[rparundekar]

What this does

Adds agent.py beside the counterparty news workflow. This is the guide's overnight fallback: articles the workflow sends to review go to Claude. Claude reads the borrowers' credit files and reruns the workflow's own event check with the link fixed. A judge in code then checks every quote, borrower and event, and allows only the outcomes that each review reason permits. Everything goes to the credit analyst's queue, and nothing becomes an alert.

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

10 of 15 jobs came out as expected.

• N-04, the shared name: 5 of 5.
• N-07, the ticker conflict: 5 of 5.
• A-01: 0 of 5. It was written to land between the alert and filed scores, but the workflow filed it as routine on every run, so it never reached the agent. It's reported as a miss, not tuned. It also means the guide's "unsure score" example hasn't fired on our data yet.

recipe.py is unchanged, and its outputs are identical before and after. 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 551eb47 and abf77ee. Each fix has a selfcheck case that fails if the fix is undone.

Important (5, all fixed)

• The judge couldn't tell shared-name candidates apart. For N-04, the event check with the link fixed reads the same sentences for CP-1004 and CP-1005, so a finding that picked the wrong Northwind still came out finding_checked. Fixed: the judge now runs the check for every candidate. finding_checked needs the pick's result to differ from every other candidate's; otherwise the status is call_not_checked. N-04's expected outcome changed to call_not_checked because review found this gap, not because of a live result, and the README says so.
• Quotes were a bare substring match, so "a" or a field label counted as evidence. Fixed: a quote must be at least 4 whole words, taken from a field value. For a link, the article quote must contain the name in question.
• suggest_not_a_borrower needed no evidence. Fixed: it now needs an article quote and a quote from the candidate's file.
• Scores were only checked on links, so a made-up 0.99 could appear in the analyst's notes. Fixed: on other outcomes the score must be empty or match the workflow's own number.
• A retry where every call failed stopped the whole overnight run. Fixed: the first-pass results are kept, and those articles go to the analyst as sie_failed.

Suggestions (all fixed)

• Tool scope: credit files and lookups reach only the row's candidates, and names or tickers that appear in the article.
• review_notes.md: text is escaped, the note is fenced, and the judge rejects a note containing a link or HTML.
• Tool input: checked against each tool's schema. Real errors are no longer reported to Claude as "bad arguments".
• Judge checks: cached per job, and only the final submission's checks are written out.
• README: the Mermaid chart shows the retry path, "every figure" now says figures of two or more digits, and the job count is honest.
• Anthropic errors: classified as transient or fatal. Tokens are priced on the model that answered, and a fallback model with no listed price shows as unpriced rather than guessed.

Reruns after the fixes (5 live runs)

• N-04 came out call_not_checked on 5 of 5 runs.
• N-07 came out left_for_person on 5 of 5 runs.
• A-01 was still filed by the workflow, so it never reached the agent, and it stays a reported miss.
• No live job reached finding_checked. That path is exercised only in selfcheck.py, and the README says so.