kev-knowledge

# Kev knowledge graph (qmd) Every Devin CLI session on this repo since 17 Sep 2026 is distilled into `~/dev/kev-knowledge`: hand-written topic notes (what we know), one note per session (what happened), a day-by-day timeline, PR and issue indexes with back-links, and the extracted transcripts. It is indexed by `qmd` as the collection `kev`. Use it instead of re-deriving history from git or guessing why a rule exists. It lives outside the repo on purpose (transcripts are noisy and may hold private names); never copy its `raw/` or `digest/` text into the repo. ## When to look - "Did we try X?", "why is Y the rule?", "what did round N find?", "what did PR #N change and why?", "what did Jared decide about Z?" - search first, then cite the note. - Before planning a round, a serving change, a data family, a release or a benchmark: read the matching topic note. - When AGENTS.md or PLAN.md states a rule without the incident behind it, the topic notes have the incident. ## How to search Two collections: `kev` (the distilled notes: index, timeline, topics, sessions, prs, issues; searched by default) and `kev-transcripts` (`digest/` + `raw/`; excluded from default queries, name it with `-c`). ```sh qmd search "<keywords>" -c kev -n 8 # BM25, instant; best for identifiers (PR numbers, flags, suite names) qmd vsearch "<question>" -c kev -n 8 # semantic, ~3 s qmd query "<question>" -c kev -n 8 # hybrid + LLM rerank, best quality; ~20 s on this Mac (CPU, no Metal) qmd query "<question>" -c kev --files --min-score 0.3 # paths only qmd get "kev/topics/<slug>.md" # read a note (line-numbered) qmd multi-get "kev/sessions/*.md" -l 30 # skim the first 30 lines of every session note qmd search "<exact phrase>" -c kev-transcripts -n 5 # quotes from the transcripts ``` Read order: `topics/` (distilled, trustworthy) -> `sessions/<id>.md` (what happened, with the exact asks and outcomes) -> `digest/<id>.md` (prose transcript, for quotes) -> `raw/<id>.md` (tool calls, truncated outputs; last resort). Transcript hits are long; read them with `qmd get "kev-transcripts/digest/<id>.md:<line>:<count>"` rather than whole. Without qmd: `rg -n "<term>" ~/dev/kev-knowledge/topics ~/dev/kev-knowledge/sessions`, then open `~/dev/kev-knowledge/index.md`. ## Map | file | use | |---|---| | `index.md` | hub: topics with session counts, sessions by date | | `timeline.md` | the story 17 Sep - 1 Oct 2026, with open threads | | `topics/architecture.md`, `training-recipe.md`, `full-weight-sft.md`, `calibration.md` | the model and how it is trained; what moved the needle | | `topics/evaluation-suites.md`, `research-rounds.md`, `autoresearch-program.md` | suites, gates, the audit that removed scienthoon/wanli/typesafe; every round's verdict; the standing rules | | `topics/serving-performance.md`, `deployment-paths.md`, `long-context.md`, `modal-infrastructure.md` | CUDA graphs, batching, fused kernels, MLX, vLLM decision, 64k states, GPUs and spend | | `topics/releases.md`, `docs-and-writing.md`, `data-and-synthetic.md`, `competitors.md`, `base-models.md` | what shipped when; README/card rules; data policy and teachers; Jev/AutoJev/Clef; Qwen -> Gemma/Inkling | | `topics/skills-and-workflow.md`, `code-quality.md` | worktrees, subagents, Devin CLI tips, the thermonuclear standard | | `sessions/warm-lute.md` | the nine-day marathon that ran rounds 4-29 and every 27B release | | `prs.md`, `issues.md` | every PR/issue -> sessions that discussed it (`prs.md#pr-<n>`) | ## Citing Quote the note path (`~/dev/kev-knowledge/topics/calibration.md`) or the session id (`devin -r warm-lute` resumes it). Numbers in the notes were written by hand from the transcripts; for a published number, prefer `docs/claims.json`, PLAN.md or the model card, and say which you used. ## Refreshing after new sessions ```sh cd ~/dev/kev-knowledge python3 tools/extract_sessions.py # read-only over ~/.local/share/devin/cli/sessions.db python3 tools/build_graph.py # exits non-zero naming any session that lacks a note qmd update && qmd embed -c kev && qmd embed -c kev-transcripts # the second is slow (tens of minutes); optional ``` If `qmd` is missing: `npm install -g --allow-scripts=node-llama-cpp @tobilu/qmd`, then `qmd collection add ~/dev/kev-knowledge --name kev --mask "*.md,topics/**/*.md,sessions/**/*.md"` and `qmd collection add ~/dev/kev-knowledge --name kev-transcripts --mask "digest/**/*.md,raw/**/*.md"` + `qmd collection exclude kev-transcripts`; the contexts are listed in `~/dev/kev-knowledge/README.md`. For each new session: read `digest/<id>.md`, add an entry to `tools/session_notes.py` (title, summary, asks, outcomes, lessons, topics), extend the topic notes it taught something new, add the day to `timeline.md`, rebuild. `tools/fetch_github.sh` refreshes `data/prs.tsv` / `data/issues.tsv` (needs `gh`). Do not edit `sessions/*.md`, `index.md`, `prs.md` or `issues.md` by hand; they are regenerated. `README.md` there documents the layout.