Fixes #9

Upload a bank/credit-card CSV and have the AI map an arbitrary layout to normalized transactions. The AI proposes rows; the user reviews/edits them in a staging table and approves per-row; only approved rows are committed via the existing direct journal writer. Builds on the merged #8 AI plumbing (shared Anthropic client + Settings-page key) — no new dependency, no config changes.

How it works

pages/import.vue → useImport → /api/import/parse  (structured output, NEVER writes)
                             → /api/import/commit (the ONLY write path, approved rows only)

• Parse (parse.post.ts): one-shot Anthropic structured output (messages.parse + json_schema) — not a tool loop. Grounds the model's account/envelope suggestions in real targets fetched from hledger; returns normalized proposals + dropdown context + any rows it couldn't parse. Never touches the journal.
• Commit (commit.post.ts): the only write path. Re-validates each row server-side, dedups against the journal, and writes via appendSimplifiedTransaction — extracted from transactions.post.ts so manual-add and import share one envelope-accounting path. Partial success ({ committed, skippedDuplicates, failed }).
• UI: editable staging grid (ImportReviewTable.vue), /import page with a persistent data-egress notice, no-API-key empty state, and a sidebar nav entry.

Notable design decisions

• Structured outputs over a tool loop — purpose-built for one-shot CSV→JSON, compatible with adaptive thinking, no forced tool_choice.
• Dedup is a flag, not a hard block — hashes date|cents|payee against the journal itself (no separate import ledger → stateless). Genuinely-identical same-day rows are surfaced as "possible duplicate" for the user to confirm rather than silently dropped; the commit-time journal check is what prevents accidental re-import of a committed batch.
• Uncategorized handling — outflows require an envelope before approval; uncategorized inflows commit as income → Ready to Assign.
• Row cap MAX_IMPORT_ROWS=200, non-streaming (streaming/chunking deferred).

Data egress (Issue #9 risk note)

CSV rows go to the Anthropic API — the one external data flow. No bank credentials, no aggregator, no stored secrets beyond ANTHROPIC_API_KEY. The import page shows a persistent notice; CSV contents and the key are never logged.

Spec

Design-first under .kiro/specs/ai-csv-import/ (design → requirements → tasks). Two small deviations recorded in the spec: the model echoes the verbatim sourceRow for reliable provenance; commit needs no API key (local write) — original R7.2 said "parse/commit".

Verification

• npx vitest run → 58 files pass (39 new tests, incl. the load-bearing "parse writes 0×" safety test and refactor-parity on existing routes).
• npx nuxi typecheck → clean (exit 0).
• npm run build → succeeds.
• No vitest.config / tsconfig / nuxt.config / package.json changes.
• Not run: the live AI round-trip (upload → real Anthropic parse → commit) — needs a configured ANTHROPIC_API_KEY and a live API call. Everything verifiable statically is green; the end-to-end UI run is left to confirm with a key configured.

🤖 Generated with Claude Code