Pickup — Build the Enhanced Records List + Promotion Checkpoint

Read this first

The spec is finished. Open and follow: [[Enhanced-Records-List-and-Promotion-Checkpoint]] — context-v/specs/Enhanced-Records-List-and-Promotion-Checkpoint.md

It is self-contained — data model, the microfrontend surface, the promotion-as-snapshot semantics, the cemented triage_states on rows, the row-level archived flag (the only way a record drops out of the lineage), the iterative loop, and 8 acceptance checkpoints.

Also relevant: [[Helpful-Links-on-Records-Captured-During-Triage]] — context-v/prompts/Helpful-Links-on-Records-Captured-During-Triage.md (the helpful-links shape, already implemented end-to-end in this session and feeding into the spec above).

Where things stand at the close of 2026-05-22

The session shipped a lot of working code, mostly uncommitted on branch rebuild/turbo-rsbuild. Only one commit landed:

• 82c7ca4 fix(record-collector, prompt-template-manager): clear a11y warnings on sidebar list items

Everything below is on disk and live in the running stack but not yet committed. The walking skeleton’s pre-flight + post-flight loop is now genuinely usable.

Prompt Templates (apps/prompt-template-manager)

• Per-row × delete in the sidebar (no need to load each prompt into the editor to delete it).
• The Create/Save button is dirty-aware — disabled when nothing has changed.
• New Apply → button: explicit handoff to Request Reviewer (was silently auto-dispatched on selection before).

Request Reviewer (apps/request-reviewer)

• A firing… indicator the instant firing flips true, with a spinner, regardless of whether progress events have landed yet.
• Cancel run button while a run is in flight — wired to a new prompt.run.cancel capability; the runner aborts the in-flight Anthropic request via AbortSignal.
• Coverage strip between the row stepper and the model knobs: 12 / 207 covered · 3 needs-rerun · 192 remaining. Auto-refreshes on response events. New Fire remaining (N) button passes row_ids = uncovered so the next pass doesn’t re-fire what’s already done.
• include needs-rerun checkbox to fold the rerun-queued set into the remaining-batch.

Response Reviewer (apps/response-reviewer)

• New needs-human triage flag (the “model said unknown but the entity is itself unclear, a human needs to clarify what this even is” bucket).
• Per-bucket counts on the filter chips at the top (all 96 · unflagged 0 · good 62 · needs-human 34 · partial 0 · wrong 0).
• The bulk 🧹 Clear button moved up next to the filter chips (top-right) — filter-scoped clear with the count visible.
• The triage flag chips moved INLINE with the stepper row (◀ response N / N ▶ flag-pill | triage: good partial wrong needs-rerun needs-human) — one mouse movement from step to flag.
• Actions row converted to icon-only with CSS tooltips (instant on hover): ✓ accept · ↻ rerun · ✦ distill (stubbed, disabled) · 🗑 delete this response.
• A bottom border separates the filter zone (top) from the current-response zone (just below).
• Critical bug fixed: response text edits now autosave to a new edited_text field on the response, on textarea blur AND before stepping to the next response. A visible save-state pill on the response heading reads saving… / unsaved (red) / saved Ns ago. beforeunload does a last-ditch flush. 75 URLs were lost before this fix — the user re-researched them.
• Helpful links on every record: capture URLs + optional notes during triage; lives in Row.fields.helpful_links (NOT in the CSV schema). New capabilities row.helpful_links.add / row.helpful_links.remove / row.get in the workspace. Survives across record-set derivations because it lives on the row. Provenance: each link stores the response_id being triaged when it was captured.

Prompt Runner (services/prompt-runner)

• Per-row diagnostic logging: row started / row completed (with elapsed_ms, chars) / row failed (with elapsed_ms, error).
• Per-row 90-second timeout via AbortController → SDK closes the HTTP request → row fails fast with [error: row timed out after 90000ms] → loop continues.
• Run-level cancellation: keyed by record_set_id, Map<record_set_id, AbortController> in server.ts, fired on prompt.run.cancel.requested.
• The runner now propagates an AbortSignal into runPrompt and through to anthropic.messages.create({ signal }).

