perf(tasks): maintain task browse summaries on trial writes

[kyle-compute]

Why this PR exists

• Loading the main /tasks page currently does database work proportional to the organization's trial history even though the screen shows only 24 task cards.
• Fixing that requires an indexed per-version summary that the page query can order and limit on before touching trial rows.
• This PR adds that summary's model, helpers, and write-through refresh hooks. It deliberately changes no read path.
• The combined re-land (perf(tasks): bound task browser page work #1152) was 936 lines of application code; the limit is 500. This is split 1 of 4.

Split structure and merge order

1. perf(tasks): maintain task browse summaries on trial writes #1185 (this PR) — summary maintenance: model, helpers, write-through refresh hooks. 432 application-code lines.
2. perf(tasks): repopulate historical task browse summary rows #1190 — historical population: migration task_browse_summary_002 rebuilds every summary row.
3. perf(tasks): summary-first /tasks/browse selection with bounded previews #1191 — backend read cutover: GET /tasks/browse becomes summary-first with bounded previews. 452 application-code lines.
4. perf(frontend): render bounded task-card trial preview #1192 — frontend card cap: render the bounded preview and "showing 24 of N". 48 application-code lines.

• Merge strictly in that order. Each later PR is based on the previous PR's branch; after a parent merges, rebase the child onto staging and retarget it before merging.
• Each piece gets its own preview deployment, staging verification, and rollback.

Exact screen, request, and frequency

• Screen: the authenticated main task-card grid at /tasks.
• Screenshot gate: attach a representative slow /tasks skeleton and loaded grid to this body before merge.
• Controlling browser request: GET /api/tasks/browse. This PR does not change its behavior; it prepares the data that perf(tasks): summary-first /tasks/browse selection with bounded previews #1191 will read.
• Write frequency: every trial create/import, start/reset, completion, cancellation, retry/supersede, scoped deletion, experiment deletion, and default-version selection refreshes the affected version's summary row inside the same transaction.
• S3 involvement: none.

Production evidence

• Production p75 for /api/tasks/browse was approximately 8.86 seconds; production /tasks LCP was approximately 9.15 seconds.
• The main task-selection SQL alone took approximately 3.09 seconds because it groups the organization's entire trial history before applying the 24-task limit.
• The summary rows this PR maintains are what lets perf(tasks): summary-first /tasks/browse selection with bounded previews #1191 order and limit first instead.

What this PR changes

• Adds TaskBrowseSummaryModel over the existing task_version_browse_summaries table: one row per task version with exact card counters, status buckets, model-grouped cost inputs, and last_run_at. The table already exists in staging and production (migration task_browse_summary_001 survived the Revert "perf(tasks): bound task browser page work" #1156 revert), so this PR contains no DDL.
• Adds refresh_task_browse_summaries and calls it in the same transaction at every trial-population mutation: create/import, start/reset, completion, cancellation, retry/supersede, scoped deletion, experiment deletion, and default-version selection.
• Serializes refreshes with sorted transaction-scoped advisory locks; concurrent trial inserts hold KEY SHARE FK locks on task_versions, and a FOR UPDATE upgrade there can deadlock (documented in AGENTS.md).
• Ships the shared cost helper with the corrected billed_cost_usd accumulator key. The first landing wrote billed_usd, which is the KeyError that 500'd /tasks for any org with a billed trial and caused the Revert "perf(tasks): bound task browser page work" #1156 revert.
• No reader change: GET /tasks/browse still aggregates on demand.
• No historical scan: no migration, no backfill. Rows exist only for versions written to after deploy.
• Safe to deploy unused: nothing reads the rows yet.

Why this is being re-landed

• perf(tasks): bound task browser page work #1152 merged and was reverted shortly afterward by Revert "perf(tasks): bound task browser page work" #1156. The production failure was the billed_usd/billed_cost_usd key mismatch above.
• CI had no billed-trial fixture, so it stayed green while production /tasks would 500. The corrected key lands here; the billed-trial coverage that exercises it end-to-end lands with the reader in perf(tasks): summary-first /tasks/browse selection with bounded previews #1191.
• The summary table and its original backfill survived the revert, which is why this PR needs no DDL and perf(tasks): repopulate historical task browse summary rows #1190 must rebuild the stale rows.

What this PR does not change

• It does not change GET /tasks/browse or any response shape. Users see zero behavior change.
• It does not populate historical summary rows. That is perf(tasks): repopulate historical task browse summary rows #1190.
• It does not optimize clicking a task title. That is perf(tasks): add bounded task-open response primitives #1167/perf(frontend): add bounded task-open resource #1168.
• It does not read, list, move, or backfill S3 artifacts.

Historical-data decision

• This PR takes no historical-data action. Population is isolated in perf(tasks): repopulate historical task browse summary rows #1190 so its runtime risk is measured on the production-sized staging clone instead of riding along with application code.

Validation completed

• backend tests/test_browse_search.py: 5 passed. The single failure, test_search_grammar_and_or_exclude, fails identically on clean staging.
• New test: experiment delete refreshes the surviving task's summary row in the delete transaction (the Bugbot finding on the combined PR).
• New test: two sessions insert trials for the same version (both holding KEY SHARE FK locks), then refresh; the holder must not block and the second refresh must serialize behind the first's commit. A lock-ordering bug hangs this test instead of flaking it.
• oddish tests/test_retry_trial_core.py and tests/test_local_runner_gate.py: 20 passed.

Rollback

• Revert the hooks. Users never saw a behavior change, because nothing reads the summary rows.
• The table stays; it already exists in production and is harmless unread.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
oddish-app	Ready	Preview	Aug 12, 2026 12:37am

[cursor]

Risk: high. Left a non-blocking comment — Cursor Bugbot finished without a clean pass and left an unresolved high-severity finding, so this was not approved. Reviewers were assigned for human review.

  

_{Sent by Cursor Approval Agent: Pull Request Router and Approver}

[github-actions]

Oddish preview

Commit: bf1223bf8609264d9277503ebfb725faa2eff31f

Surface	Link	Target
Frontend	https://pr-1185.oddish.app	Vercel preview for bf1223b
Backend	oddish-pr-1185	oddish-pr-1185
Database	project lxfqpxpcttlaqlhfwsoh	project lxfqpxpcttlaqlhfwsoh

Vercel deployment URL: https://oddish-7goxerv0n.oddish.app

Plan:

• Frontend deploy: true
• Backend deploy: true
• Migrations: false

This comment is updated by the PR Preview workflow.

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit bf1223b. Configure here.}