diff --git a/.agent/phases/todo/20_sources_midstream_bug/00_phase.md b/.agent/phases/todo/20_sources_midstream_bug/00_phase.md
new file mode 100644
index 0000000..06287da
--- /dev/null
+++ b/.agent/phases/todo/20_sources_midstream_bug/00_phase.md
@@ -0,0 +1,122 @@
+# Phase 20 — Sources Mid-Stream: an in-flight answer is not lost on navigation
+
+**Source:** `TODO.md` L3 — *"Clicking "sources" while chat is generating
+clears chat and result will never show up"*
+**Story:** `.agent/user_stories/sources-midstream.md` (created by task 02)
+**Context:** `frontend/assets/app.js` — the phase-14 persistence block
+(`STORAGE_KEY = "bor.chat.v1"`, `conversation`, `saveConversation`,
+`rememberBrainTurn`), the turn state machine (`UI_STATE.thinking` /
+`.streaming`), the streaming accumulators (`acc` / `thinkingAcc` /
+`sawThinking`), and the phase-14 restore path; `frontend/index.html`
+(`#nav-sources` link); `frontend/assets/header.js` (`clearChatStorage` —
+the deliberate New-Chat clear, NOT this bug).
+
+## Objective
+When the user leaves the chat page (the Sources nav link, the document
+viewer, any link) while a turn is still in flight, the answer generated so
+far must not vanish. Today the brain message is persisted only on `done`,
+so navigating away aborts the stream and the partial answer is lost — the
+user returns to their own question with no result, ever. After this phase,
+returning to the chat shows the question **and** the partial answer that
+had already streamed (rendered like any brain message, thinking block
+restored if any).
+
+## Owner-confirmed (2026-08-24, roadmap A1)
+1. **A partial answer is persisted as a plain brain message** — no
+   "(partial)" marker, no sources/suggestions (the turn is dead; the user
+   can re-ask for the full answer).
+2. Navigation **before the first answer token** (pure thinking) persists
+   nothing brain-side: the question is restored, no empty/partial bubble.
+3. The deliberate **New Chat** `clearChatStorage()` (sources/viewer pages)
+   is untouched — that clear is by design (phase 14/19).
+4. No server-side resume (A10 stays stateless) and no
+   "leave page?" confirmation dialog.
+
+## Design
+- **`app.js` — one new `pagehide` handler** (`window.addEventListener(
+  "pagehide", …)` — fires on navigate-away and bfcache store):
+  - Guard: only when a turn is in flight (current `uiState` is
+    `UI_STATE.thinking` or `UI_STATE.streaming`) **and** `acc` is
+    non-empty.
+  - Action: `rememberBrainTurn(acc, { thinking: thinkingAcc ||
+    undefined })` — reuse the existing save-point helper, so the partial
+    text is stored raw (the restore path re-renders through the
+    escape-first markdown renderer; the phase-17 `thinking` field
+    restores the collapsed Thinking block).
+  - **Idempotency guard:** a turn-local `persistedOnLeave` flag so a
+    second `pagehide` (or bfcache store+restore churn) never appends the
+    same partial message twice. The `done` save point is unaffected
+    (navigation means the stream is dead; if the user returns via
+    bfcache the turn is already aborted by the unloading page).
+- **Restore path:** unchanged — a stored partial message is a well-formed
+  brain message and renders exactly like a completed one (minus
+  sources/deflection, which it simply doesn't carry).
+- **Non-goals:** no resume of the SSE stream, no API changes, no changes
+  to the New Chat buttons, sign-out, or the document-viewer back link.
+
+## Dependencies
+- `14_chat_persistence` (complete) — `bor.chat.v1` shape, save points,
+  restore, and the `rememberBrainTurn` helper this phase reuses.
+- `17_thinking_display` (complete) — the `thinking` field on persisted
+  brain messages and the `thinkingAcc` accumulator.
+- `19_shared_header` (complete) — the `#nav-sources` link (admin-only)
+  the bug report clicks.
+- `18_follow_bottom_scroll` (complete) — no overlap (scroll gating only).
+
+## Tasks
+1. `01_persist_inflight_turn.md` — the `pagehide` partial-persistence
+   handler in `app.js` + source-level unit pins.
+2. `02_e2e_story_suite_commit.md` — `tests/e2e/test_sources_midstream_bug.py`
+   (the story gate, isolated), regression suites, story file, final
+   validation, the single atomic commit, phase move to `complete/`.
+
+## Locked decisions
+- **A10 untouched** — API stays stateless; no resume. **A11 untouched** —
+  vanilla JS, no CDN. **A16 honored** — one new story E2E suite +
+  adapted regressions. No anchor changed.
+
+## Testing & Quality
+- **Unit (source-level, new `tests/unit/test_sources_midstream.py`,
+  following the repo's source-pin pattern):** `app.js` registers a
+  `pagehide` listener; the guard references the in-flight `uiState` and a
+  non-empty `acc`; the partial path calls `rememberBrainTurn` with
+  `thinking: thinkingAcc || undefined`; a turn-local idempotency flag
+  exists; `STORAGE_KEY`/save-point comments updated to list the new
+  save point.
+- **Integration:** none (no `app/` changes) — the `uv run pytest
+  --cov=app` number must stay at today's.
+- **Coverage:** frontend-only; the >90% `app/` gate is unaffected,
+  re-run to prove it.
+- **E2E:** `tests/e2e/test_sources_midstream_bug.py` (task 02), green
+  **in isolation** (prereq `podman compose up -d db`).
+- **Lint/types:** `uv run ruff check . && uv run pyright` clean.
+
+## Completion Criteria
+- [ ] Admin, mid-stream, clicks **Sources** → returns to `/`: the
+      question **and** the already-streamed partial answer are both
+      rendered; no error banner; `bor.chat.v1` holds the partial brain
+      message.
+- [ ] Navigate away before the first token → back: question restored,
+      no empty/partial brain bubble.
+- [ ] A completed turn is persisted exactly as before (sources,
+      deflection, suggestions intact).
+- [ ] New Chat from the sources page still clears the conversation.
+- [ ] `uv run pytest` green; `uv run pytest --cov=app
+      --cov-report=term-missing` ≥ today's number.
+- [ ] `uv run pytest tests/e2e/test_sources_midstream_bug.py -v --no-cov`
+      green in isolation; regressions green in isolation (one command
+      each): `test_chat_persistence.py`, `test_thinking_display.py`,
+      `test_shared_header.py`.
+- [ ] `uv run ruff check . && uv run pyright` clean.
+- [ ] UI Structure Check (AGENTS.md rule 5): no new UI surface — the
+      restored partial renders through the existing bubble/thinking
+      contract.
+- [ ] `.agent/user_stories/sources-midstream.md` exists.
+- [ ] One `--no-gpg-sign` commit (below);
+      `.agent/phases/todo/20_sources_midstream_bug/` moved to
+      `.agent/phases/complete/`.
+
+## Commit
+```bash
+git add -A .agent/ frontend/ tests/ && git commit --no-gpg-sign -m "fix(chat): keep the in-flight answer when navigating away mid-turn — partial answer restored on return"
+```
diff --git a/.agent/phases/todo/20_sources_midstream_bug/01_persist_inflight_turn.md b/.agent/phases/todo/20_sources_midstream_bug/01_persist_inflight_turn.md
new file mode 100644
index 0000000..0ef26bc
--- /dev/null
+++ b/.agent/phases/todo/20_sources_midstream_bug/01_persist_inflight_turn.md
@@ -0,0 +1,83 @@
+# Task 01 — app.js: persist the partial answer on navigate-away (pagehide)
+
+**Phase:** `20_sources_midstream_bug` · **Source:** `TODO.md` L3 —
+*"Clicking "sources" while chat is generating clears chat and result will
+never show up"*
+
+## Objective
+`frontend/assets/app.js` gains exactly one new save point: on `pagehide`
+(navigate-away / bfcache), if a turn is in flight and answer text has
+already streamed, the partial answer is persisted via the existing
+`rememberBrainTurn` helper — so returning to the chat restores the
+question **and** what had been generated.
