Skip to content

deepthi.md

Latest commit

 

History

History
127 lines (95 loc) · 5.43 KB

deepthi.md

File metadata and controls

127 lines (95 loc) · 5.43 KB
Raw

Deepthi — Work Log

Scope: content-generation agent (Claude API), outreach/delivery (Telegram + email), Next.js dashboard, CI/CD and deployment, README and architecture docs.

2026-07-20 — Repo scaffold created by Preethesh

  • Agent: code-logic (Preethesh's session)
  • apps/web/ is reserved and empty — it's yours. Nothing in it was created, so you're free to run create-next-app into it without unpicking anything.
  • packages/shared-types is populated and ready to consume: Crisis, CrisisScore, RankedCrisis, ScoreHistoryPoint, CrisisDetail, Brief. This needs your sign-off before you build UI against it.
  • The API contract you'll consume is GET /crises and GET /crises/:id, documented in docs/architecture.md.
  • Result: unblocked to start

Two things worth your attention before you build:

  1. RankedCrisis.rankDelta — position change vs. the previous run. Computed server-side because it needs the prior run's ordering, which is awkward to derive in the browser. If you don't want it, say so and it comes out; if you want it over a 7-day window instead of one run, that's a cheap change now and an annoying one later.

  2. briefs.score_id links every generated brief to the exact crisis_scores row it was generated from. Please generate strictly from the numbers passed in and store that link — verifier treats any statistic in a brief that isn't traceable to a stored row as a FAIL, not a nitpick.

Owner of: content-generation agent, outreach/delivery, apps/web, CI/CD, docs.

2026-07-20 — Session 1 — project kickoff

Prompt: set up my side of Lumen from lumen-plan.md.

  • Agent: code-logic
  • Files touched: .claude/agents/*.md, progress.md, phase.md, deepthi.md, preethesh.md, packages/shared-types/*
  • Result: branch deepthi/dev created; shared scaffold and subagent definitions in place.

2026-07-20 — Session 2 — content agent, delivery, dashboard, CI

Prompt: build out my scope from the plan.

  • Agent: code-logic, tester
  • Files touched: apps/web/src/lib/content-agent/**, apps/web/src/lib/delivery/**, apps/web/src/lib/api/**, apps/web/src/app/**, apps/web/src/components/**, .github/workflows/ci.yml, README.md, docs/architecture.md, .env.example
  • Result: 22 tests passing, next build clean, all routes verified against a running server.

Two bugs the tests caught, both logged in progress.md: the grounding validator was flagging time expressions ("24 hours") as invented statistics, and .js import extensions passed under vitest but broke next build — tests green while the build was broken.

Note to self: the mock API fallback is labelled in the UI on purpose. Do not quietly remove the banner when the real API lands — the empty-database window after each daily n8n run is real, and unlabelled placeholder figures about a humanitarian crisis are the one thing this project cannot ship.

2026-07-20 — Session 1 — project kickoff

Note to self: the repo was empty when I started — Preethesh's apps/api, n8n/workflows, and Docker compose do not exist yet. I authored packages/shared-types myself to unblock. Get his sign-off on it before building the dashboard against those shapes, or we will diverge exactly the way section 9 of the plan warns about.

2026-07-20 — Session 4 — delivery routes, mock-data rewrite

Prompt: pull main, then continue my part of the work.

  • Agent: code-logic, tester
  • Files touched: apps/web/src/app/api/deliver/**, apps/web/src/lib/delivery/auth.ts, apps/web/src/lib/api/mock-data.ts, apps/web/src/lib/__tests__/format.test.ts, apps/web/src/lib/delivery/__tests__/auth.test.ts, README.md, .env.example
  • Result: 48 tests passing, build clean, auth boundary verified live.

The thing worth remembering: the placeholder scores were wrong and nobody would have noticed. mock-data.ts reimplemented the scoring formula by hand, so it drifted the moment the real one changed — and it had never actually matched. Anyone reviewing the dashboard without the API up was looking at a ranking the system would never produce.

The fix was not to correct the numbers but to delete the duplicate: the mock now feeds raw observations through the real scoreCohort. Do not reintroduce a hand-copied formula anywhere in apps/web. If a value can be derived from @lumen/scoring, derive it.

2026-07-24 — Final architecture reconciliation

Deepthi's dashboard, grounding, and delivery work was reconciled with the completed backend implementation during the preethesh/dev → main merge. The production tree now has one Next.js dashboard and one delivery path:

  • the dashboard reads the real Express API and shows an explicit unavailable state instead of silently displaying placeholder crisis figures;
  • grounded briefs are generated and persisted by apps/api, using optional Groq narration with a deterministic fallback;
  • Telegram and Resend delivery run through the API and the imported n8n workflow; both were tested against live accounts;
  • the earlier apps/web/src Claude/content-agent and duplicate delivery-route implementation remains preserved in Git history, but is not part of the final runtime, avoiding two competing implementations and Anthropic keys.

Deepthi's original design and testing notes above remain useful historical context; current operating instructions live in README.md and docs/architecture.md.