nbody\n</skill>", "s1")
+            + &msg("user", "input_text", "<hook_prompt hook_run_id=\"1\">keep going</hook_prompt>", "h1")
+            + &msg("assistant", "output_text", "ok", "a3");
+        let (_d, p) = rollout(&body);
+        let scan = Codex.scan(&p, 0, true).unwrap();
+        assert_eq!(scan.turns.len(), 1);
+    }
+
+    #[test]
+    fn handoff_and_voice_threads_are_read() {
+        for source in ["chatgpt_handoff", "voice_chat"] {
+            let body = line(&json!({"type":"session_meta","payload":{"id":"x","thread_source":source}}))
+                + &msg("assistant", "output_text", "Plan.", "a")
+                + &msg("user", "input_text", "no, that's wrong", "u")
+                + &msg("assistant", "output_text", "You're right.", "b");
+            let (_d, p) = rollout(&body);
+            assert_eq!(Codex.scan(&p, 0, true).unwrap().turns.len(), 1, "{source}");
+        }
     }
 
     #[test]
     fn skips_guardian_threads() {
         assert!(!human_thread(&serde_json::json!({"source":{"subagent":{"other":"guardian"}}})));
         assert!(human_thread(&serde_json::json!({"source":"vscode","thread_source":"user"})));
-        assert!(!human_thread(&serde_json::json!({"thread_source":"agent"})));
+        assert!(!human_thread(&serde_json::json!({"thread_source":"guardian_review"})));
+        assert!(!human_thread(&serde_json::json!({"thread_source":"subagent"})));
+        for s in ["chatgpt_handoff", "voice_chat"] {
+            assert!(human_thread(&serde_json::json!({"thread_source":s})), "{s}");
+        }
     }
 
     fn msg(role: &str, kind: &str, text: &str, id: &str) -> String {
@@ -556,7 +604,7 @@ mod tests {
         assert_eq!((light.session.as_str(), light.prompts.len()), ("rollout-2026-10-20-abc", 1));
         assert_eq!(light.prompts[0].key, whole.prompts[1].key);
         assert_eq!(light.prompts[0].usage, Usage::default());
-        let guardian = line(&json!({"type":"session_meta","payload":{"id":"abc","thread_source":"agent"}}));
+        let guardian = line(&json!({"type":"session_meta","payload":{"id":"abc","thread_source":"guardian_review"}}));
         let (_d2, p2) = rollout(&(guardian.clone() + &early + &late));
         let skipped = Codex.view_from(&p2, (guardian.len() + early.len()) as u64, false).unwrap().unwrap();
         assert!(skipped.prompts.is_empty(), "a model-to-model thread has no person's prompts");
diff --git a/crates/sources/src/export.rs b/crates/sources/src/export.rs
index 8ca995a..7075d8c 100644
--- a/crates/sources/src/export.rs
+++ b/crates/sources/src/export.rs
@@ -1,53 +1,144 @@
-//! One-off data exports: a `conversations.json` array, read whole every time it changes.
+//! One-off data exports: a `conversations.json` array, read in full every time it changes.
 
 use crate::{pair, Error, Event, FileScan, Result, MAX_READ_BYTES};
+use serde::de::{Deserializer, SeqAccess, Visitor};
 use serde_json::Value;
-use std::io::Read;
+use std::io::{BufReader, Read};
 use std::path::Path;
 
-/// Reads `path` whole, or says it is over `limit` bytes without reading past that.
-fn read_capped(path: &Path, limit: u64) -> Result<Vec<u8>> {
-    let mut bytes = Vec::new();
-    let file = std::fs::File::open(path).map_err(Error::io(path))?;
-    file.take(limit + 1).read_to_end(&mut bytes).map_err(Error::io(path))?;
-    if bytes.len() as u64 > limit {
-        return Err(Error::TooBig { path: path.to_path_buf(), limit });
+/// Calls `each` with every element of the top-level array in `reader`, one at a time, so only one
+/// conversation is held in memory (not the whole export as a tree of values).
+fn each_element<R: Read>(reader: R, mut each: impl FnMut(Value)) -> serde_json::Result<()> {
+    struct Elements<'a, F>(&'a mut F);
+    impl<'de, F: FnMut(Value)> Visitor<'de> for Elements<'_, F> {
+        type Value = ();
+        fn expecting(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
+            f.write_str("an array")
+        }
+        fn visit_seq<A: SeqAccess<'de>>(self, mut seq: A) -> std::result::Result<(), A::Error> {
+            while let Some(v) = seq.next_element::<Value>()? {
+                (self.0)(v);
+            }
+            Ok(())
+        }
     }
-    Ok(bytes)
+    let mut de = serde_json::Deserializer::from_reader(reader);
+    de.deserialize_seq(Elements(&mut each))?;
+    de.end()
 }
 
