review · segments
[SOLO ORCHESTRATION CONTEXT] You are running inside Solo (process ID 909, project lounge-wt3 / ID 34). SOLO_PROCESS_ID=909, SOLO_PROJECT_ID=34. Call whoami() first; use scratchpad/todo/kv tools. If Solo MCP is NOT available, say so on stdout. [END CO
codex 263 events 1 segments wip/generative-personas
segment 1 of 1
Implement the P4 post-generation pipeline (ExemplarRetriever, PersonaContextAssembler, GenerationQc, RunArtifactWriter, PostGenerator, and genpersona:generate-post command)
The agent locked todo 611, performed discovery by reading Solo scratchpad 973, canonical docs (README, SPEC, DECISIONS), seed cohort KV, existing generation client/factory, models (GenerativePersona, Archetype, MinHash, BannedString, RedditPostBody, TextMinhash), schema, and prior commands (ModelSmoke, ClassifyCohort). The agent gathered context to align with decisions D11 (transient QC), D13 (JSON artifacts), and D17 (provider-agnostic client), but did not create any new service classes or the command.
outcome
No new code written; only reconnaissance and context gathering completed.
next steps
- Create ExemplarRetriever service to fetch exemplar posts from the DB using seed cohort account names and subreddit.
- Create PersonaContextAssembler to build system/user messages for the LLM from persona data + exemplars.
- Create GenerationQc service with transient MinHash similarity (no DB writes) and banned-string flagging.
- Create RunArtifactWriter to write JSON artifacts under storage/app/genpersona/runs.
- Create PostGenerator orchestrator that ties the above steps together and calls GenerationClientFactory::make()->chat().
- Create genpersona:generate-post command with --persona, --sub, --count options.
- Validate by generating 2 posts for each of nyc-lowkey-weeknight-couple and charleston-social-soft-sell-couple and confirming distinctness.
- Write findings to scratchpad and comment on todo 611.
key decisions
- Follow D11: QC is transient/non-blocking, no new DB rows created.
- Follow D13: Generated posts live in JSON run artifacts under storage/app/genpersona/runs.
- Follow D17: Generation client must be resolved via GenerationClientFactory::make() and remain swappable through config.
- Use DB reads for exemplar retrieval (via raw queries or DB facade) to avoid Eloquent event dispatch.
- Read banned strings from existing BannedString model rows and match with contains/like checks in PHP.
- Compute MinHash signature in memory and compare with text_minhashes rows only via LSH candidate fetch (no writes).
open questions
- Exact shape of the JSON artifact schema (fields to include for a generated post).
- How to handle the cold start Latency of the first model call (~10s) during validation.
- Whether the PersonaContextAssembler should include exemplar post bodies or just titles.
- The specific banned-string filter behavior: should it flag or reject? Per spec it's non-blocking, so flag only.
1 week ago → 1 week ago