emrg: implement tool result sliding window

[argszero]

Tool Result Sliding Window (rant 2026-08-22T11:33:54)

Problem: every LLM request re-sends the full session history including all complete tool results (bash/read/grep outputs). Long sessions (dozens of tool rounds, up to max_tool_rounds=270) grow token usage linearly and hit auto-compact (lossy summarization) early. Early tool results are rarely useful but dominate the token budget.

Solution: fold tool results older than the most recent N rounds (default 7, config tool_window_rounds, 0 = off = old behavior) into a placeholder assistant message before each LLM request.

What changed

• emrg/config.py — new LlmConfig.tool_window_rounds: int = 7 (reads [llm] tool_window_rounds from config.toml; 0 disables).
• emrg/server/daemon.py:
  • _apply_tool_window(messages, keep_rounds, history_path) — pure function (no session/disk access). Atomic group = assistant-with-tool_calls + its following tool messages (OpenAI pairing constraint). Keeps the most recent N groups in full; each older group is replaced in-place by a placeholder assistant message. Never folds system/user/plain-text-assistant/summary messages.
  • _fold_tool_group() — builds the placeholder: [Tool results omitted — older than recent N rounds] + executed tool counts + tool_call_ids + absolute history.jsonl path + grep backtrack hint.
  • Called at the top of the tool-loop while True: (before round-budget checks and the auto-compact token estimate), so both the session history and same-request accumulated rounds are covered. On-disk history.jsonl stays complete — nothing is lost, only the sent payload is folded.
  • _build_system_prompt injects tool_window_rounds into the render context.
• emrg/server/prompts/system.j2 — fold notice + "summarize valuable info proactively" advice block, rendered only when tool_window_rounds > 0 (zero-config behavior identical to before). The gitignored packaged copy dist/runtime/source/emrg/server/prompts/system.j2 was synced on disk as well (build artifact, regenerated by packaging).
• tests/test_daemon.py — +5 tests: 10-group fold (3 oldest → placeholders, 7 complete, order preserved), keep_rounds=0 identity, within-window no-op, multi-tool count aggregation, non-foldable preservation + OpenAI-pairing validation via _validate_tool_messages.
• Agent.md — test count 1003 → 1008 (doc-count guard).

Verification

• Full suite: 1007 passed, 1 skipped (was 1002; +5 new).
• Import check + emrg --help OK.
• Placeholder backtrack path = <cwd>/.emrg/sessions/<session_id>/history.jsonl (verified against Session._history_path).

[argszero]

✅ LGTM — cycle. Re-verified head ecad963 against design doc (tool-window-design.md):

• _apply_tool_window is a pure fold (no session/disk access); atomic groups = assistant-with-tool_calls + following tool messages, so OpenAI pairing (session._validate_tool_messages) always holds — verified by the new validation test.
• In-place placement preserves message order; only the most recent keep_rounds groups stay complete; keep_rounds<=0 = identity (old behavior).
• Called at the top of the tool-loop while True before the round-budget check and the auto-compact token estimate — both history and same-request accumulation are covered; disk history.jsonl stays complete.
• Placeholder carries executed counts + tool_call_ids + absolute history path + grep hint for backtracking; role=assistant without tool_calls → API-valid.
• system.j2 renders the notice only when tool_window_rounds > 0 and is only rendered via _build_system_prompt (which always sets the ctx key) — no undefined-variable path.
• Full suite: 1007 passed, 1 skipped; CI test + test-windows green (run 32561317425).

[argszero]

✅ LGTM — cycle (2nd). Re-verified merged head 0c2f5de (merge of master d56cc2d):

• Conflict with #935 (Agent.md doc counts) resolved: pytest count kept at 1008 (my +5 tests, verified: 1007 passed + 1 skipped collected), GUI counts updated to master's 20 conn-manager / 129 renderer smoke from #935.
• Local verification on merged head: 1007 passed, 1 skipped; import + CLI OK.
• CI re-run on merged head: test + test-windows both pass (run 32562164734); mergeStateStatus now MERGEABLE.
• Functional review unchanged from cycle 1259: _apply_tool_window pure fold, atomic group = assistant-with-tool_calls + following tool messages, keep_rounds<=0 identity, placeholder carries counts + tool_call_ids + history path, system.j2 gated on tool_window_rounds>0 with key always set.

[argszero]

✅ LGTM — cycle (3rd). Fresh-eyes re-verification of head 0c2f5de:

• Diff vs master: 5 files +246/−2 — Agent.md (test count 1008), config.py (tool_window_rounds=7, parsed from llm config, 0=off), daemon.py (+106: constructor wiring, ctx in _build_system_prompt always-set, _apply_tool_window/_fold_tool_group pure functions, call as first statement of tool-loop while True), system.j2 (fold notice gated on tool_window_rounds>0), tests (+5: 10-group fold, zero identity, within-window no-op, count aggregation, pairing validation).
• Folding preserves OpenAI message pairing: atomic group = assistant-with-tool_calls + following tool messages; placeholder is role=assistant without tool_calls + carries executed counts + tool_call_ids + absolute history.jsonl path + grep hint for backtracking; disk history stays complete.
• keep_rounds<=0 = identity (old behavior); in-place order preserved.
• Local: 1007 passed + 1 skipped (=1008 collected); import + CLI OK. CI: test + test-windows green (run 32562164734); mergeStateStatus CLEAN.