 /// Reads an export: each conversation's events (from `events`) become turns in the session
 /// `session(conv)` names. The checkpoint is the file size, since exports are never appended to.
+/// The file is read as a stream and never past [`MAX_READ_BYTES`].
 pub(crate) fn scan_export(
     source: &'static str,
     path: &Path,
     session: fn(&Value) -> &str,
     events: fn(&Value) -> Vec<Event>,
 ) -> Result<FileScan> {
-    let bytes = read_capped(path, MAX_READ_BYTES)?;
-    let convs: Vec<Value> =
-        serde_json::from_slice(&bytes).map_err(|source| Error::Json { path: path.to_path_buf(), source })?;
-    let mut scan = FileScan { checkpoint: bytes.len() as u64, ..FileScan::default() };
-    for c in &convs {
-        let one = pair(source, session(c), events(c), 0, true);
+    scan_export_capped(source, path, MAX_READ_BYTES, session, events)
+}
+
+fn scan_export_capped(
+    source: &'static str,
+    path: &Path,
+    limit: u64,
+    session: fn(&Value) -> &str,
+    events: fn(&Value) -> Vec<Event>,
+) -> Result<FileScan> {
+    let file = std::fs::File::open(path).map_err(Error::io(path))?;
+    let mut counted = Counted { inner: BufReader::new(file.take(limit + 1)), read: 0 };
+    let mut scan = FileScan::default();
+    let parsed = each_element(&mut counted, |c| {
+        let one = pair(source, session(&c), events(&c), 0, true);
         scan.turns.extend(one.turns);
+        scan.sessions.extend(one.sessions);
         scan.cards.extend(one.cards);
+    });
+    // Past the cap the stream is cut off, which also looks like a broken file: say which it was.
+    if counted.read > limit {
+        return Err(Error::TooBig { path: path.to_path_buf(), limit });
     }
+    parsed.map_err(|source| Error::Json { path: path.to_path_buf(), source })?;
+    scan.checkpoint = counted.read;
     Ok(scan)
 }
 
+/// A reader that counts the bytes taken through it.
+struct Counted<R> {
+    inner: R,
+    read: u64,
+}
+
+impl<R: Read> Read for Counted<R> {
+    fn read(&mut self, buf: &mut [u8]) -> std::io::Result<usize> {
+        let n = self.inner.read(buf)?;
+        self.read += n as u64;
+        Ok(n)
+    }
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;
+    use serde_json::json;
+
+    fn id(c: &Value) -> &str {
+        c["id"].as_str().unwrap_or_default()
+    }
+
+    fn events(c: &Value) -> Vec<Event> {
+        let ts = chrono::DateTime::from_timestamp(1, 0).unwrap();
+        let text = |k: &str| c[k].as_str().unwrap_or_default().to_string();
+        vec![
+            Event::Ai { text: text("ai"), model: None, ts: Some(ts), key: None, pos: 0 },
+            Event::Human { text: text("human"), ts, key: Some(text("id")), pos: 1 },
+            Event::Ai { text: "You're right.".into(), model: None, ts: Some(ts), key: None, pos: 2 },
+        ]
+    }
+
+    fn write(dir: &tempfile::TempDir, body: &[u8]) -> std::path::PathBuf {
+        let p = dir.path().join("conversations.json");
+        std::fs::write(&p, body).unwrap();
+        p
+    }
 