+
+## Work
+1. `frontend/assets/app.js`
+   - Next to the other turn-local state (around the `acc` /
+     `thinkingAcc` / `sawThinking` / `aborted` declarations, ~line 920+),
+     declare a turn-local flag:
+     ```js
+     let persistedOnLeave = false; // pagehide partial-persist at most once
+     ```
+     and reset it to `false` at the top of `runTurn` (where `acc`,
+     `thinkingAcc`, `sawThinking`, `wrap` are initialized), so it is
+     turn-scoped like the rest.
+   - Register one handler at module boot (next to the other
+     `window.addEventListener` calls):
+     ```js
+     /* Navigate-away save point (phase 20, owner choice 2026-08-24 A1):
+      * leaving the chat mid-turn would otherwise drop the in-flight
+      * answer — the brain message persists only on `done`, and
+      * navigation aborts the stream. On `pagehide`, if a turn is in
+      * flight and answer text has streamed, persist the partial raw text
+      * (reusing the save-point helper, so restore re-renders it exactly
+      * like a completed answer — no "(partial)" marker, no sources).
+      * Thinking-only (no answer tokens yet) persists nothing brain-side:
+      * the question is already saved on send and the user can re-ask.
+      * `persistedOnLeave` makes this idempotent across pagehide/bfcache
+      * churn. */
+     window.addEventListener("pagehide", () => {
+       if (persistedOnLeave) return;
+       if (uiState !== UI_STATE.thinking && uiState !== UI_STATE.streaming)
+         return;
+       if (!acc) return; // nothing brain-side to save yet
+       persistedOnLeave = true;
+       rememberBrainTurn(acc, { thinking: thinkingAcc || undefined });
+     });
+     ```
+     **ASSUMPTION (owner-confirmed A1):** the partial is a plain brain
+     message (no marker, no sources); the variables `uiState`, `acc`,
+     `thinkingAcc` are the turn's existing ones — match the names actually
+     in scope (the file keeps `let acc = ""` / `let thinkingAcc = ""`
+     turn-locals inside `runTurn`; if the handler needs them outside that
+     scope, hoist the turn-locals to module scope *without* changing any
+     behavior — smallest diff wins).
+   - Update the phase-14 persistence block comment (~line 631): the
+     "Save points:" sentence now lists three — the user message on send,
+     the brain message on `done`, and the **partial** brain message on
+     navigate-away (`pagehide`, phase 20).
+2. `tests/unit/test_sources_midstream.py` (new — follow the repo's
+   source-level pin pattern used by `tests/unit/test_shared_header.py`):
+   - `app.js` contains `addEventListener("pagehide"` exactly once.
+   - The pagehide body is guarded by the in-flight states (`thinking`
+     and `streaming`) and a non-empty `acc` check.
+   - The pagehide body calls `rememberBrainTurn(acc,` with
+     `thinking: thinkingAcc || undefined`.
+   - `persistedOnLeave` is declared and reset in `runTurn`.
+   - The persistence block comment lists the `pagehide` save point.
+   - `clearChatStorage` (header.js) is untouched — still the only
+     deliberate clear.
+
+## Testing & Quality
+- `uv run pytest tests/unit/test_sources_midstream.py -v` green.
+- `uv run ruff check . && uv run pyright` clean.
+- Manual smoke (dev server, DEBUGPY optional): send a question against
+  the real LLM (or any slow stream), click Sources mid-stream, return —
+  the partial answer is visible.
+
+## Completion Criteria
+- [ ] The `pagehide` handler exists, is turn-scoped and idempotent, and
+      reuses `rememberBrainTurn` (no duplicated storage code).
+- [ ] No other save point changed: send + `done` behave byte-identically
+      to before (existing persistence unit pins still pass).
+- [ ] Unit pins green; lint/types clean.
diff --git a/.agent/phases/todo/20_sources_midstream_bug/02_e2e_story_suite_commit.md b/.agent/phases/todo/20_sources_midstream_bug/02_e2e_story_suite_commit.md
new file mode 100644
index 0000000..28933f1
--- /dev/null
+++ b/.agent/phases/todo/20_sources_midstream_bug/02_e2e_story_suite_commit.md
@@ -0,0 +1,76 @@
+# Task 02 — E2E story suite, story file, validation, commit
+
+**Phase:** `20_sources_midstream_bug` · **Source:** `TODO.md` L3
+
+## Objective
+The story gate: `tests/e2e/test_sources_midstream_bug.py` proves the bug
+is fixed end-to-end (navigate away mid-stream, come back, the partial
+answer is there), plus the regression suites, the story file, final
+validation, and the single atomic commit.
+
+## Work
+1. `tests/e2e/test_sources_midstream_bug.py` (new — mirror
+   `test_chat_persistence.py`'s scaffolding: fixture import via
+   `_import_fixtures`/`_run_in_thread`, mock LLM on a thread, `login`
+   from `e2e.auth_helpers`). The mock LLM must stream **slowly enough**
+   that the turn is still in flight when the test navigates (reuse the
+   streaming pattern from `test_thinking_display.py` / `mock_llm.py`;
+   tune the per-chunk delay until the navigation lands mid-stream).
+   Tests (Playwright Mapping Rule — one per numbered scenario):
+   1. `test_partial_answer_survives_sources_nav_midstream` — admin
+      (`login(page, app_url, next="/")`), send a question, wait for the
+      first streamed chunk to render (expect the first chunk's text in
+      the answer bubble), **click `#nav-sources`** (the actual nav link —
+      admin sees it), land on `/sources.html`, then `page.goto("/")`:
+      expect the question text AND the first-chunk text present, no
+      `role="alert"` banner. Read `localStorage` `bor.chat.v1`: the
+      messages contain a brain message whose text starts with the first
+      chunk.
+   2. `test_no_orphan_brain_message_when_navigated_before_first_token` —
+      mock streams a `thinking` event, then a long pre-token pause;
+      navigate (direct `page.goto("/sources.html")` is fine here) during
+      the pause, return to `/`: the question is present, exactly one
+      user message and **zero** brain messages in both the DOM and
+      `bor.chat.v1`.
+   3. `test_completed_turn_unaffected` — a turn that finishes normally
+      (`done`), then navigate to sources and back: full answer, sources
+      chips, and the done-metadata (sources array) intact in storage.
+   4. `test_new_chat_still_clears_conversation` — regression: completed
+      turn → `/sources.html` → click the sources-page New Chat button →
+      lands on `/` with the empty state and `bor.chat.v1` removed.
+2. `.agent/user_stories/sources-midstream.md` (new) — the short story
+   file matching the repo's story format (goal, the bug report verbatim
+   from `TODO.md` L3, the owner-confirmed A1 decisions from
+   `00_phase.md`, the E2E mapping table test-name → scenario).
+3. Run the suite **in isolation** (prereq `podman compose up -d db`):
+   `uv run pytest tests/e2e/test_sources_midstream_bug.py -v --no-cov`.
+4. Regressions, in isolation, one command each (all must stay green):
+   - `uv run pytest tests/e2e/test_chat_persistence.py -v --no-cov`
+   - `uv run pytest tests/e2e/test_thinking_display.py -v --no-cov`
+   - `uv run pytest tests/e2e/test_shared_header.py -v --no-cov`
+5. Final validation: `uv run pytest` green; `uv run pytest --cov=app
+   --cov-report=term-missing` ≥ today's number (>90% gate);
+   `uv run ruff check . && uv run pyright` clean.
