alee/6krrt
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat(admin): Lift A.2, surface loops and stalling models with one-click block #104

Merged
alee merged 17 commits from feat/no-progress-lift-a2 into main 2026-09-28 23:56:34 +00:00
Conversation 0 Commits 17 Files Changed 9 +1419 -28
alee commented 2026-09-28 23:52:16 +00:00
Owner
Copy Link

Lift A.2: surfaces watchdog loops and stalling models in the admin portal, with the lever beside the evidence (North Star #4). Detection is unchanged; this is the surfacing half Lift A left thin.

What ships

  • Loops panel (Home, #loops): a table of open alerts with severity, session title, model, top repeated target, calls and $ since the last landed change, time open, and Block model. Alerts resolved in the last 24h are listed below, dimmed. GET /admin/api/watchdog/loops now joins each open alert to its root's latest flagged verdict and returns {alerts, resolved}.
  • Stalls by model (Models): 1h / 24h / 7d radio. Per model: distinct stalled sessions, $ in stalled sessions (MAX per session, then SUM, because the column is cumulative), traffic share as its own column, availability, and Block. New GET /admin/api/watchdog/rollup?hours=N.
  • Block with a reason, inline: an inline confirm row with a prefilled, editable reason (stalled N sessions in 24h, looping session <title>), no browser dialogs. The dropdown path now records set from Models dropdown.
  • Blocked models card: reason, blocked-since and one-click Unblock. New GET /admin/api/models/blocked.
  • Alert text carries the evidence: <title>: <calls> calls, $<cost> since last landed change on <model>. <url>#loops.

Verification

  • Full suite: 2441 passed, 1 failed (test_incumbent_routing::test_debug_log_emits_incumbent_identity, also failing on main).
  • Pinned ruff@0.16.9: no new findings in touched Python files.
  • Visual QA at 1400 and 800 px on an 8081 sandbox. The Block, Confirm, Blocked list, Unblock round trip was clicked through by hand on a sqlite-backup copy of the live DB. The live DB was never written.
  • Every time window uses julianday() on both sides; stored timestamps are ISO with T/+00:00.

Known follow-ups (not in this PR)

  • Argument-less tools such as Playwright browser_snapshot count as exact repeats. That caused one real false alert (a QA worker), which triggered and resolved correctly.
  • calls_since_landed resets on any landing in the whole session tree, not just the session's own descendants.

After merge: systemctl --user restart llm-router.service so 8080 serves the new endpoints and pages. The watchdog picks up the new alert text at its next tick.

🤖 Generated with Claude Code

https://claude.ai/code/session_01N9biTbFC63yDfYfUsZmhgd

Lift A.2: surfaces watchdog loops and stalling models in the admin portal, with the lever beside the evidence (North Star #4). Detection is unchanged; this is the surfacing half Lift A left thin. **What ships** - **Loops panel** (Home, `#loops`): a table of open alerts with severity, session title, model, top repeated target, calls and $ since the last landed change, time open, and **Block model**. Alerts resolved in the last 24h are listed below, dimmed. `GET /admin/api/watchdog/loops` now joins each open alert to its root's latest flagged verdict and returns `{alerts, resolved}`. - **Stalls by model** (Models): 1h / 24h / 7d radio. Per model: distinct stalled sessions, $ in stalled sessions (MAX per session, then SUM, because the column is cumulative), traffic share as its own column, availability, and **Block**. New `GET /admin/api/watchdog/rollup?hours=N`. - **Block with a reason, inline**: an inline confirm row with a prefilled, editable reason (`stalled N sessions in 24h`, `looping session <title>`), no browser dialogs. The dropdown path now records `set from Models dropdown`. - **Blocked models** card: reason, blocked-since and one-click Unblock. New `GET /admin/api/models/blocked`. - **Alert text carries the evidence**: `<title>: <calls> calls, $<cost> since last landed change on <model>. <url>#loops`. **Verification** - Full suite: 2441 passed, 1 failed (`test_incumbent_routing::test_debug_log_emits_incumbent_identity`, also failing on `main`). - Pinned `ruff@0.16.9`: no new findings in touched Python files. - Visual QA at 1400 and 800 px on an 8081 sandbox. The Block, Confirm, Blocked list, Unblock round trip was clicked through by hand on a sqlite-backup copy of the live DB. The live DB was never written. - Every time window uses `julianday()` on both sides; stored timestamps are ISO with `T`/`+00:00`. **Known follow-ups (not in this PR)** - Argument-less tools such as Playwright `browser_snapshot` count as exact repeats. That caused one real false alert (a QA worker), which triggered and resolved correctly. - `calls_since_landed` resets on any landing in the whole session tree, not just the session's own descendants. **After merge:** `systemctl --user restart llm-router.service` so 8080 serves the new endpoints and pages. The watchdog picks up the new alert text at its next tick. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01N9biTbFC63yDfYfUsZmhgd
alee added 17 commits 2026-09-28 23:52:17 +00:00
plans: Lift A.2 brief, surfacing loops and stalling models acde5fd396 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01N9biTbFC63yDfYfUsZmhgd
feat(admin): enrich watchdog loops endpoint with verdict evidence and resolved list 235e2e8b19
feat(admin): enrich loops panel UI with evidence columns and formatting helpers 3658adcb57
feat(admin): add watchdog rollup endpoint for per-model stall aggregates d015958ddb
feat(admin): add stalls-by-model card with window picker to models page d05ca52479
feat(admin): add blocked models list endpoint 05024ccecb
feat(admin): add blocked models card with unblock button to models page b0ff5943c2
feat(watchdog): enrich alert summary with calls, cost, and model evidence 2270ed3770
feat(admin): add inline block confirm with reason, shared across panels 5d5c4fca2c
fix(admin): import-sort in test_admin_watchdog.py (I001) fc2b8fa250
fix: import sort order in test_admin_watchdog.py c60402b5ac
fix: use nth-of-type(1) selector for inline confirm buttons; fix cost NULL→None propagation in watchdog 67106852cd
fix(admin): Loops panel as real table with header+top_what+num, fix stalls num CSS fc1b81f55d
fix(admin): use data-* attrs for Block button, fix title-with-quotes break 6be41ff19d
fix(admin): wire stalls table Block onclick, fix toast id missing quote 319efc1d66 
- Add onclick='blockModelConfirm()' to stalls table Block buttons;
  commit 6be41ff added data-* attrs but never wired the handler
- Fix toast div: id="toast class=" → id="toast" class=" (was missing
  closing quote, causing toast() JS to crash)
fix(admin): wire Block button onclick to blockModelConfirm(event) de67be5e65
fix(watchdog): cost_since_landed_map values may be None 1938a461bf 
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01N9biTbFC63yDfYfUsZmhgd
alee merged commit 8d532069d3 into main 2026-09-28 23:56:34 +00:00
Sign in to join this conversation.
Reviewers
No Reviewers
Labels
No items
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
alee
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: alee/6krrt#104
Delete Branch "feat/no-progress-lift-a2"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?