     #[test]
     fn an_export_over_the_cap_is_refused_with_a_plain_error() {
         let dir = tempfile::tempdir().unwrap();
-        let p = dir.path().join("conversations.json");
-        std::fs::write(&p, "[1, 2, 3, 4, 5]").unwrap();
-        assert_eq!(read_capped(&p, 15).unwrap().len(), 15);
-        let err = read_capped(&p, 14).unwrap_err();
+        let p = write(&dir, b"[1, 2, 3, 4, 5]");
+        assert!(scan_export_capped("x", &p, 15, id, events).is_ok(), "exactly at the cap is fine");
+        assert!(matches!(scan_export_capped("x", &p, 14, id, events), Err(Error::TooBig { .. })));
+        let body = serde_json::to_vec(&json!([{"id":"a","ai":"Use x","human":"no, wrong"}])).unwrap();
+        let p = write(&dir, &body);
+        assert_eq!(scan_export_capped("x", &p, body.len() as u64, id, events).unwrap().checkpoint, body.len() as u64);
+        let err = scan_export_capped("x", &p, body.len() as u64 - 1, id, events).unwrap_err();
         assert!(matches!(err, Error::TooBig { .. }));
         assert!(err.to_string().contains("conversations.json: too large to read"), "{err}");
     }
+
+    #[test]
+    fn many_conversations_stream_through_one_at_a_time() {
+        let dir = tempfile::tempdir().unwrap();
+        let convs: Vec<Value> =
+            (0..50).map(|i| json!({"id": format!("c{i}"), "ai": "Use x", "human": format!("no, wrong {i}")})).collect();
+        let body = serde_json::to_vec_pretty(&convs).unwrap();
+        let scan = scan_export("x", &write(&dir, &body), id, events).unwrap();
+        assert_eq!(scan.checkpoint, body.len() as u64);
+        let want: Vec<_> = convs.iter().flat_map(|c| pair("x", id(c), events(c), 0, true).turns).collect();
+        assert_eq!(scan.turns.len(), 50);
+        let key = |t: &crate::SourceTurn| (t.session.clone(), t.key.clone(), t.turn.human.clone());
+        assert_eq!(scan.turns.iter().map(key).collect::<Vec<_>>(), want.iter().map(key).collect::<Vec<_>>());
+        assert_eq!(scan.turns[49].session, "c49");
+    }
+
+    #[test]
+    fn a_broken_or_wrong_shaped_export_is_a_plain_json_error() {
+        let dir = tempfile::tempdir().unwrap();
+        for body in [&b"{not json"[..], b"{}", b"[{\"id\":\"a\"}", b"[] trailing", b""] {
+            let err = scan_export("x", &write(&dir, body), id, events).unwrap_err();
+            assert!(matches!(err, Error::Json { .. }), "{err}");
+        }
+        assert!(scan_export("x", &write(&dir, b"[]"), id, events).unwrap().turns.is_empty());
+    }
 }
diff --git a/crates/sources/src/jsonl.rs b/crates/sources/src/jsonl.rs
index 80f6bd1..a725b95 100644
--- a/crates/sources/src/jsonl.rs
+++ b/crates/sources/src/jsonl.rs
@@ -2,7 +2,7 @@
 
 use crate::{Error, Result};
 use std::fs::File;
-use std::io::{self, BufRead, BufReader, Seek, SeekFrom};
+use std::io::{self, BufRead, BufReader, Read, Seek, SeekFrom};
 use std::path::Path;
 
 /// Opens `path` for line reading, positioned at byte `from`.
@@ -12,16 +12,57 @@ pub(crate) fn open_at(path: &Path, from: u64) -> Result<BufReader<File>> {
     Ok(BufReader::with_capacity(1 << 16, f))
 }
 
+/// The longest line read, newline included. A longer one is a corrupt or hostile file, not a message.
+const MAX_LINE: usize = 64 << 20;
+
 /// Calls `f(line_pos, line)` for each complete line (newline included) of `reader`, whose first
 /// byte sits at offset `start`. Stops at EOF or at a last line still being written, and returns
 /// the offset just past the last complete line: where the next pass should resume.
-pub(crate) fn for_each_line<R: BufRead>(mut reader: R, start: u64, mut f: impl FnMut(u64, &[u8])) -> io::Result<u64> {
+///
+/// A line over [`MAX_LINE`] is skipped to its newline and reported as an empty `line`, which no real
+/// line is. If the file ends inside one, the pass resumes at the end of the file, so the unfinished
+/// line isn't read into memory again on every pass.
+pub(crate) fn for_each_line<R: BufRead>(reader: R, start: u64, f: impl FnMut(u64, &[u8])) -> io::Result<u64> {
+    for_each_line_capped(reader, start, MAX_LINE, f)
+}
+
+fn for_each_line_capped<R: BufRead>(
+    mut reader: R,
+    start: u64,
+    cap: usize,
+    mut f: impl FnMut(u64, &[u8]),
+) -> io::Result<u64> {
     let mut pos = start;
     let mut buf = Vec::new();
     loop {
         buf.clear();
-        let n = reader.read_until(b'\n', &mut buf)?;
-        if n == 0 || buf.last() != Some(&b'\n') {
+        let n = Read::take(&mut reader, cap as u64 + 1).read_until(b'\n', &mut buf)?;
+        if n == 0 {
+            return Ok(pos);
+        }
+        if n > cap {
+            let mut ended = buf.last() == Some(&b'\n');
+            drop(std::mem::take(&mut buf)); // give the memory back before reading on
+            let mut skipped = n as u64;
+            while !ended {
+                let chunk = reader.fill_buf()?;
+                if chunk.is_empty() {
+                    break;
+                }
+                let (used, found) =
+                    chunk.iter().position(|&b| b == b'\n').map_or((chunk.len(), false), |i| (i + 1, true));
+                reader.consume(used);
+                skipped += used as u64;
+                ended = found;
+            }
+            f(pos, b"");
+            pos += skipped;
+            if !ended {
+                return Ok(pos); // EOF inside it
+            }
+            continue;
+        }
+        if buf.last() != Some(&b'\n') {
             return Ok(pos); // EOF, or a line still being written: resume here next time
         }
         f(pos, &buf);
@@ -42,6 +83,29 @@ mod tests {
         assert_eq!(for_each_line(&b""[..], 4, |_, _| {}).unwrap(), 4);
     }
 