+6. **UI Structure Check** (AGENTS.md rule 5): the restored partial
+   renders through the existing bubble/thinking contract — no new
+   surface, no new ids, focus/contrast unchanged.
+7. Write the phase report
+   (`.agent/reports/20_sources_midstream_bug/` — what changed, E2E
+   results, the manual-smoke note from task 01).
+8. Commit (one atomic commit) and move the phase:
+   ```bash
+   git add -A .agent/ frontend/ tests/
+   git commit --no-gpg-sign -m "fix(chat): keep the in-flight answer when navigating away mid-turn — partial answer restored on return"
+   mv .agent/phases/todo/20_sources_midstream_bug .agent/phases/complete/
+   ```
+
+## Testing & Quality
+- Story suite green **in isolation**; the three regression suites green
+  in isolation; full unit+integration suite green; `app/` coverage at or
+  above today's number (>90%); ruff + pyright clean.
+
+## Completion Criteria
+- [ ] `test_sources_midstream_bug.py` 4/4 in isolation.
+- [ ] Regressions (persistence, thinking display, shared header) green.
+- [ ] Story file + phase report exist.
+- [ ] One `--no-gpg-sign` commit; phase directory in `complete/`.
diff --git a/.agent/phases/todo/21_thinking_no_scroll/00_phase.md b/.agent/phases/todo/21_thinking_no_scroll/00_phase.md
new file mode 100644
index 0000000..51a4a47
--- /dev/null
+++ b/.agent/phases/todo/21_thinking_no_scroll/00_phase.md
@@ -0,0 +1,104 @@
+# Phase 21 — Thinking Window: No Scroll Back, Just Live
+
+**Source:** `TODO.md` L4 — *"Disable scroll in the thinking window.
+Users don't need to scroll back through thinking, just see it live."*
+**Story:** `.agent/user_stories/thinking-no-scroll.md` (created by task 02)
+**Context:** `frontend/assets/styles.css` —
+`details.thinking .thinking-text` (today: `max-height: 320px;
+overflow-y: auto;`, i.e. a user-scrollable 320px window, ~line 444);
+`frontend/assets/app.js` — the phase-17 thinking block (`ensureThinkingBlock`,
+the streaming `thinking` branch that appends chunks and already pins the
+stream to the bottom: `textEl.scrollTop = textEl.scrollHeight`, ~line 886),
+and the phase-14 restore path (stored `thinking` re-renders a
+**collapsed** block).
+
+## Objective
+The live Thinking block is a scratchpad, not a transcript. The user must
+not be able to scroll back through it — the 320px window always shows the
+**live tail** of the reasoning stream (the existing per-chunk
+bottom-pinning stays). Wheel, drag, and keyboard scrolling on
+`.thinking-text` stop working; the stream itself keeps pinning to the
+bottom as chunks arrive.
+
+## Owner-confirmed (2026-08-24, roadmap A2)
+1. **Keep the 320px clip** — "just see it live" means the window stays a
+   fixed 320px viewport showing the newest lines; no auto-height growth,
+   no "↓ more" affordance.
+2. **The answer bubble is untouched** — final answers keep their existing
+   scroll behavior (phase 11 long answers).
+3. **Restored (collapsed) Thinking blocks are untouched** — the phase-14
+   restore renders them collapsed, where overflow is moot.
+
+## Design
+- **CSS (the whole functional change):**
+  `details.thinking .thinking-text` — `overflow-y: auto` →
+  `overflow-y: hidden`; keep `max-height: 320px`.
+  `overflow: hidden` still permits **programmatic** scrolling
+  (`scrollTop`), so the phase-17 pin
+  (`textEl.scrollTop = textEl.scrollHeight` on every thinking chunk)
+  keeps the window glued to the live tail — no JS change needed.
+  Add a CSS comment: *no user scroll back (owner choice 2026-08-24):
+  the window is a live tail only — the JS bottom-pin is the sole
+  scroller*.
+- **No JS change** — the pin already exists; nothing else touches
+  `.thinking-text` scroll.
+- **Non-goals:** no change to the answer bubble, the collapsed restore
+  state, the summary/chevron, or the auto-collapse on first delta
+  (phase 17).
+
+## Dependencies
+- `17_thinking_display` (complete) — the block, the pin, the restore.
+- `18_follow_bottom_scroll` (complete) — no overlap (chat-page scroll
+  gate only; the thinking window is a separate inner element).
+- `11_long_answers` (complete) — the untouched answer-bubble behavior.
+
+## Tasks
+1. `01_disable_thinking_scroll.md` — the CSS change + source-level unit
+   pins.
+2. `02_e2e_story_suite_commit.md` — `tests/e2e/test_thinking_no_scroll.py`
+   (the story gate, isolated), regression suites, story file, final
+   validation, the single atomic commit, phase move to `complete/`.
+
+## Locked decisions
+- **A11 untouched** — no CDN, pure CSS. **A16 honored** — one new story
+  E2E suite + adapted regressions. No anchor changed.
+
+## Testing & Quality
+- **Unit (source-level, new `tests/unit/test_thinking_no_scroll.py`,
+  repo source-pin pattern):** `styles.css` carries
+  `details.thinking .thinking-text` with `overflow-y: hidden` and
+  `max-height: 320px`; the phase-17 pin line
+  (`textEl.scrollTop = textEl.scrollHeight`) still present in `app.js`
+  (the live-tail mechanism must not be lost).
+- **Integration:** none (no `app/` changes).
+- **Coverage:** frontend-only; the >90% `app/` gate is unaffected.
+- **E2E:** `tests/e2e/test_thinking_no_scroll.py` (task 02), green
+  **in isolation** (prereq `podman compose up -d db`).
+- **Lint/types:** `uv run ruff check . && uv run pyright` clean.
+
+## Completion Criteria
+- [ ] With a long thinking stream, wheel/mouse-drag/keyboard on
+      `.thinking-text` do **not** move it; the visible content is always
+      the live tail (`scrollTop === scrollHeight` after each chunk,
+      within 1px).
+- [ ] Computed style: `overflow-y: hidden`, `max-height: 320px`.
+- [ ] A long **answer** bubble still scrolls normally; a restored
+      collapsed Thinking block still renders (phase 17 regression).
+- [ ] `uv run pytest` green; `uv run pytest --cov=app
+      --cov-report=term-missing` ≥ today's number.
+- [ ] `uv run pytest tests/e2e/test_thinking_no_scroll.py -v --no-cov`
+      green in isolation; regressions green in isolation:
+      `test_thinking_display.py`, `test_long_answers.py`.
+- [ ] `uv run ruff check . && uv run pyright` clean.
+- [ ] UI Structure Check (AGENTS.md rule 5): no new surface; the block
+      keeps its focus-visible summary, aria contract, and reduced-motion
+      behavior.
+- [ ] `.agent/user_stories/thinking-no-scroll.md` exists.
+- [ ] One `--no-gpg-sign` commit (below);
+      `.agent/phases/todo/21_thinking_no_scroll/` moved to
+      `.agent/phases/complete/`.
+
+## Commit
+```bash
+git add -A .agent/ frontend/ tests/ && git commit --no-gpg-sign -m "fix(ui): thinking window no longer scrolls — live 320px view pinned to the stream tail"
+```
diff --git a/.agent/phases/todo/21_thinking_no_scroll/01_disable_thinking_scroll.md b/.agent/phases/todo/21_thinking_no_scroll/01_disable_thinking_scroll.md
new file mode 100644
index 0000000..0a7ba6c
--- /dev/null
+++ b/.agent/phases/todo/21_thinking_no_scroll/01_disable_thinking_scroll.md
@@ -0,0 +1,57 @@
+# Task 01 — styles.css: `.thinking-text` overflow hidden (no user scroll)
+
+**Phase:** `21_thinking_no_scroll` · **Source:** `TODO.md` L4 —
+*"Disable scroll in the thinking window. Users don't need to scroll back
+through thinking, just see it live."*
+
+## Objective
+One CSS property change makes the Thinking window a live-tail-only view:
+`overflow-y: hidden` instead of `auto`, keeping the 320px clip. The
+phase-17 JS bottom-pin (which keeps working under `overflow: hidden`) is
+the sole scroller.
