emrg: session — recompute message_count on compact, count message records only (rant 2026-08-31T14:18:14)

[argszero]

Fix TUI/GUI message count inflation after /compact (rant 2026-08-31T14:18:14).

Problem

The status bar showed "5770 msgs" for a session whose history.jsonl held only 192 records. Root cause: session.py's counter only ever grows:

1. append_message() incremented _message_count for every record — including tool_result records persisted from the tool loop (daemon.py:2778) and summaries.
2. Session.compact() replaced the compacted messages with one summary but never decremented _message_count — the inflated value survived (e.g. 5806 = ~5700 pre-compact + ~100 after, matching last_compact_at).
3. TUI status bar (client/app.py:1018) and GUI session list (widgets.py:333) display meta.message_count directly.

Meanwhile rewind (daemon.py:2083) already recomputed the count as sum(r.type == "message") — compact simply missed the same reconciliation.

Fix (Plan A — confirmed by host)

• Session.compact(): after writing the new history, recompute _message_count from the surviving records (message-type only) — aligned with rewind semantics.
• Session.append_message(): increment only for type == "message" records (missing type defaults to message, preserving existing callers) — tool_result/summary records persist but no longer inflate the user-facing count, so the count cannot re-inflate between compacts.

Tests

• test_append_tool_result_does_not_increment_count — tool_results persist but are excluded from the count.
• test_compact_recomputes_message_count — mixed message/tool_result history: count shrinks to the surviving message records after compact (regression for the rant).
• test_compact_count_persists_after_reload — recomputed count survives Session.load() from meta.json.

Verification: pytest 1182 passed + 1 skipped; import + CLI green; Agent.md Python count synced to 1183.

Note: a display-side fallback (Plan B — fall back to the actual record count when message_count deviates wildly) was mentioned as optional in the rant; Plan A alone fixes the reported bug, so B is deferred.

[argszero]

✅ LGTM — cycle (1/3)

Reviewed head ba65cee against master (1fdc606):

• Root cause confirmed: append_message() incremented _message_count for every persisted record including tool_result, and Session.compact() never decremented it — the TUI/GUI 'msgs' counter inflated monotonically (e.g. 5770 shown vs 192 records in history.jsonl).
• Fix is sound: compact() recomputes _message_count from the surviving message-type records (same semantics as the rewind handler in daemon.py), and append_message() only counts type == "message" records (defaulting to message when absent, so existing callers are preserved); tool_result/summary records still persist to history.
• Tests: 3 new (tool_result exclusion; compact recompute regression with mixed message/tool_result history; recompute survives reload from meta.json). test_session.py 66 passed; CI test + test-windows green (run 33366280252); mergeable_state clean.
  No issues found.

[argszero]

✅ LGTM — cycle (2/3)

Re-verified head ba65cee against current master (4fd2b63) this cycle:

• diff unchanged and matches the intended fix: append_message() increments _message_count only for type == "message" records (defaulting to message when absent — existing callers preserved; tool_result/summary still persist without inflating the count); Session.compact() recomputes _message_count from the surviving message-type records, aligned with the rewind handler's semantics in daemon.py.
• The fix directly addresses the reported symptom (5770 msgs shown vs 192 records in history.jsonl) and prevents re-inflation between compacts.
• Tests: 3 new regression tests (tool_result exclusion; compact recompute; reload persistence) — branch test_session.py 66 passed; master baseline still green (63 passed). CI test + test-windows green (run 33366280252); mergeable_state clean.
  No issues found.

[argszero]

✅ LGTM — cycle (3/3)

Re-verified head ba65cee against current master (4fd2b63) this cycle:

• diff unchanged: append_message() counts only type == "message" records (tool_result/summary persist without inflating the count); Session.compact() recomputes _message_count from surviving message-type records (rewind-aligned).
• Local test_session.py + test_doc_counts.py all pass (68 passed); CI test + test-windows green (run 33366280252); mergeable_state clean.
  3 consecutive ✅ from distinct cycles (07:04:12Z / 07:11:06Z / this cycle), no ❌ — good to merge.