Description
A committed corpus is finite, so someone watching the demo for a long time will spot repeats. The chosen hybrid keeps the corpus committed and reviewable, and adds a tool that uses a cheap model (Haiku on Bedrock) to propose new reader questions, follow-ups and developer prompts in the corpus formats. A person reviews the output before it is committed, so nothing unreviewed reaches a live demo. Spend must be bounded and visible.
Acceptance Criteria
- #1 A just recipe generates candidate corpus entries for a named persona, team or scenario and writes them to a review file, not straight into the corpus
- #2 Generated entries are validated against the corpus schema and deduplicated against existing entries
- #3 The generator runs with a hard spend cap, and its calls show up as a traced agent in Agent Observability
- #4 docs describe the review-then-commit workflow
Definition of Done
- #1 just check
Implementation Plan
- just recipe runs a Haiku-on-Bedrock generator for a persona, team or scenario, writing a review file. 2. Validate against corpus schema and dedupe. 3. Hard spend cap, traced via the Agent Observability SDK. 4. Docs for review-then-commit.
Implementation Notes
Live check 2026-09-29: grow.mjs –kind reader-intent –name best-price –count 5 –max-usd 0.10 against the lab Haiku application inference profile wrote 5 candidates, 0 rejected, 0.0017 USD of the 0.10 cap, to a review file (not the corpus). Generation 47c50cef landed in Agent Observability as agent touchline-corpus-gen with its own conversation and usage. CodeRabbit fixes applied before commit: malformed model output rejects only that attempt; accept.mjs checks every target and duplicate stems before writing anything. Recipes sit in the dev group (fleet justfile rule allows six fixed groups). just check: code gates green; its only failure is a scrub hit in the parent AIO-0001 plan text, cleared separately.
Final Summary
Added just corpus-grow / corpus-accept (apps/agents/corpus-gen): Haiku on Bedrock proposes reader phrasings, follow-ups, intents or developer prompt files for a named persona, intent, team or scenario; candidates are schema-validated and deduplicated (exact and token-Jaccard >= 0.8) into a review file, and only reviewer-accepted entries are merged. Hard spend cap per run (default 0.50, ceiling 2.00 USD), traced as touchline-corpus-gen. Verified by 52 unit tests with an injected model and one live run whose generation shows in Agent Observability. Docs: docs/traffic-corpus.md.