emrg: vibe check & slowdown refactor — recommend_slowdown-only throttle + unified work/slowdown_reason fields (rant 2026-08-20T10:58:55)

[argszero]

vibe check & slowdown refactor (rant 2026-08-20T10:58:55, host design-finalized)

The host finalized a redesign of the vibe-check / slowdown mechanism:

• Delete the empty-cycle counter (_empty_cycles) and the slowdown-vote machinery (_slowdown_hits, thresholds 30/10) entirely.
• Slowdown is now decided solely by the vibe check's recommend_slowdown flag: true → _slowdown_active=True (next run at heartbeat interval), false → back to normal interval. Vibe unavailable / truncated / aborted → state unchanged (conservative). Manual trigger (/trigger) clears the flag.
• The throttle flag is persisted to ~/.emrg/saturation/<task>.json ({"slowdown_active": bool}) so it survives daemon restarts; legacy files read as not-throttled (no migration).
• Fields unified to work / recommend_slowdown / slowdown_reason: summary/done → work, meaningful deleted (agent no longer self-rates value), reason → slowdown_reason.
• status() drops last_cycle_summary; GUI primary row shows only last_run_at + throttled badge (heartbeat interval), the expandable secondary list shows work / throttle flag / slowdown_reason (no idle badge — meaningful is gone).

Background: the agent's self-rating meaningful was indirect and error-prone (e.g. a deepseek-harness session with 44 tool calls was judged "empty" by a naive heuristic).

Changes

• emrg/server/scheduler.py — state machine rewritten (init, _load/_save_saturation_state, _load_task_runs/_append_task_run fields, run() wait mode + manual-trigger reset, status(), _saturation_heartbeat_active(), _request_vibe_check returns work/recommend_slowdown/slowdown_reason, _run_evolution_cycle log construction).
• emrg/server/prompts/vibe_check.j2 — prompt rewritten for the 3-field JSON output.
• emrg/protocol.py — EvolutionLog fields: summary→work, meaningful deleted, reason→slowdown_reason.
• emrg/server/daemon.py — _task_vibe_check result parsing updated.
• emrg/gui/renderer/js/dialogs.js + i18n.js — task run detail uses work/slowdown_reason; saturation badge text → app.taskThrottled ("已降频 · heartbeat {m}s").
• tests/ — test_scheduler.py (state machine, persistence, heartbeat tests rewritten for the flag model), test_ws_e2e.py (3-field vibe result), renderer.smoke.test.js (new mock fields + badge assertions).

Verification

• uv run pytest tests/ -q → 986 passed, 1 skipped
• uv run python -c "from emrg.client.app import run_client" + python -m emrg --help → OK
• GUI npm test → 253/260 (7 by-design local daemon-spawn fails, green in CI)

[argszero]

✅ LGTM — cycle (1/3)

Full implementation of host rant 2026-08-20T10:58:55 (design-finalized vibe check & slowdown refactor):

• empty-cycle counter + slowdown-vote machinery deleted; throttle decided solely by vibe check recommend_slowdown
• fields unified to work / recommend_slowdown / slowdown_reason (summary/meaningful/reason removed) across protocol, daemon, scheduler, GUI
• slowdown flag persisted to ~/.emrg/saturation/.json, survives daemon restarts (legacy files read as not-throttled)
• status() drops last_cycle_summary; GUI primary row shows throttled badge (heartbeat interval), secondary list shows work/slowdown_reason

Verification:

• pytest: 986 passed, 1 skipped (test + test-windows CI green on head via re-triggered dispatch run 32329212493)
• import + CLI OK; GUI npm test 253/260 (7 by-design local daemon-spawn fails, green in CI)

[argszero]

✅ LGTM — cycle (2/3)

Fresh-cycle re-review of the full diff (head 83925e0):

• protocol.py: EvolutionLog fields unified — summary→work, meaningful deleted, reason→slowdown_reason. Clean, no stale refs.
• vibe_check.j2: prompt rewritten for the 3-field JSON; agent no longer self-rates meaningful; slowdown_reason required only when recommend_slowdown=true. Matches daemon parsing (_task_vibe_check).
• scheduler.py: empty-cycle counter + vote machinery fully removed; _slowdown_active is the single switch (vibe recommend_slowdown only); persisted to saturation/.json across restarts; truncated/vibe-unavailable/aborted leave state unchanged; manual trigger clears. status() drops last_cycle_summary, saturation = {heartbeat_interval, heartbeat_active}.
• GUI: dialogs.js renders work/slowdown_reason + throttled badge on heartbeat_active; i18n taskThrottled; renderer smoke updated for the new model.
• Tests rewritten comprehensively (state machine, persistence incl. legacy-file no-migration, heartbeat, ws_e2e 3-field).

CI: test + test-windows SUCCESS on head via re-triggered dispatch run 32329212493 (original PR run cancelled after Windows runner hang — runner issue, not code).

[argszero]

✅ LGTM — cycle (3/3)

Fresh-cycle re-verification: head 83925e0 unchanged since 2/3. Both platforms green on head (test + test-windows SUCCESS via dispatch run 32329212493; the PR's own windows run was cancelled only due to a GitHub runner hang — re-triggered and green, same as #873). Full diff reviewed across the earlier cycles: protocol/daemon/scheduler/GUI field unification, flag-only slowdown state machine with persistence, comprehensive test rewrite. No issues found — merging.