+    #[test]
+    fn a_line_over_the_cap_is_skipped_and_reported_empty() {
+        let mut seen = Vec::new();
+        let body = &b"ab\nTOO-LONG-LINE\nok\n"[..];
+        let end = for_each_line_capped(body, 0, 5, |pos, line| seen.push((pos, line.to_vec()))).unwrap();
+        assert_eq!(seen, vec![(0, b"ab\n".to_vec()), (3, vec![]), (17, b"ok\n".to_vec())]);
+        assert_eq!(end, 20);
+        // Exactly at the cap is still a line; one more byte is not.
+        let mut seen = Vec::new();
+        for_each_line_capped(&b"abcd\nabcde\n"[..], 0, 5, |pos, line| seen.push((pos, line.len()))).unwrap();
+        assert_eq!(seen, vec![(0, 5), (5, 0)]);
+    }
+
+    #[test]
+    fn an_unfinished_line_over_the_cap_is_skipped_once() {
+        let mut seen = Vec::new();
+        let end =
+            for_each_line_capped(&b"ab\nXXXXXXXXXXXX"[..], 0, 5, |pos, line| seen.push((pos, line.len()))).unwrap();
+        assert_eq!((seen, end), (vec![(0, 3), (3, 0)], 15), "resumes past it, not at its start");
+        // A short unfinished line still waits for its newline.
+        assert_eq!(for_each_line_capped(&b"ab\nXX"[..], 0, 5, |_, _| {}).unwrap(), 3);
+    }
+
     #[test]
     fn open_at_resumes_mid_file_and_names_missing_files() {
         let dir = tempfile::tempdir().unwrap();
diff --git a/crates/sources/src/lib.rs b/crates/sources/src/lib.rs
index 46e3278..becfdd4 100644
--- a/crates/sources/src/lib.rs
+++ b/crates/sources/src/lib.rs
@@ -105,6 +105,9 @@ pub struct SourceTurn {
 pub struct FileScan {
     /// Complete turns, in file order.
     pub turns: Vec<SourceTurn>,
+    /// Every chat this pass read, turns or not. A chat whose only messages weren't the person's has
+    /// no turns, but it was still read, so a rebuild mustn't put back what older rules counted in it.
+    pub sessions: Vec<String>,
     /// Where the next pass should resume. Only ever past turns that are complete.
     pub checkpoint: u64,
     /// Record blocks found in AI replies, in file order. A pass that re-reads a reply (after a
diff --git a/crates/sources/src/pair.rs b/crates/sources/src/pair.rs
index 10526fb..c8fb5dc 100644
--- a/crates/sources/src/pair.rs
+++ b/crates/sources/src/pair.rs
@@ -112,7 +112,7 @@ fn pair_at(
     let tips = read_tips(session, &events);
     // A tip on its own is no reply: what came before it is still what the person answered.
     events.retain(|e| !matches!(e, Event::Ai { text, .. } if only_a_tip(text)));
-    let mut scan = FileScan { checkpoint: end, tips, ..FileScan::default() };
+    let mut scan = FileScan { checkpoint: end, tips, sessions: vec![session.to_string()], ..FileScan::default() };
     // The chat's model near each message: the last real one before it, else the first after it.
     let real = |e: &Event| match e {
         Event::Ai { model: Some(m), .. } if engine::is_real_model(m) => Some(m.clone()),
diff --git a/crates/sources/src/testing.rs b/crates/sources/src/testing.rs
index 4a898c2..17b679c 100644
--- a/crates/sources/src/testing.rs
+++ b/crates/sources/src/testing.rs
@@ -15,6 +15,27 @@ pub fn cc_user(uuid: &str, ts: &str, text: &str) -> String {
     line(&json!({"type":"user","uuid":uuid,"timestamp":ts,"message":{"content":text}}))
 }
 