CSV ingest (services/ingest)

• Now drops fully-blank rows on import (matches xlsx-ingest’s existing behaviour). Logs the dropped count.
• The current 261 partial-blank rows (only Probability (auto)-style auto-formula cells populated, no real prospect data) were surgically removed from row-store + response-store in this session — see .backups/ for the pre-surgery snapshots.

Response store

• edited_text + edited_at added to ResponseRecord; backfilled to null on load for older records.
• response.set_text capability + response.edited broadcast.
• response.delete + response.delete_all capabilities + response.deleted broadcast.
• response.coverage capability — returns { covered_row_ids, needs_rerun_row_ids } for a given prompt × record set.

Row store

• row.get capability (was missing).
• row.helpful_links.add / row.helpful_links.remove capabilities + handlers; both broadcast row.updated.

Data state

• 4 record sets in row-store: parent (96 rows, blank-org removed), RSVP list (39, untouched), and two derived + url sets (25 + 71 = 96).
• 96 responses in response-store, all flagged (62 good + 34 needs-human). Zero unflagged.
• 21 row cells in the parent set have accepted URLs written; the remaining 75 are where edits-were-lost-before-autosave-landed.

Pre-surgery backups still on disk

augment-it/.backups/responses_2026-05-22_180004.json + rows_2026-05-22_180004.json (the pre-cleanup snapshot) and responses_cleaned.json + rows_cleaned.json (what got copied back in). Keep until the next commit is sanity-checked.

Concrete first step for the next session

Per the spec, the smallest forward move is the record_uuid plumbing:

1. Define the field in Row.fields (side-channel; not a schema column).
2. At ingest (services/ingest/src/parse.ts and services/xlsx-ingest/src/parse.ts): mint a rec_<base36> for every new row.
3. At derivation (services/prompt-runner/src/run.ts): propagate record_uuid from parent row to derived row alongside the existing field-spread.
4. One-shot backfill script in services/row-store/scripts/backfill-record-uuid.ts:
   • Parent rows in each non-derived set get fresh record_uuids.
   • Derived rows match to their parent by within-set order (the derivation preserves order); inherit the parent’s record_uuid.
   • Run against the live row-store via NATS (or directly mutate the JSON with the row-store stopped, the same pattern as the blank-row cleanup).

Once record_uuid lands everywhere, the enhanced-records-list microfrontend can be scaffolded — Svelte 5 + workspace-singleton pattern, port 3006 likely, register in shell/src/remotes.ts, pair with Record Collector per the spec’s PAIRINGS note.

Housekeeping when you open the next session

1. Commit the uncommitted stack. The 2026-05-22 work is in one branch’s worth of unrelated-but-coherent changes. Split into logical commits per the git-conventions skill:
   • feat(prompt-template-manager): Save / Apply ergonomics + sidebar deletes
   • feat(request-reviewer): firing indicator, cancel, coverage strip, fire-remaining
   • feat(response-reviewer, response-store): edited_text autosave, needs-human flag, triage chip relocation, icon actions with CSS tooltips, delete + clear all, filter counts
   • feat(row-store, response-reviewer): helpful_links on rows
   • feat(prompt-runner): per-row logging, timeout, cancellation
   • feat(ingest): drop fully-blank rows on import
   • Plus a changelog entry — Phase-5 is the natural framing: “the post-flight pre-checkpoint round.”
2. Check the running stack — docker ps should show all 8 containers up. The cleaned data is live.
3. Read the spec before writing any code: context-v/specs/Enhanced-Records-List-and-Promotion-Checkpoint.md.
4. Open this pickup again if anything below is unclear.

See also

• [[Enhanced-Records-List-and-Promotion-Checkpoint]] — the spec, the build target
• [[Helpful-Links-on-Records-Captured-During-Triage]] — already implemented, feeds into the spec
• [[Pickup-2026-05-23]] — the previous session’s pickup (now superseded by the work that’s landed since)
• [[Response-Reviewer-and-Response-Store]] / [[Request-Reviewer-Pre-Flight-Surface]] — the per-stage specs the new surface complements