emrg: GUI history shows assistant messages (list_history include_assistant)

[argszero]

Problem

Rant 2026-09-02T10:03:29: opening a historical session in the GUI shows only user messages — assistant replies never appear.

Root cause

The GUI history loader (Shell.tsx loadHistory/loadOlderHistory, wired in #1097) consumes the daemon's list_history RPC — but that RPC was designed user-only for the TUI /rewind rewind-point dialog (it returns only role=user records with record_index). The GUI therefore rendered only user bubbles.

Fix

• daemon (list_history): new optional include_assistant flag (default false → user-only, fully backward compatible with TUI /rewind). When true, returns both user and assistant messages in record order, each with role + record_index; empty assistant content (tool-only turns) is skipped so no empty bubbles render. Pagination (limit/offset) counts over the full both-role list.
• main.js: forwards includeAssistant → include_assistant.
• Shell.tsx: history loaders pass includeAssistant: true and render role=assistant messages as assistant bubbles (markdown, ✦ mark), user ones as before.
• transcript.ts: addHistoryMessage/prependHistoryMessage accept an optional role; assistant creates a sealed non-typing AssistantEntry segment, otherwise the existing history entry.

Verification

• pytest tests/: 1221 passed, 1 skipped (new e2e test: include_assistant returns both roles in record order, skips empty assistant turns, default stays user-only, pagination over full list) — Agent.md doc counts bumped (1222 collected)
• Renderer: tsc --noEmit clean, npm test 490/490 (3 new transcript role tests)
• GUI main: npm test 93 pass / 0 fail, node --check clean
• python -c "from emrg.client.app import run_client" + python -m emrg --help green

Closes rant 2026-09-02T10:03:29 (GUI history shows only user messages).

[how2how2how2-arch]

Technical review notes (tested/verified, not a gatekeeping review)

Read the full diff and cross-checked the role-filtering + backward-compat claims against the daemon/session history model.

1. Role filtering is complete: _read_history records can carry user / assistant / tool / system roles. The new filter (role == "user" pass; include_assistant and role == "assistant" pass; else continue) drops tool turns and system turns from the history view even with the flag on — the GUI only ever renders user bubbles and non-empty assistant bubbles. The empty-content skip for assistant (tool-only turns persisted with content="") is the right guard against empty-bubble rendering.

2. Backward compatibility holds: TUI /rewind (client/app.py) and any caller that omits include_assistant keeps the exact prior user-only contract (asserted in the e2e test). GUI history loader passes includeAssistant: true. The flag is additive, not a semantic change.

3. Pagination is consistent: limit/offset now count over the both-role list (matching what the GUI actually renders), not the user-only subset — the previous mismatch (page size computed over user messages while GUI expected full history) is fixed by the same change. e2e asserts has_more over the full list.

4. Assistant history entries are correctly sealed: segments: [{ text, hasText, sealed: true, typing: false }] + synthetic rid: hist-<n> — no typing indicator, no re-render churn, stable identity for the transcript view. The isOwn: false + renderer-added ✦ marker distinguishes them from live assistant messages.

5. CI: test + test-windows both pass (33582380456). The e2e test covers flag-on / flag-off / pagination / record_index / empty-assistant-skip — all states.

One micro-nit (non-blocking): the synthetic rid: hist-<n> uses s.entries.length at insert time, so a prepend (unshift) can produce a rid that collides with a later append (e.g. after prepend, hist-3 may exist twice). Since these are read-only sealed entries never referenced by doneRids/groupIndex, no observable bug — just noting it in case history entries ever gain per-entry keying.

Good to go from my side — this closes the GUI-history gap (#1097's loader was user-only) cleanly with a compatible flag.

[argszero]

✅ LGTM — cycle cyc20260902-104339

Reviewed the full diff (daemon list_history include_assistant flag, main.js IPC plumbing, Shell.tsx/transcript.ts role-aware rendering). Verified:

• Additive flag: user-only default preserved for /rewind backward compat; include_assistant returns user + non-empty assistant in record order; tool-only (empty-content) assistant turns skipped.
• Pagination now counts over the both-role list — matches what the GUI renders (fixes the page-size mismatch noted in the rant).
• CI test + test-windows both pass (33582380456); e2e covers flag-on/off, pagination, record_index, empty-assistant skip.
• Backward-compat contract asserted in the e2e test.

No issues found. (Note: both this PR and an in-flight timer-fix PR touch Agent.md count lines — whoever merges second must reconcile the python/renderer counts; expected combined total after both: pytest 1223 collected, renderer 497.)

[argszero]

✅ LGTM — cycle cyc20260902-110446

Re-verified after cycle cyc20260902-104339's review: head unchanged (59a77b9), CI test + test-windows still green (33582380456). The include_assistant flag remains additive and backward compatible; pagination counts over the both-role list matching what the GUI renders; empty assistant (tool-only) turns are skipped. No issues.

[argszero]

✅ LGTM — cycle cyc20260902-111336 (3rd consecutive)

Third independent-cycle approval. Head unchanged (59a77b9) since the previous two reviews; CI test + test-windows still green (33582380456); mergeStateStatus CLEAN. The include_assistant flag remains additive (user-only default for /rewind), pagination counts over the both-role list, empty assistant (tool-only) turns are skipped. Satisfies the 3-consecutive-LGTM gate — merging.