+/// A Claude Code record the person typed, with the `promptId` the app's hooks and report cards go by.
+pub fn cc_prompt(uuid: &str, prompt_id: &str, ts: &str, text: &str) -> String {
+    line(&json!({"type":"user","uuid":uuid,"promptId":prompt_id,"timestamp":ts,"message":{"content":text}}))
+}
+
+/// A Claude Code reply from `model` with its own uuid and message id, and its token counts as
+/// (new input, cache write, cache read, output).
+pub fn cc_reply_with_usage(
+    model: &str,
+    uuid: &str,
+    message_id: &str,
+    ts: &str,
+    text: &str,
+    (input, cache_write, cache_read, output): (u64, u64, u64, u64),
+) -> String {
+    line(&json!({"type":"assistant","uuid":uuid,"timestamp":ts,"message":{"id":message_id,"model":model,
+        "content":[{"type":"text","text":text}],
+        "usage":{"input_tokens":input,"cache_creation_input_tokens":cache_write,"cache_read_input_tokens":cache_read,
+            "output_tokens":output}}}))
+}
+
 /// A Claude Code reply from `model`.
 pub fn cc_ai(model: &str, ts: &str, text: &str) -> String {
     line(&json!({"type":"assistant","timestamp":ts,"message":{"model":model,"content":[{"type":"text","text":text}]}}))
diff --git a/crates/store/src/card_tests.rs b/crates/store/src/card_tests.rs
index ba07f50..305419e 100644
--- a/crates/store/src/card_tests.rs
+++ b/crates/store/src/card_tests.rs
@@ -7,7 +7,7 @@ use crate::rebuild::SCORE_TABLES;
 use chrono::{NaiveDate, SecondsFormat, TimeZone, Utc};
 use engine::{habit_bit, seal, Usage};
 use serde_json::{json, Value};
-use sources::testing::{cc_user, line, set_mtime_ago};
+use sources::testing::{cc_reply_with_usage, cc_user, line, set_mtime_ago};
 use sources::{ChatView, Kind, Source};
 use std::io::Write;
 use std::path::PathBuf;
@@ -43,9 +43,7 @@ fn reply(uuid: &str, ts: &str, text: &str, usage: (u64, u64, u64, u64)) -> Strin
 
 /// The same, from `model`.
 fn reply_on(model: &str, uuid: &str, ts: &str, text: &str, usage: (u64, u64, u64, u64)) -> String {
-    line(&json!({"type":"assistant","uuid":uuid,"timestamp":ts,"message":{"id":format!("m-{uuid}"),
-        "model":model,"content":[{"type":"text","text":text}],"usage":{"input_tokens":usage.0,
-        "cache_creation_input_tokens":usage.1,"cache_read_input_tokens":usage.2,"output_tokens":usage.3}}}))
+    cc_reply_with_usage(model, uuid, &format!("m-{uuid}"), ts, text, usage)
 }
 
 /// Four prompts ten minutes apart, `p1` to `p4`: a request, two follow-ups and a sign-off.
diff --git a/crates/store/src/ingest.rs b/crates/store/src/ingest.rs
index 73c996c..092c8e0 100644
--- a/crates/store/src/ingest.rs
+++ b/crates/store/src/ingest.rs
@@ -139,6 +139,15 @@ impl Store {
         let rebuilding = self.rebuild_pending()?;
         let tx = self.conn.transaction()?;
         let (mut messages, mut corrections) = (0, 0);
+        if rebuilding {
+            // Read again, even with no turns of the person's in it: what older rules counted there goes.
+            for chat in &scan.sessions {
+                tx.execute(
+                    "INSERT OR IGNORE INTO rebuild_read (id, source) VALUES (?1, ?2)",
+                    params![hash(&[source, chat]), source],
+                )?;
+            }
+        }
         for st in scan.turns.iter().filter(|st| state.wants(st.ts)) {
             if let Some(corrected) = count_turn(&tx, source, st, level, rebuilding)? {
                 messages += 1;
diff --git a/crates/store/src/rebuild.rs b/crates/store/src/rebuild.rs
index a1efd7a..000f17d 100644
--- a/crates/store/src/rebuild.rs
+++ b/crates/store/src/rebuild.rs
@@ -59,8 +59,9 @@ impl Store {
     }
 
