docs: extraction blueprint — the ollama calling layer as a consumable package - #231
Merged
Merged
Conversation
… package Synthesis of four independent max-effort reviews (wire contract on both sides of the API, ladder+lifecycle, capture+provenance, coupling audit) against main @ 95715da, blocker claims re-verified by hand. The calling discipline is extractable — salvage.py, sampling.py and summarize_engine_compare.py lift unchanged — but each best mechanism has unsanctioned rivals: four capped definitions (two of which pool capped cells into published means), identity and provenance living in filenames and stdout, a second payload builder in preflight, and a ladder whose correctness rests on cold restarts nothing verifies (MLX runners never reload on option changes). The blueprint fixes first in-repo (P0/P1), extracts second, and documents the wire contract a consumer must know — including fork-vs-stock deltas invisible in the response shape (pass-summed eval_count, silently dropped fork options) and three measurement follow-ups (M1 challenges the 2026-08-19 endpoint entry's interpretation; measurement decides, per the log's REFUTED convention). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Synthesis of four independent Opus max-effort reviews of the ollama calling layer (wire contract on both sides of the API, ladder+lifecycle, capture+provenance, coupling audit), run against main @ 95715da with blocker claims re-verified by hand before synthesis.
Verdict: extractable. Three modules lift unchanged today (salvage.py, sampling.py, summarize_engine_compare.py); client.py needs shimming, not rewriting — its omission sentinels and tri-state think are the API. What blocks a clean package is enforcement, not mechanisms: four "capped" implementations (two pool capped cells into published means), block identity living in filenames, provenance in stdout, a second payload builder in preflight, and a ladder whose correctness rests on cold restarts nothing verifies (MLX runners never reload on option changes — verified at sched.go).
The blueprint: package layers + design rules; the condensed wire contract a consumer must know (incl. fork-vs-stock deltas invisible in the response shape: pass-summed eval_count, silently dropped fork options, the format:null chat two-pass trigger); a ranked fix-first list (P0 correctness / P1 robustness / P2 measurement follow-ups, including M1 which challenges the 2026-08-19 endpoint entry's interpretation — to be settled by re-measurement per the log's REFUTED convention, not edits); and the strengths inventory to preserve verbatim.
🤖 Generated with Claude Code