[rparundekar]

What this does

Adds agent.py beside the invoice check. An investigator asks about one repair shop's year of claims. The agent runs the invoice workflow on every claim as its verifier. Code builds the per-claim table from those results, and Claude only reports patterns, each labelled a finding or a question. Code then checks that:

• the claim list is complete;
• every quote is a value the workflow read;
• every finding rests on a failed check.

That last check means a closed claim the workflow refused to check can't be turned into a finding.

Live results (5 runs, using SIE Serverless, 10 Nov 2026)

5 of 5 as expected.

• The refusal: on one run the agent called the closed claim a finding. The code sent it back, and the revision made it a question.
• Noise: every run also asked a needless question about the shop's own address. It's reported under "Where it fails".

recipe.py and data/ are unchanged. Depends on the docs PR.

🤖 Generated with Claude Code

[rparundekar]

Self-review

Two review passes ran on this PR. One read the code: correctness, silent failures, tests and comments. The other covered security and checked the README against the code. Every finding below was fixed in b6afbb2, and each fix has a selfcheck case that fails if the fix is undone.

Important (2, both fixed)

• agent.py claim_rows / check_report: a field the reader missed counted as a finding against the shop. A check flagged only because the OCR missed a field ("Not read: tax", conflicting_values, "no claim number read") was filed as a finding against the shop. A finding pattern resting on it passed, and the job ended report_ready. Fixed: those claims are now not_checked with the reason "a field wasn't read", the job ends incomplete, and no finding can rest on them. A mix of unread and real mismatches stays a finding, citing only the real failures.
• agent.py report_markdown: invoice text went into siu-report.md as raw Markdown or HTML. A shop could print an image link in its bill-to, and it would load when the analyst opened the report. Fixed: every cell goes through md_safe, and the summary, notes and request are fenced.

Suggestions (all fixed)

• The tools could reach other shops and claims. claims_for_shop and get_claim now refuse anything outside the job's vendor, year and claim list, so the README's "can't reach siu-09" holds by construction.
• A claim with two invoices kept only the last one. Each claim now keeps a list of results.
• Anthropic errors: 408, 409 and 5xx count as transient, and any other error stops the run with its own message first. Tokens are priced on the model that actually answered.
• README: a finding may name a clear claim next to the failed one, and the README now says so. The token figure is exact, not an average.

Reruns after the fixes (5 live runs, 11 Nov 2026)

All 5 runs came out as expected, with the same per-claim table each time. On run 1 the agent filed the closed claim as a finding; the code sent it back, and the revision made it a question. The noise question about the shop's own address now shows up on 3 runs in 5, down from 5 in 5. Claude cost about $0.09 a job.