     /// Ends a rebuild once history has been read again. A chat that wasn't read again comes back as
-    /// it was counted: its file is gone, or its source is stopped. Counts per day have no chat on
-    /// them, so a day with nothing read again comes back whole, and a day that a missing chat
+    /// it was counted: its file is gone, or its source is stopped. A chat read again with nothing of
+    /// the person's left in it stays gone. Counts per day have no chat on them, so a day with nothing
+    /// read again comes back whole if a missing chat (or no known chat) touched it, and a day that a missing chat
     /// touched keeps the larger message count and adds back that chat's corrections. Other days take
     /// the new count. Message keys of missing chats come back too, so a file that returns can't
     /// count twice. A correction that a fork already brought back isn't restored a second time.
@@ -96,7 +97,16 @@ impl Store {
             ),
             [],
         )?;
-        tx.execute(&format!("INSERT INTO daily SELECT k.* FROM kept_daily k WHERE NOT {READ_DAY}"), [])?;
+        // A day with nothing read again comes back whole, unless every chat known on it was read again
+        // and simply holds nothing of the person's under the new rules.
+        tx.execute(
+            &format!(
+                "INSERT INTO daily SELECT k.* FROM kept_daily k WHERE NOT {READ_DAY} AND ({MISSING_CHAT_DAY}
+                   OR NOT EXISTS (SELECT 1 FROM kept_sessions s WHERE s.source = k.source
+                     AND k.day BETWEEN s.first_day AND s.last_day))"
+            ),
+            [],
+        )?;
         tx.execute(
             &format!(
                 "INSERT INTO events SELECT NULL, ts, day, source, session, provider, family, version,
diff --git a/crates/store/src/tests.rs b/crates/store/src/tests.rs
index c11cb63..9a09c63 100644
--- a/crates/store/src/tests.rs
+++ b/crates/store/src/tests.rs
@@ -594,6 +594,29 @@ fn a_rebuild_keeps_what_it_cant_read_again() {
     assert_eq!(s.get_setting(REBUILDING_KEY).unwrap(), None);
 }
 
