peptide-repo/prautoblogger
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

fix(phase2b): Authority hold-on-quorum-miss no longer crashes + pipeline_mode label (#184, v0.32.2)

Browse Source 
Canary-found P1: hold with empty content (research quorum miss) recorded a HOLD decision without calling Publisher (no more empty-content RuntimeException → run held, not failed). Content-present holds unchanged. Authority posts now labelled pipeline_mode=authority. QA APPROVE f759d75; CI green (phpcs+phpunit VPS-verified).

Agent-Session: prab-phase2b-holdfix-merge
This commit is contained in:
peptiderepo
2026-06-24 19:59:50 +07:00
committed by GitHub
parent 81aba2ac23
commit 181e20fb17
10 changed files with 437 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
CONTEXT.md
View File

@@ -171,7 +171,7 @@ into the live Economy path until P2b.4).
| **Research_Fanout** | `PRAutoBlogger_Research_Fanout` — the orchestrator class that implements the fan-out pattern. Reserves the SUMMED worst-case cost of all N agents from the cost governor before dispatch, applies quorum enforcement after, and returns per-agent result arrays to the caller. Not live until P2b.4 wires it into the tier router. |
| **curate stage** | The pipeline stage (`stage = 'curate'`) that follows fan-out. Implemented by `PRAutoBlogger_Research_Judge`. Deduplicates sources from all agents, scores each by `quality_score`, and writes the top-12 results to `run_sources`. Already present in `Stage_Display_Map` since v0.18.0. |
| **Research_Judge** | `PRAutoBlogger_Research_Judge` — the class that executes the curate stage. Dedup chain: URL-canonical → semantic cosine (OpenRouter embeddings) → keyword-overlap fallback. Scores via `Research_Source_Scorer`. Writes `kept=1` / `kept=0` rows to `run_sources` with `quality_score` and `reason`. |
| **quorum** | Minimum number of research agents that must return usable results for the fan-out to proceed. Formula: ⌈N/2⌉+1 (one strict majority). Example: N=3 → quorum=3; N=5 → quorum=4. If fewer than quorum agents return valid results, `Research_Fanout::dispatch()` returns an empty array and the caller **must hold the run** rather than proceeding on thin research. |
| **quorum** | Minimum number of research agents that must return usable results for the fan-out to proceed. Formula: ⌈N/2⌉+1 (one strict majority). Example: N=3 → quorum=3; N=5 → quorum=4. If fewer than quorum agents return valid results, `Research_Fanout::dispatch()` returns an empty array and the pipeline holds the run (status `held-quorum`). v0.32.2: quorum-miss hold no longer calls Publisher — only a `run_decisions` row is written (no draft post created; Publisher would throw on empty content). |
| **specialist role** | One of `{mechanisms, clinical, safety, comparison, practical}`. Each agent in the fan-out is assigned a distinct role and receives a role-specific user prompt. The first N roles in priority order are used (N = `prautoblogger_research_agent_count`). Written to `run_stages` as `agent_role = 'researcher:<role>'`. |
| **quality_score** | A composite score for a source: `relevance × authority_weight`. `relevance` comes from the agent's self-reported score (0.0–1.0). `authority_weight` is the source-type multiplier from `Research_Source_Scorer`. Higher scores sort the source closer to the top-12 kept set. Stored in `run_sources.quality_score`. |
| **authority weight** | A source-type multiplier applied by `PRAutoBlogger_Research_Source_Scorer` to produce `quality_score`. Values: DOI/PubMed/NCBI peer-reviewed → 1.0; .gov/.edu/WHO/FDA institutional → 0.85; preprint → 0.70; HTTPS → 0.60; HTTP → 0.40. Reflects the editorial preference for peer-reviewed sources in Authority-tier articles. |