Parser Refine

---
description: Recursive parser refinement loop. Opus designs heuristic → haiku applies + runs tests → repeat until accuracy target hit.
allowed-tools: Read, Edit, Write, Bash, Grep, Glob, Agent, Skill
argument-hint: <pdf-path> [target-accuracy=0.95]
---

# /parser-refine

Recursive accuracy-loop for parser heuristics. Capped at 5 iterations to prevent runaway opus cost.

## Inputs

- `$1` — PDF path
- `$2` — target accuracy (default 0.95)

## Loop

iteration = 0 while accuracy < target AND iteration < 5:

1. haiku: run /parser-test cli $1 → record count, totals
2. haiku: run python3 Scripts/compare_parsers.py $1 → record expected count
3. compute accuracy = swift_count / python_count
4. if accuracy >= target → DONE
5. opus: read at most 2 missing-txn diff samples, propose ONE heuristic change
6. haiku: apply opus's diff
7. haiku: rebuild + retest iteration++

## Hard rules

- opus is invoked AT MOST 5 times per loop
- opus produces a diff < 30 lines per iteration
- if accuracy decreases between iterations → REVERT haiku's last change + STOP
- if iteration 5 hits and accuracy < target → write `parser_refine_log.md` and STOP

## Workflow agent invocations

Agent(subagent_type: haiku-agent, prompt: "run /parser-test cli $1, report Txns=N Dr=X Cr=Y") Agent(subagent_type: opus-agent, prompt: "given delta of M missing txns in , propose ONE heuristic change to . Output unified diff only.") Agent(subagent_type: haiku-agent, prompt: "apply this diff and rerun /parser-test cli $1: ")

## Output

=== /parser-refine $1 === iter 0: 94 / 429 → 21.9% iter 1: 142 / 429 → 33.1% (heuristic: column-X tolerance widening) iter 2: 268 / 429 → 62.5% (heuristic: narration-line dedup) ... final: 419 / 429 → 97.7% PASS

## Cost guard

If opus is invoked 5x and target not hit, the loop ends with a written failure report. Do NOT silently keep retrying.