+#[test]
+fn a_rebuild_drops_a_chat_that_no_longer_has_any_of_your_messages() {
+    let dir = tempfile::tempdir().unwrap();
+    let ts = (Utc::now() - Duration::hours(1)).to_rfc3339();
+    let file = dir.path().join("a.jsonl");
+    write_settled(&file, cc_ai("claude-opus-5-5", &ts, "hi") + &cc_correction("u1", &ts));
+    let all = claude_code(&dir);
+    let mut s = Store::open_in_memory().unwrap();
+    s.allow("claude-code", None).unwrap();
+    sync(&mut s, &all).unwrap();
+    assert_eq!(kept_counts(&s).1, 1);
+
+    // Newer rules see the same chat's only message as another agent's, not the person's.
+    let agent = "Another Claude session sent a message:\nNo, that's wrong.";
+    write_settled(
+        &file,
+        cc_ai("claude-opus-5-5", &ts, "hi") + &cc_user("u1", &ts, agent) + &cc_ai("claude-opus-5-5", &ts, "ok"),
+    );
+    s.set_setting("rules_version", "1").unwrap();
+    assert!(sync(&mut s, &all).unwrap().rescored);
+    assert_eq!(kept_counts(&s), (0, 0, 0, 0), "the chat was read again, so what older rules counted goes");
+}
+
 #[test]
 fn a_day_with_some_of_its_chats_gone_keeps_them() {
     let dir = tempfile::tempdir().unwrap();
@@ -1276,3 +1299,183 @@ fn a_file_that_cant_be_read_is_tried_again_only_when_it_changes() {
     std::fs::write(&file, "still not json!").unwrap();
     assert_eq!(sync(&mut s, &all).unwrap().errors.len(), 1);
 }
+
+// ---- the privacy steps PRIVACY.md promises ----------------------------------------------------
+
+#[test]
+fn a_sync_steps_old_text_down_in_the_same_pass() {
+    let dir = tempfile::tempdir().unwrap();
+    let ts = (Utc::now() - Duration::days(40)).to_rfc3339();
+    write_settled(&dir.path().join("a.jsonl"), cc_ai("claude-opus-5-5", &ts, "hi") + &cc_correction("u1", &ts));
+    let all = claude_code(&dir);
+    let mut s = Store::open_in_memory().unwrap();
+    s.set_retention(Retention { level: Level::Full, full_days: 30, snippet_days: 365 }).unwrap();
+    s.allow("claude-code", None).unwrap();
+    assert_eq!(sync(&mut s, &all).unwrap().corrections, 1);
+    // Ingested at Full, so only the sweep at the end of the sync can have taken the full text away.
+    let (level, full, snip): (String, Option<String>, Option<String>) = s
+        .conn
+        .query_row("SELECT level, human_full, human_snip FROM events", [], |r| Ok((r.get(0)?, r.get(1)?, r.get(2)?)))
+        .unwrap();
+    assert_eq!(level, "snippet");
+    assert_eq!(full, None, "the full text is gone");
+    assert!(snip.is_some(), "the snippet stays until it is older than its own limit");
+}
+
+#[test]
+fn a_sweep_that_deletes_text_empties_the_write_ahead_log() {
+    let dir = tempfile::tempdir().unwrap();
+    let wal_path = dir.path().join("t.db-wal");
+    let wal = || std::fs::metadata(&wal_path).map_or(0, |m| m.len());
+    let mut s = Store::open(&dir.path().join("t.db")).unwrap();
+    s.set_retention(Retention { level: Level::Full, full_days: 30, snippet_days: 365 }).unwrap();
+    ingest(&mut s, "f", vec![turn("no, that's wrong", "You're right.", "m", &now_ts(60), "old")]);
+    assert!(wal() > 0, "the text is in the log before the sweep");
+    assert_eq!(s.sweep(Utc::now()).unwrap(), 1, "the full text was stepped down");
+    assert_eq!(wal(), 0, "and no copy of it lingers in the log");
+}
+
+#[test]
+fn redaction_runs_again_over_the_corrections_a_rebuild_set_aside() {
+    let mut s = Store::open_in_memory().unwrap();
+    s.set_retention(Retention { level: Level::Full, ..Retention::default() }).unwrap();
+    ingest(&mut s, "f", vec![turn("no, that's wrong", "You're right.", "m", &now_ts(1), "k")]);
+    // A rebuild's set-aside copy, written with an older, weaker redaction.
+    s.conn.execute("INSERT INTO kept_events SELECT * FROM events", []).unwrap();
+    s.conn.execute("DELETE FROM events", []).unwrap();
+    s.conn
+        .execute("UPDATE kept_events SET human_full = 'mail pat@example.com', human_snip = 'mail pat@example.com'", [])
+        .unwrap();
+    s.set_setting(REDACTION_KEY, "1").unwrap();
+    s.redact_stored_text().unwrap();
+    let (full, snip): (String, String) =
+        s.conn.query_row("SELECT human_full, human_snip FROM kept_events", [], |r| Ok((r.get(0)?, r.get(1)?))).unwrap();
+    assert_eq!((full.as_str(), snip.as_str()), ("mail [email]", "mail [email]"));
+}
+
+// ---- Forget touches only its own source -------------------------------------------------------
+
+/// Every table with a `source` column (the sources table itself keeps it as `id`), with its columns
+/// and declared types, read from the schema so a table added later is covered too.
+fn tables_with_a_source(s: &Store) -> Vec<(String, Vec<(String, String)>)> {
+    let names: Vec<String> = {
+        let mut q = s
+            .conn
+            .prepare("SELECT name FROM sqlite_master WHERE type = 'table' AND name NOT LIKE 'sqlite_%' ORDER BY name")
+            .unwrap();
+        q.query_map([], |r| r.get(0)).unwrap().collect::<rusqlite::Result<_>>().unwrap()