+
+## Work
+1. `frontend/assets/styles.css` — in the phase-17 thinking block
+   section (~line 444):
+   ```css
+   details.thinking .thinking-text {
+     padding: 0 0.75rem 0.75rem;
+     color: var(--ink-soft); /* 6.9:1 on --surface */
+     font-size: 0.875rem;
+     line-height: 1.55;
+     max-height: 320px;
+     overflow-y: hidden; /* no scroll back (owner choice 2026-08-24):
+                            the window is a live tail only — the phase-17
+                            JS bottom-pin (scrollTop = scrollHeight per
+                            chunk) is the sole scroller */
+   }
+   ```
+   (Only the `overflow-y` value + comment change; every other declaration
+   stays byte-identical.)
+2. `frontend/assets/app.js` — **no change expected.** Verify the pin is
+   intact: the streaming `thinking` branch still does
+   `textEl.scrollTop = textEl.scrollHeight` on every chunk (~line 886).
+   If (and only if) the pin were missing/broken, fix it — do not remove
+   or alter any other scrolling behavior.
+3. `tests/unit/test_thinking_no_scroll.py` (new — repo source-pin
+   pattern):
+   - `styles.css`: the `details.thinking .thinking-text` rule contains
+     `overflow-y: hidden` and `max-height: 320px` (no `overflow-y: auto`
+     left in that rule).
+   - `app.js`: the bottom-pin line
+     `textEl.scrollTop = textEl.scrollHeight` is still present (the
+     live-tail mechanism).
+4. Manual smoke (dev server): stream a long thinking turn; try to wheel /
+   drag / Tab+ArrowDown inside the Thinking block — it must not move;
+   the newest chunk is always the one visible at the bottom.
+
+## Testing & Quality
+- `uv run pytest tests/unit/test_thinking_no_scroll.py -v` green.
+- `uv run ruff check . && uv run pyright` clean.
+
+## Completion Criteria
+- [ ] `.thinking-text` is `overflow-y: hidden`, `max-height: 320px`, with
+      the owner-choice comment.
+- [ ] The JS bottom-pin is verified intact (no app.js diff unless the
+      pin was broken).
+- [ ] Unit pins green; lint/types clean; manual smoke passed.
diff --git a/.agent/phases/todo/21_thinking_no_scroll/02_e2e_story_suite_commit.md b/.agent/phases/todo/21_thinking_no_scroll/02_e2e_story_suite_commit.md
new file mode 100644
index 0000000..e92d2c7
--- /dev/null
+++ b/.agent/phases/todo/21_thinking_no_scroll/02_e2e_story_suite_commit.md
@@ -0,0 +1,70 @@
+# Task 02 — E2E story suite, story file, validation, commit
+
+**Phase:** `21_thinking_no_scroll` · **Source:** `TODO.md` L4
+
+## Objective
+The story gate: `tests/e2e/test_thinking_no_scroll.py` proves the window
+can't be user-scrolled but always tracks the live tail, plus regressions,
+story file, final validation, and the single atomic commit.
+
+## Work
+1. `tests/e2e/test_thinking_no_scroll.py` (new — reuse
+   `test_thinking_display.py`'s mock-LLM streaming scaffolding; the mock
+   must stream a **long** thinking body, in many chunks, so
+   `.thinking-text` overflow exceeds its 320px box). Tests:
+   1. `test_thinking_window_not_user_scrollable` — open the block, wait
+      until `scrollHeight > clientHeight`; focus `.thinking-text`
+      (`el.focus()`), dispatch mouse wheel over it
+      (`page.mouse.wheel(0, -200)` after moving the mouse over the
+      element) and press `Home`/`ArrowUp`: `scrollTop` must not decrease
+      (assert `scrollTop` unchanged within 1px between actions).
+   2. `test_thinking_window_tracks_live_tail` — while chunks stream,
+      after the 2nd-to-last and last chunk:
+      `scrollTop === scrollHeight` (within 1px) — the visible window is
+      the live tail; the **last** chunk's text is within the visible
+      rectangle (its offsetTop + scrollTop geometry check, or
+      `elementFromPoint` at the box's bottom).
+   3. `test_thinking_window_css_contract` — computed style of
+      `.thinking-text`: `overflow-y === "hidden"`,
+      `max-height === "320px"`.
+   4. `test_answer_bubble_still_scrollable` (regression, phase 11) — a
+      long answer (use the long-answer mock from
+      `test_long_answers.py`): the answer bubble is still
+      user-scrollable (scrollTop moves on wheel) and
+      `overflow-y` is not `hidden` there.
+   5. `test_restored_collapsed_thinking_unaffected` (regression,
+      phase 17) — a turn with stored `thinking`, reload: the collapsed
+      Thinking block renders with its text (existing pin from
+      `test_thinking_display.py` — replicate, don't duplicate the file).
+2. `.agent/user_stories/thinking-no-scroll.md` (new) — story file per
+   the repo format: goal, the bug report verbatim from `TODO.md` L4, the
+   owner-confirmed A2 decisions from `00_phase.md`, E2E mapping table.
+3. Run the suite **in isolation** (prereq `podman compose up -d db`):
+   `uv run pytest tests/e2e/test_thinking_no_scroll.py -v --no-cov`.
+4. Regressions, in isolation, one command each:
+   - `uv run pytest tests/e2e/test_thinking_display.py -v --no-cov`
+   - `uv run pytest tests/e2e/test_long_answers.py -v --no-cov`
+5. Final validation: `uv run pytest` green; `uv run pytest --cov=app
+   --cov-report=term-missing` ≥ today's number (>90% gate);
+   `uv run ruff check . && uv run pyright` clean.
+6. **UI Structure Check** (AGENTS.md rule 5): no new surface; the block
+   keeps its summary chevron, focus-visible ring, aria-live/label
+   contract, and the reduced-motion stillness (styles.css ~line 686).
+7. Write the phase report (`.agent/reports/21_thinking_no_scroll/`).
+8. Commit (one atomic commit) and move the phase:
+   ```bash
+   git add -A .agent/ frontend/ tests/
+   git commit --no-gpg-sign -m "fix(ui): thinking window no longer scrolls — live 320px view pinned to the stream tail"
+   mv .agent/phases/todo/21_thinking_no_scroll .agent/phases/complete/
+   ```
+
+## Testing & Quality
+- Story suite green **in isolation**; both regression suites green in
+  isolation; full unit+integration suite green; `app/` coverage at or
+  above today's number (>90%); ruff + pyright clean.
+
+## Completion Criteria
+- [ ] `test_thinking_no_scroll.py` 5/5 in isolation.
+- [ ] Regressions (thinking display, long answers) green in isolation.
+- [ ] Story file + phase report exist.
