Prevent + net the unescaped-quote bug in the durable prompts/pipeline

The escape-ASCII-quote rule previously lived only in ephemeral Agent-call
strings. Bake it into the durable artifacts so the next session doesn't
re-derive it:
- SUBAGENT_PROMPT.md + ENRICHMENT_PROMPT.md: explicit rule to escape any
  ASCII " inside JSON string values (Romanian „cuvânt" is the trap).
- run_enrichment.py collect_enrichment: repair malformed parts with
  escape_stray_quotes instead of dropping them — the enrichment path had no
  repair net (bad parts were silently dropped, losing that activity's
  enrichment). Extraction already had one; now both do.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This commit is contained in:
Claude Agent
2026-05-29 18:16:04 +00:00
parent bcfb6841eb
commit d6971e47f8
3 changed files with 46 additions and 3 deletions

View File

@@ -74,6 +74,21 @@ The file is one JSON object: a `header` plus an `activities` array.
- Do **not** paraphrase the `source_excerpt` — copy it character for character.
- Better to extract fewer activities accurately than to pad the output.
## Escaping quotes inside JSON strings (CRITICAL)
Any ASCII double-quote (`"`, U+0022) that appears **inside a string value** must
be written escaped as `\"`. This is the single most common way these extractions
break: Romanian source text uses typographic quotes like `„cuvânt"` where the
closing mark is a plain ASCII `"`. Written raw, it terminates the JSON string
early and corrupts the whole file. So:
- `"description": "grupul cântă „Unu\" în cor"` ← correct (inner `"` escaped)
- `"description": "grupul cântă „Unu" în cor"` ← BROKEN (unescaped `"`)
Prefer keeping the source's typographic quotes (`„ "`), but whenever a literal
ASCII `"` lands inside a value, escape it. After writing, re-read the file and
confirm it parses as valid JSON.
## Writing large outputs in batches (IMPORTANT)
A single Write tool call has a hard ~32K output-token limit. Dense chunks