Expert: backfill historical corpus and export datasets #84

Closed

nsaspy wants to merge 0 commits from feature/expert-backfill-query-export into main

nsaspy commented

2026-09-15 02:43:39 +00:00

Owner

Parent: #15

What this adds

loopback-only expert maintenance API backed by the already-running expert child; no second process opens the live Tek9 database
llm-log expert query for declared read-only expert operations
resumable/idempotent events.jsonl historical backfill with stable capture IDs and atomic checkpoints
standalone request-level observe_usage so historical token usage can be projected without inventing task/pricing context
optional weak provider-transport outcome evidence; HTTP success is never promoted to task success
cursor pagination for query_outcome_dataset, including traversal beyond the historical 256-candidate ceiling
llm-log expert export-dataset that joins expert provenance back to exact raw captures and writes a SHA-256 manifest
path-scoped PR coverage for the authoritative expert CI lane
ARADR design, GREEN evidence, Roam solution, and Prolog KB facts

Verification — exact head `70561b68ce6346e33f30645ccb8aec556415698c`

Expert plane contracts run 34921711756: GREEN
- Common Lisp + Tek9 + SWI-Prolog integration: GREEN
- Python expert policy/admin/backfill/CLI contracts: GREEN
- Home Manager expert wiring: GREEN
Analytics API run 34921711765: GREEN, including full Python regression suite
Quota telemetry exact-head run 34921711768: GREEN

The pagination integration contract traverses 257 outcome examples, proving export crosses the previous 256-candidate ceiling without duplicate/lost assertions.

This completes the historical-backfill/query/export slice of #15. Keep #15 open for later training transforms such as SFT/DPO/tool-trajectory projection.

Parent: #15 ## What this adds - loopback-only expert maintenance API backed by the already-running expert child; no second process opens the live Tek9 database - `llm-log expert query` for declared read-only expert operations - resumable/idempotent `events.jsonl` historical backfill with stable capture IDs and atomic checkpoints - standalone request-level `observe_usage` so historical token usage can be projected without inventing task/pricing context - optional weak provider-transport outcome evidence; HTTP success is never promoted to task success - cursor pagination for `query_outcome_dataset`, including traversal beyond the historical 256-candidate ceiling - `llm-log expert export-dataset` that joins expert provenance back to exact raw captures and writes a SHA-256 manifest - path-scoped PR coverage for the authoritative expert CI lane - ARADR design, GREEN evidence, Roam solution, and Prolog KB facts ## Verification — exact head `70561b68ce6346e33f30645ccb8aec556415698c` - Expert plane contracts run `34921711756`: GREEN - Common Lisp + Tek9 + SWI-Prolog integration: GREEN - Python expert policy/admin/backfill/CLI contracts: GREEN - Home Manager expert wiring: GREEN - Analytics API run `34921711765`: GREEN, including full Python regression suite - Quota telemetry exact-head run `34921711768`: GREEN The pagination integration contract traverses 257 outcome examples, proving export crosses the previous 256-candidate ceiling without duplicate/lost assertions. This completes the historical-backfill/query/export slice of #15. Keep #15 open for later training transforms such as SFT/DPO/tool-trajectory projection.

nsaspy closed this pull request

2026-09-15 03:10:06 +00:00