+- [ ] One `--no-gpg-sign` commit; phase directory in `complete/`.
diff --git a/.agent/phases/todo/22_background_animation/00_phase.md b/.agent/phases/todo/22_background_animation/00_phase.md
new file mode 100644
index 0000000..a9007b9
--- /dev/null
+++ b/.agent/phases/todo/22_background_animation/00_phase.md
@@ -0,0 +1,114 @@
+# Phase 22 — Animated Background: Make It Actually Animate
+
+**Source:** `TODO.md` L5 — *"Fix background animation not working, just
+blinking"*
+**Story:** `.agent/user_stories/background-animation.md` (created by task 02)
+**Context:** `frontend/assets/styles.css` — the phase-08 animated
+background block (~lines 60–100): `body::before` (44px drifting grid,
+1px lines at ~35% `--line` alpha, radial mask
+`radial-gradient(120% 90% at 50% 0%, black 25%, transparent 78%)`,
+`animation: bg-grid-drift 60s linear infinite`) and `body::after` (two
+soft radial glows, `animation: bg-glow-breathe 14s ease-in-out infinite
+alternate`, opacity 0.65↔1 + scale 1↔1.05). Both layers are
+`position: fixed; inset: 0; z-index: -1; pointer-events: none`. `html`
+owns the `var(--bg)` canvas and `body` is `background: transparent`
+(~lines 42–54) — if any later rule occludes that, the layers vanish.
+The **phase-08 design comments are the spec** for what "working" means.
+
+## Objective
+Owner report 2026-08-24: the background "just blinks" — i.e. the motion
+the phase-08 design promised (a slow, seamless grid drift + a gentle
+glow breathe) is not perceived; at most a flicker/blink is visible.
+Diagnose which layer(s) actually fail in a real Chromium viewport, fix
+the CSS, and leave a background that visibly and smoothly animates as
+designed — no blink, no static frame, no jank.
+
+## Owner-confirmed (2026-08-24, roadmap A3)
+1. **Intended effect = the phase-08 design comments:** seamless 60s grid
+   drift (one cell per loop) + 14s ease glow breathing. The fix serves
+   that design, not a redesign.
+2. **Pure CSS, zero JS** (phase-08 anchor) — no animation JS, no new
+   assets, no `filter: blur` (perf note in the block).
+
+## Design / diagnostic plan
+The fix is found, not guessed — work through this checklist in a real
+Chromium window (dev server, full page, ~15s of observation):
+1. **Per-layer visibility:** toggle each pseudo-element (DevTools
+   generated-content / a temp outline) and screenshot — is the grid
+   visible at all? Is only the glow (the "blink" the user perceives)
+   alive?
+2. **Grid layer:** sample `background-position` on `body::before` at two
+   timestamps — is it actually moving? Is the radial mask fading the
+   visible region so small that the 44px/60s drift is imperceptible?
+   (If the drift is real but too faint: raise the grid line alpha and/or
+   the mask's visible radius — smallest change that reads as "smooth
+   drift".)
+3. **Glow layer:** is the 14s breathe reading as a *blink*? (If the
+   opacity swing 0.65↔1 is perceived as pulsing: lengthen the period
+   and/or narrow the opacity delta so it reads as breathing.)
+4. **Occlusion check:** confirm nothing later in `styles.css` (or in
+   `html`/`body` rules) paints an opaque background over the
+   `z-index: -1` layers — the phase-08 comment at ~line 42 is the
+   contract.
+5. **Apply the fix in `styles.css`** — document the found root cause in
+   the phase report (screenshot before/after in
+   `.agent/screenshots/22_background_animation/`).
+
+## Dependencies
+- `08_story_dark_tech_theme` (complete) — owns the layers, the palette,
+  and the "pure CSS, zero JS" anchor this phase must respect.
+- `07_story_responsive_polish` (complete) — no new overflow at 360px
+  (both layers are `fixed; inset: 0` — keep it that way).
+
+## Tasks
+1. `01_fix_background_animation.md` — diagnosis + the CSS fix +
+   source-level unit pins.
+2. `02_e2e_story_suite_commit.md` — `tests/e2e/test_background_animation.py`
+   (the story gate, isolated), regression suites, story file, final
+   validation, the single atomic commit, phase move to `complete/`.
+
+## Locked decisions
+- **Phase-08 anchor honored** — pure CSS, zero JS, no `filter: blur`,
+  WCAG AA palette untouched (background layers carry no text).
+  **A11 untouched** — no CDN, no new assets. **A16 honored** — one new
+  story E2E suite + adapted regressions. No anchor changed.
+
+## Testing & Quality
+- **Unit (source-level, new `tests/unit/test_background_animation.py`,
+  repo source-pin pattern):** `styles.css` still defines
+  `@keyframes bg-grid-drift` and `@keyframes bg-glow-breathe`;
+  `body::before` animates `bg-grid-drift` with `linear infinite`;
+  `body::after` animates `bg-glow-breathe`; both layers remain
+  `position: fixed; z-index: -1; pointer-events: none`; `html` keeps
+  `background: var(--bg)` and `body` keeps `background: transparent`
+  (the no-occlusion contract). Pin the **final** values the fix lands
+  on (durations/opacities may move per the design plan).
+- **Integration:** none (no `app/` changes).
+- **Coverage:** frontend-only; the >90% `app/` gate is unaffected.
+- **E2E:** `tests/e2e/test_background_animation.py` (task 02), green
+  **in isolation** (prereq `podman compose up -d db`).
+- **Lint/types:** `uv run ruff check . && uv run pyright` clean.
+
+## Completion Criteria
+- [ ] In a real Chromium viewport, the background visibly and smoothly
+      animates (grid drift + glow breathe) — screenshot before/after in
+      the phase report; owner's "just blinking" perception gone.
+- [ ] Root cause documented in `.agent/reports/22_background_animation/`.
+- [ ] `uv run pytest` green; `uv run pytest --cov=app
+      --cov-report=term-missing` ≥ today's number.
+- [ ] `uv run pytest tests/e2e/test_background_animation.py -v --no-cov`
+      green in isolation; regressions green in isolation:
+      `test_dark_tech_theme.py`, `test_responsive_polish.py`.
+- [ ] `uv run ruff check . && uv run pyright` clean.
+- [ ] UI Structure Check (AGENTS.md rule 5): layers stay behind content
+      (`z-index: -1`, `pointer-events: none`), no text/contrast impact,
+      no 360px overflow.
+- [ ] `.agent/user_stories/background-animation.md` exists.
+- [ ] One `--no-gpg-sign` commit (below);
+      `.agent/phases/todo/22_background_animation/` moved to
+      `.agent/phases/complete/`.
+
+## Commit
+```bash
+git add -A .agent/ frontend/ tests/ && git commit --no-gpg-sign -m "fix(ui): animated background actually animates — grid drift and glow breathe per the phase-08 design"
+```
diff --git a/.agent/phases/todo/22_background_animation/01_fix_background_animation.md b/.agent/phases/todo/22_background_animation/01_fix_background_animation.md
new file mode 100644
index 0000000..22b9e41
--- /dev/null
+++ b/.agent/phases/todo/22_background_animation/01_fix_background_animation.md
@@ -0,0 +1,60 @@
+# Task 01 — Diagnose and fix the animated background (styles.css)
+
+**Phase:** `22_background_animation` · **Source:** `TODO.md` L5 —
+*"Fix background animation not working, just blinking"*
+
+## Objective
+Find why the phase-08 animated background reads as "just blinking" and
+fix `styles.css` so the grid drift and the glow breathe are both visibly
+and smoothly alive, per the phase-08 design comments (pure CSS, zero JS).
+
+## Work
+1. **Reproduce:** `uv run uvicorn app.main:app --reload` (db up), open `/`
+   in a visible Chromium window (Playwright or the interactive browser),
+   observe ≥15s. Note exactly what moves and what doesn't.
+2. **Diagnose per the `00_phase.md` checklist** (per-layer visibility
+   toggles, `background-position` samples on `body::before`, mask
+   inspection, occlusion check against `html`/`body` rules, glow
+   opacity-swing perception). Record findings + before-screenshot in
+   `.agent/reports/22_background_animation/` and
+   `.agent/screenshots/22_background_animation/`.
+   **ASSUMPTION (to verify, not assume):** the likely culprits, in
+   order — (a) the masked grid drift is too faint/slow to perceive,
+   (b) only the glow opacity swing is visible and it reads as a blink,
+   (c) a later rule occludes the `z-index: -1` layers. Confirm which
+   one actually fires before touching CSS; the fix must match the found
+   cause.