+    };
+    names
+        .into_iter()
+        .filter_map(|name| {
+            let mut q = s.conn.prepare(&format!("PRAGMA table_info({name})")).unwrap();
+            let cols: Vec<(String, String)> = q
+                .query_map([], |r| Ok((r.get(1)?, r.get::<_, String>(2)?.to_uppercase())))
+                .unwrap()
+                .collect::<rusqlite::Result<_>>()
+                .unwrap();
+            cols.iter().any(|(c, _)| c == "source").then_some((name, cols))
+        })
+        .collect()
+}
+
+/// One made-up row for `source` in every table that has a `source` column, each value different.
+fn seed_every_source_table(s: &Store, source: &str) {
+    for (n, (table, cols)) in tables_with_a_source(s).into_iter().enumerate() {
+        let values: Vec<String> = cols
+            .iter()
+            .map(|(col, ty)| match (col.as_str(), ty.contains("INT")) {
+                ("source", _) => format!("'{source}'"),
+                (_, true) => format!("{}", 1000 + n * 10 + usize::from(source == "b")),
+                _ => format!("'{col}-{table}-{source}'"),
+            })
+            .collect();
+        s.conn.execute(&format!("INSERT INTO {table} VALUES ({})", values.join(", ")), []).unwrap();
+    }
+}
+
+fn rows_of(s: &Store, table: &str, source: &str) -> i64 {
+    s.conn.query_row(&format!("SELECT COUNT(*) FROM {table} WHERE source = ?1"), [source], |r| r.get(0)).unwrap()
+}
+
+#[test]
+fn forgetting_one_source_leaves_every_row_of_the_other() {
+    let mut s = Store::open_in_memory().unwrap();
+    for id in ["a", "b"] {
+        s.allow(id, None).unwrap();
+        seed_every_source_table(&s, id);
+    }
+    let named = [
+        "events",
+        "daily",
+        "sessions",
+        "tasks",
+        "seen",
+        "tips",
+        "tip_chats",
+        "report_cards",
+        "files",
+        "rebuild_read",
+        "kept_events",
+        "kept_daily",
+        "kept_sessions",
+        "kept_tasks",
+        "kept_seen",
+        "kept_tips",
+        "kept_tip_chats",
+    ];
+    for table in named {
+        assert_eq!((rows_of(&s, table, "a"), rows_of(&s, table, "b")), (1, 1), "{table} was seeded");
+    }
+    s.forget("a").unwrap();
+    for table in named {
+        assert_eq!((rows_of(&s, table, "a"), rows_of(&s, table, "b")), (0, 1), "{table}");
+    }
+    let ids: Vec<String> = {
+        let mut q = s.conn.prepare("SELECT id FROM sources").unwrap();
+        q.query_map([], |r| r.get(0)).unwrap().collect::<rusqlite::Result<_>>().unwrap()
+    };
+    assert_eq!(ids, ["b"], "b's consent stays");
+}
+
+#[test]
+fn forget_clears_the_source_in_every_table_that_has_one_and_nothing_else() {
+    let mut s = Store::open_in_memory().unwrap();
+    for id in ["a", "b"] {
+        s.allow(id, None).unwrap();
+        seed_every_source_table(&s, id);
+    }
+    let tables = tables_with_a_source(&s);
+    assert!(
+        tables.len() >= 15,
+        "the schema has many tables with a source: {:?}",
+        tables.iter().map(|t| &t.0).collect::<Vec<_>>()
+    );
+    s.forget("a").unwrap();
+    for (table, _) in &tables {
+        assert_eq!(rows_of(&s, table, "a"), 0, "{table} still has rows from the forgotten source");
+        assert_eq!(rows_of(&s, table, "b"), 1, "{table} lost rows from the other source");
+    }
+}
+
+// ---- consent cutoffs ----------------------------------------------------------------------------
+
+#[test]
+fn a_message_on_the_backfill_day_is_kept_and_one_just_before_it_is_not() {
+    let day = NaiveDate::from_ymd_opt(2026, 6, 15).unwrap();
+    let midnight =
+        Local.from_local_datetime(&day.and_hms_opt(0, 0, 0).unwrap()).earliest().unwrap().with_timezone(&Utc);
+    let state = SourceState { backfill_since: Some(day), ..test_source() };
+    assert!(state.wants(midnight), "the first moment of the day counts");
+    assert!(state.wants(midnight + Duration::hours(12)));
+    assert!(!state.wants(midnight - Duration::seconds(1)), "the last moment of the day before does not");
+}
+
+#[test]
+fn a_message_at_the_from_now_on_moment_is_kept_and_one_just_before_it_is_not() {
+    let at = Utc.with_ymd_and_hms(2026, 6, 15, 12, 0, 0).unwrap();
+    let state = SourceState { start_after: Some(at), ..test_source() };
+    assert!(!state.wants(at - Duration::seconds(1)));
+    assert!(state.wants(at), "exactly at the moment counts");
+    assert!(state.wants(at + Duration::seconds(1)));
+}
diff --git a/plugins/grade/.claude-plugin/plugin.json b/plugins/grade/.claude-plugin/plugin.json
index d272e54..293f702 100644
--- a/plugins/grade/.claude-plugin/plugin.json
+++ b/plugins/grade/.claude-plugin/plugin.json
@@ -1,6 +1,6 @@
 {
   "name": "grade",
-  "version": "0.5.6",
+  "version": "0.5.7",
   "description": "Report cards on how you prompt: type /grade:me for a letter grade, why, and how to ask so it's clear the first time. Works with the No, Actually... app to track your progress.",
   "author": {
     "name": "HTMLGuy LLC",