+3. **Fix in `frontend/assets/styles.css`** (smallest change that makes
+   the design read):
+   - grid: raise line alpha and/or the mask's visible radius and/or the
+     drift speed as needed for a clearly visible, seamless drift
+     (drift delta must still equal one 44px cell for a seamless loop —
+     if the speed changes, keep `background-position` 0→44px and only
+     move the duration);
+   - glow: if the breathe reads as a blink, narrow the opacity delta
+     (e.g. 0.8↔1) and/or lengthen the period — it must read as
+     breathing, not pulsing;
+   - keep: both layers `position: fixed; inset: 0; z-index: -1;
+     pointer-events: none`; no `filter: blur`; no JS; palette/contrast
+     untouched.
+4. **After-screenshot** (same viewport, two frames a few seconds apart
+   showing motion) into the same screenshots dir.
+5. `tests/unit/test_background_animation.py` (new — repo source-pin
+   pattern): pin the **final** `styles.css` values — both
+   `@keyframes` present, `body::before` → `bg-grid-drift linear
+   infinite`, `body::after` → `bg-glow-breathe`, both layers
+   `fixed`/`z-index: -1`/`pointer-events: none`, `html` keeps
+   `background: var(--bg)`, `body` keeps `background: transparent`.
+6. Manual re-verify: the "just blinking" perception is gone — smooth
+   drift + gentle breathe, no jank, no static frame.
+
+## Testing & Quality
+- `uv run pytest tests/unit/test_background_animation.py -v` green.
+- `uv run ruff check . && uv run pyright` clean.
+
+## Completion Criteria
+- [ ] Root cause documented (with before/after screenshots) in the
+      phase report dir.
+- [ ] Both layers visibly animate as the phase-08 design describes;
+      pure CSS, zero JS, no blur.
+- [ ] Unit pins green against the final values; lint/types clean.
diff --git a/.agent/phases/todo/22_background_animation/02_e2e_story_suite_commit.md b/.agent/phases/todo/22_background_animation/02_e2e_story_suite_commit.md
new file mode 100644
index 0000000..8124519
--- /dev/null
+++ b/.agent/phases/todo/22_background_animation/02_e2e_story_suite_commit.md
@@ -0,0 +1,68 @@
+# Task 02 — E2E story suite, story file, validation, commit
+
+**Phase:** `22_background_animation` · **Source:** `TODO.md` L5
+
+## Objective
+The story gate: `tests/e2e/test_background_animation.py` proves both
+background layers are actually running animations (not just declared),
+plus regressions, story file, final validation, and the single atomic
+commit.
+
+## Work
+1. `tests/e2e/test_background_animation.py` (new). The layers are CSS
+   pseudo-elements, so assert via computed style + the Web Animations
+   API (Chromium reports pseudo-element CSS animations through
+   `element.getAnimations()`):
+   1. `test_grid_layer_animation_running` —
+      `getComputedStyle(document.body, "::before").animationName` is the
+      grid-drift keyframe (final name from task 01), timing function
+      `linear`, iteration count `infinite`; and a matching entry in
+      `document.body.getAnimations()` with `playState === "running"`.
+   2. `test_glow_layer_animation_running` — same for `"::after"` with
+      the glow-breathe keyframe; `playState === "running"`.
+   3. `test_animations_advance` — sample `animation.currentTime` (or
+      the `getAnimations()` entry's `currentTime`) for both layers,
+      wait ~500ms (`page.wait_for_timeout`), assert both advanced —
+      the animations are truly running, not paused.
+   4. `test_background_layers_contracts` — both pseudo-elements:
+      `position: fixed`, `z-index: -1`, `pointer-events: none`;
+      `document.documentElement` computed `background-color` is the
+      palette bg (the canvas stays on `html`); `document.body` computed
+      `background-color` is `rgba(0, 0, 0, 0)` (no occlusion).
+   5. `test_no_horizontal_overflow_with_layers` (regression, 360px) —
+      viewport 360px: `document.documentElement.scrollWidth <=
+      clientWidth` (the phase-07 pin, replicated locally).
+2. `.agent/user_stories/background-animation.md` (new) — story file per
+   the repo format: goal, the bug report verbatim from `TODO.md` L5, the
+   owner-confirmed A3 decisions + the found root cause (from task 01's
+   report), E2E mapping table.
+3. Run the suite **in isolation** (prereq `podman compose up -d db`):
+   `uv run pytest tests/e2e/test_background_animation.py -v --no-cov`.
+4. Regressions, in isolation, one command each:
+   - `uv run pytest tests/e2e/test_dark_tech_theme.py -v --no-cov`
+   - `uv run pytest tests/e2e/test_responsive_polish.py -v --no-cov`
+5. Final validation: `uv run pytest` green; `uv run pytest --cov=app
+   --cov-report=term-missing` ≥ today's number (>90% gate);
+   `uv run ruff check . && uv run pyright` clean.
+6. **UI Structure Check** (AGENTS.md rule 5): layers stay behind
+   content, no text/contrast impact, no overflow at 360px.
+7. Finish the phase report (`.agent/reports/22_background_animation/` —
+   E2E results + the task-01 screenshots).
+8. Commit (one atomic commit) and move the phase:
+   ```bash
+   git add -A .agent/ frontend/ tests/
+   git commit --no-gpg-sign -m "fix(ui): animated background actually animates — grid drift and glow breathe per the phase-08 design"
+   mv .agent/phases/todo/22_background_animation .agent/phases/complete/
+   ```
+
+## Testing & Quality
+- Story suite green **in isolation**; both regression suites green in
+  isolation; full unit+integration suite green; `app/` coverage at or
+  above today's number (>90%); ruff + pyright clean.
+
+## Completion Criteria
+- [ ] `test_background_animation.py` 5/5 in isolation.
+- [ ] Regressions (dark tech theme, responsive polish) green in
+      isolation.
+- [ ] Story file + phase report (with screenshots) exist.
+- [ ] One `--no-gpg-sign` commit; phase directory in `complete/`.
diff --git a/.agent/phases/todo/23_containerfile_build/00_phase.md b/.agent/phases/todo/23_containerfile_build/00_phase.md
new file mode 100644
index 0000000..3160eec
--- /dev/null
+++ b/.agent/phases/todo/23_containerfile_build/00_phase.md
@@ -0,0 +1,120 @@
+# Phase 23 — Containerfile: Build the Whole App Image Again
+
+**Source:** `TODO.md` L6 — *"Fix Containerfile build not working"*
+**Story:** `.agent/user_stories/containerfile-build.md` (created by task 02)
+**Context:** `Containerfile` (3 stages: node:22-alpine + esbuild
+0.25.5 frontend bundle → uv/python deps → slim runtime serving
+`/app/static`); `frontend/` (4 pages: `index.html`, `sources.html`,
+`document.html`, `login.html`; assets: `styles.css`, `markdown.js`
+(classic script), `header.js`/`app.js`/`sources.js`/`document.js`/
+`login.js` (ES modules)); `scripts/entrypoint.sh`.
+
+## Verified diagnosis (2026-08-24, this conversion — not a guess)
+1. **Root cause of the build failure:** phase 19 switched the page
+   scripts to `import … from "/assets/header.js"` (an absolute URL).
+   esbuild resolves that as the *filesystem* path `/assets/header.js`
+   and the stage-1 bundle dies:
+   `✘ [ERROR] Could not resolve "/assets/header.js"`
+   (reproduced with esbuild **0.25.5**, the exact pinned version, on a
+   copy of `frontend/`).
+2. **Secondary gap (image would be broken even if it built):** stage 1
+   bundles only `app.js` + `sources.js` and copies only `index.html` +
+   `sources.html`. Missing from the image: `document.html` +
+   `login.html` (phases 10/16), `document.js` + `login.js`, and
+   `markdown.js` (classic script loaded by `index.html` +
+   `document.html`).
+3. **Verified fix:** with relative imports (`from "./header.js"`) all
+   four page scripts bundle cleanly with esbuild 0.25.5.
+4. **Latent double-evaluation trap:** all four HTML pages also load
+   `<script type="module" src="/assets/header.js">` directly while the
+   page script imports it. In dev the browser dedupes (same module
+   URL) — but in the image the bundled page script already contains the
+   header code, so shipping a raw `header.js` too would evaluate the
+   module **twice** (duplicate sign-out listener, double init). The
+   direct tags are redundant: the page script's `import` is hoisted and
+   guarantees `header.js` evaluates before the page script's body calls
+   `initSharedHeader()`, in dev and in the bundle alike.
+
+## Objective
+`podman build -f Containerfile .` succeeds, and the resulting image
+serves the **whole app** — all four pages with their bundled, minified,
+local-only assets (No CDN rule) — with `header.js` evaluated exactly
+once per page.
+
+## Owner-confirmed (2026-08-24, roadmap A4)
+1. **Relative imports** (`./header.js`) over an esbuild alias — simpler,
+   verified working, dev-server behavior unchanged (files are
+   side-by-side).
+2. **Remove the four redundant direct `header.js` script tags** (the
+   design above) rather than ship a raw `header.js` into the image —
+   single module evaluation, no duplicate listeners.
+3. The image must cover **all four pages + all local assets** they
+   reference — the integration test (task 02) enforces this coverage so
+   the gap cannot silently reappear.
+
+## Dependencies
+- `19_shared_header` (complete) — introduced the absolute imports (root
+  cause) and the direct `header.js` tags.
+- `10_story_document_viewer` / `16_admin_auth` (complete) — the pages
+  missing from the image.
+- `08_story_dark_tech_theme` (complete) — No CDN rule the image must
+  honor.
+
+## Tasks
+1. `01_fix_containerfile_build.md` — relative imports, tag removal,
+   stage-1 asset coverage, green `podman build`, image smoke test.
+2. `02_integration_test_commit.md` — `tests/integration/
+   test_containerfile_assets.py` (hermetic coverage pin), regression
+   suites, story file, final validation, the single atomic commit,
+   phase move to `complete/`.
+
+## Locked decisions
+- **A11 honored** — vanilla JS, no CDN, static serving from FastAPI.
+  **A16 honored** — integration test for the new build coverage; story
+  file + report; no Playwright suite required (this phase is
+  build/infrastructure — the phase gate is the hermetic integration
+  test + the real `podman build` + image smoke recorded in the report,
+  plus the dev-server E2E regressions). No anchor changed.
+
+## Testing & Quality
+- **Integration (new `tests/integration/test_containerfile_assets.py`,
+  hermetic — no podman, no network):** every `frontend/*.html` is
+  copied into stage 1's `/out`; every local `src`/`href` asset
+  referenced by the four pages is produced by a stage-1 line (esbuild
+  `--outfile` or `cp`); the four page module scripts are the exact set
+  esbuild bundles; `markdown.js` is produced; no HTML references
+  `/assets/header.js` directly (single-evaluation design pin); the
+  esbuild version stays pinned.
+- **Unit:** none (no `app/` changes).
+- **Coverage:** the >90% `app/` gate is unaffected, re-run to prove it.
+- **Build gate (manual, recorded in the report):** `podman build
+  -f Containerfile .` green; image smoke (task 01 step 6) results +
+  log excerpt in `.agent/reports/23_containerfile_build/`.
+- **Dev regressions (E2E, isolated):** `test_smoke.py`,
+  `test_shared_header.py`, `test_chat_persistence.py` (the HTML tag
+  removal touches dev page load).
+- **Lint/types:** `uv run ruff check . && uv run pyright` clean.
+
+## Completion Criteria
+- [ ] Local esbuild 0.25.5 bundles all four page scripts cleanly.
+- [ ] `podman build -f Containerfile .` green (log excerpt in the
+      report).
+- [ ] Image smoke: container runs (throwaway Postgres 17 + pgvector);
+      `GET /`, `/sources.html`, `/document.html`, `/login.html` → 200;
+      `/assets/app.js` minified and contains the header code;
+      `/assets/markdown.js` 200; no `http(s)://` asset reference in any
+      served page (No CDN rule).
+- [ ] Dev server unchanged in behavior: the three regression E2E suites
+      green in isolation.
+- [ ] `uv run pytest` green; `uv run pytest --cov=app
+      --cov-report=term-missing` ≥ today's number.
+- [ ] `uv run ruff check . && uv run pyright` clean.
+- [ ] `.agent/user_stories/containerfile-build.md` exists.
+- [ ] One `--no-gpg-sign` commit (below);
+      `.agent/phases/todo/23_containerfile_build/` moved to
+      `.agent/phases/complete/`.
+
+## Commit
+```bash
+git add -A .agent/ Containerfile frontend/ tests/ && git commit --no-gpg-sign -m "fix(build): Containerfile builds again — relative module imports, all four pages and shared assets in the image"
+```
diff --git a/.agent/phases/todo/23_containerfile_build/01_fix_containerfile_build.md b/.agent/phases/todo/23_containerfile_build/01_fix_containerfile_build.md
new file mode 100644
index 0000000..6b2dbdd
--- /dev/null
+++ b/.agent/phases/todo/23_containerfile_build/01_fix_containerfile_build.md
@@ -0,0 +1,91 @@
+# Task 01 — Fix the build: relative imports, tag removal, full stage-1 asset coverage
+
+**Phase:** `23_containerfile_build` · **Source:** `TODO.md` L6 —
+*"Fix Containerfile build not working"*
+
+## Objective
+Make `podman build -f Containerfile .` succeed and ship the **complete**
+frontend in the image: all four pages, all four bundled page modules,
+the classic `markdown.js`, and the minified `styles.css` — with
+`header.js` evaluated exactly once per page.
+
+## Work
+1. **Reproduce the failure** and record it in
+   `.agent/reports/23_containerfile_build/` (log excerpt):
+   - fast: `npx -y esbuild@0.25.5` on a copy of `frontend/` → the
+     `Could not resolve "/assets/header.js"` error (root cause, already
+     reproduced during conversion);
+   - authoritative: `podman build -f Containerfile .` → stage 1 fails
+     at the same line.
+2. **`frontend/assets/{app,sources,document,login}.js`** — change the
+   header import from absolute URL to relative (one line each; the
+   specifiers are currently `from "/assets/header.js"`):
+   ```js
+   import { … } from "./header.js";
+   ```
+   (owner-confirmed A4 — relative over esbuild alias; dev-server
+   behavior is unchanged since the files are side-by-side and the
+   module URL resolves to the same file.)
+3. **Remove the four redundant direct `header.js` tags** (owner-confirmed
+   A4-2 — the single-evaluation design from `00_phase.md`):
+   - `frontend/index.html` (~line 119) —
+     `<script type="module" src="/assets/header.js"></script>`;
+   - `frontend/sources.html` (~line 125), `frontend/document.html`
+     (~line 80), `frontend/login.html` (~line 67) — same tag.
+   - Update the surrounding HTML comments that describe the
+     header-before-page-script load order (e.g. index.html ~lines
+     115–119): the order is now guaranteed by the page script's own
+     `import` (hoisted, evaluated before the page script body calls
+     `initSharedHeader()`).
+4. **`Containerfile` stage 1** — cover the whole app (keep the pinned
+   `esbuild@0.25.5` and the existing flags):
+   ```dockerfile
+   RUN mkdir -p /out/assets \
+       && esbuild ./assets/app.js      --bundle --minify --format=esm --target=es2022 --outfile=/out/assets/app.js \
+       && esbuild ./assets/sources.js  --bundle --minify --format=esm --target=es2022 --outfile=/out/assets/sources.js \
+       && esbuild ./assets/document.js --bundle --minify --format=esm --target=es2022 --outfile=/out/assets/document.js \
+       && esbuild ./assets/login.js    --bundle --minify --format=esm --target=es2022 --outfile=/out/assets/login.js \
+       && esbuild ./assets/markdown.js --minify --outfile=/out/assets/markdown.js \
+       && esbuild ./assets/styles.css  --minify --outfile=/out/assets/styles.css \
+       && cp ./index.html ./sources.html ./document.html ./login.html /out/
+   ```
+   (`markdown.js` is a classic script — minify only, **no** `--bundle`;
+   it exposes globals used by the pages.)
+5. **Verify locally (no podman):** with esbuild 0.25.5, all four module
+   bundles + the markdown minify succeed on the real `frontend/` (not a
+   copy — the copy was only for the diagnosis).
+6. **`podman build -f Containerfile .`** → green.
+7. **Image smoke test** (results + log excerpt into the report dir):
+   - throwaway Postgres 17 + pgvector (`podman compose up -d db` and
+     point the container at it, or a one-off container with the same
+     env as `compose.yaml`);
+   - run the built image (migrations run via the entrypoint);
+   - `GET /`, `/sources.html`, `/document.html`, `/login.html` → 200;
+   - `GET /assets/app.js` → 200, minified (single-line-ish), and
+     contains the header code (e.g. the `clearChatStorage` function
+     body); `GET /assets/markdown.js`, `/styles.css`, the other three
+     page modules → 200;
+   - No CDN rule: none of the four served pages contain an `http(s)://`
+     `src`/`href` asset reference.
+   - Teardown the throwaway containers when done.
+8. **Dev-server regression check** (the tag removal touches dev page
+   load — confirm boot order still holds): `uv run uvicorn
+   app.main:app --reload`, load all four pages, check the sign-out
+   binding exists exactly once (DevTools: no duplicate listener — one
+   `POST /api/logout` per click) and `initSharedHeader()` ran. (The
+   isolated E2E regressions run in task 02.)
+
+## Testing & Quality
+- Steps 5–8 above; `uv run ruff check . && uv run pyright` clean
+  (no Python changes, but keep the gate green).
+
+## Completion Criteria
+- [ ] The recorded build failure is fixed at the root cause (relative
+      imports) — not masked by an alias/patch.
+- [ ] All four direct `header.js` tags removed + comments updated; the
+      page scripts' `import "./header.js"` is the only header load.
+- [ ] Stage 1 produces: 4 HTML pages, 4 bundled modules, minified
+      `markdown.js`, minified `styles.css`.
+- [ ] `podman build` green; image smoke all-200 + No CDN + single
+      header evaluation; dev-server boot unchanged (step 8).
+- [ ] Log/screenshot evidence in `.agent/reports/23_containerfile_build/`.
diff --git a/.agent/phases/todo/23_containerfile_build/02_integration_test_commit.md b/.agent/phases/todo/23_containerfile_build/02_integration_test_commit.md
new file mode 100644
index 0000000..812a8a8
--- /dev/null
+++ b/.agent/phases/todo/23_containerfile_build/02_integration_test_commit.md
@@ -0,0 +1,70 @@
+# Task 02 — Integration coverage test, story file, validation, commit
+
+**Phase:** `23_containerfile_build` · **Source:** `TODO.md` L6
+
+## Objective
+Pin the stage-1 asset coverage so it can't silently rot again (a new
+page/script/asset without a matching Containerfile line fails CI), plus
+regressions, story file, final validation, and the single atomic commit.
+
+## Work
+1. `tests/integration/test_containerfile_assets.py` (new — **hermetic**:
+   parses `Containerfile` + `frontend/` as text, no podman, no network).
+   Tests:
+   1. `test_every_html_page_is_copied_into_stage1` — for each
+      `frontend/*.html` in the repo, a stage-1 line copies it into
+      `/out` (regex over the `cp` line; the set must be exactly the
+      four current pages — a new page added to `frontend/` fails this).
+   2. `test_every_local_asset_reference_is_produced` — collect every
+      local `src=`/`href=` under `assets/` or `/assets/` from the four
+      HTML files; each basename must be produced by a stage-1 line
+      (an `esbuild … --outfile=/out/assets/<name>` or a `cp` of it).
+      (This is what catches a missing `markdown.js`-style gap.)
+   3. `test_page_module_scripts_are_bundled` — the set of `type="module"`
+      page scripts referenced by the HTML (basenames) equals the set of
+      scripts esbuild bundles in stage 1 (`app.js`, `sources.js`,
+      `document.js`, `login.js`).
+   4. `test_header_module_is_imported_not_directly_loaded` — no HTML
+      file contains a `<script … src="/assets/header.js">` (or
+      `assets/header.js`) tag (the single-evaluation design pin,
+      owner-confirmed A4-2); and each of the four page scripts imports
+      it relatively (`from "./header.js"`).
+   5. `test_markdown_js_is_a_produced_classic_script` — `markdown.js`
+      has a stage-1 minify line **without** `--bundle` (it is a classic
+      global script) and no `import`/`export` statements at its top
+      level (source pin of that assumption).
+   6. `test_esbuild_stays_pinned` — the frontend stage pins a concrete
+      `esbuild@X.Y.Z` version (no floating version).
+2. `.agent/user_stories/containerfile-build.md` (new) — story file per
+   the repo format: goal, the bug report verbatim from `TODO.md` L6, the
+   verified diagnosis (root cause + missing-asset gap + double-eval
+   trap), the owner-confirmed A4 decisions, and the test mapping table.
+3. Regressions, in isolation, one command each (prereq
+   `podman compose up -d db`):
+   - `uv run pytest tests/e2e/test_smoke.py -v --no-cov`
+   - `uv run pytest tests/e2e/test_shared_header.py -v --no-cov`
+   - `uv run pytest tests/e2e/test_chat_persistence.py -v --no-cov`
+4. Final validation: `uv run pytest` green (includes the new
+   integration test); `uv run pytest --cov=app --cov-report=term-missing`
+   ≥ today's number (>90% gate); `uv run ruff check . && uv run pyright`
+   clean.
+5. Finish the phase report (`.agent/reports/23_containerfile_build/` —
+   build log excerpt, smoke results, regression results).
+6. Commit (one atomic commit) and move the phase:
+   ```bash
+   git add -A .agent/ Containerfile frontend/ tests/
+   git commit --no-gpg-sign -m "fix(build): Containerfile builds again — relative module imports, all four pages and shared assets in the image"
+   mv .agent/phases/todo/23_containerfile_build .agent/phases/complete/
+   ```
+
+## Testing & Quality
+- New integration suite green within `uv run pytest`; the three
+  regression E2E suites green in isolation; full suite green; `app/`
+  coverage at or above today's number (>90%); ruff + pyright clean.
+
+## Completion Criteria
+- [ ] `test_containerfile_assets.py` 6/6 within the full suite.
+- [ ] Regressions (smoke, shared header, chat persistence) green in
+      isolation.
+- [ ] Story file + phase report (build log + smoke evidence) exist.
+- [ ] One `--no-gpg-sign` commit; phase directory in `complete/`.
diff --git a/TODO.md b/TODO.md
new file mode 100644
index 0000000..e7f79b6
--- /dev/null
+++ b/TODO.md
@@ -0,0 +1,5 @@
+# TODO
+
+All items from this file were converted to the phased-execution roadmap
+on 2026-08-24 — see `.agent/phases/todo/` (phases 20–23). Each phase
+file carries a `Source:` line citing the original item and line.