--- name: kev-knowledge description: Search the Kev knowledge graph (private repo jaredpalmer/kev-knowledge) before re-deriving history. Use when asked what was tried, decided, measured or learned in earlier Kev sessions, why a rule or number exists, what a round or PR did, or when starting a research, serving, release or data task that earlier sessions may have covered. Also covers how the graph is refreshed (nightly Devin Cloud automation) and how to add to it. allowed-tools: - read - grep - glob - exec permissions: allow: - Exec(gh repo clone jaredpalmer/kev-knowledge) - Exec(rg) - Exec(qmd) - Read(/tmp/kev-knowledge/**) - Read(~/dev/kev-knowledge/**) title: kev-knowledge canonical_url: https://skilld.dev/gh/jaredpalmer/kev/kev-knowledge last_updated: 2026-10-06T05:23:00.000Z --- > Index of every page on this site: [llms.txt](https://skilld.dev/llms.txt) > **Skill from skilld.dev.** Follow the user's requested action. > If the user asked to fork this Skill, follow the fork workflow below. Do not execute the Skill instructions while copying it. > Otherwise, follow the instructions below for this session. You do not need to install anything. > > If the user asked to install this Skill, run `npx skilld install jaredpalmer/kev/kev-knowledge`. Install writes the Skill files into the project, so every session loads them. > > ## Fork workflow > > A fork creates an editable local Skill with its original author and licence. The request authorizes copying and local installation. > 1. Check `./skills/kev-knowledge`, the project lockfile, and selected Agent targets together. If the local directory or installed Skill exists, stop. Never overwrite an existing directory or Agent target. > 2. Read [source metadata](https://skilld.dev/api/v1/skills/jaredpalmer/kev/kev-knowledge) once. Use sourceUrl, sourceCommit, skillPath, sourceGone, and license. If the source is gone or its path is missing, stop. If license is null, read licence files at the source commit. > 3. Fetch only the source commit into a temporary Git repository. Do not clone full history. Derive repository_url from sourceUrl, including repository renames. If sourceCommit is absent, resolve the sourceUrl ref once. Set source_commit to that actual commit. Run these commands in one shell call: > > ```sh > git init --quiet "$temporary_dir" > git -C "$temporary_dir" fetch --quiet --depth=1 "$repository_url" "$source_commit" > git -C "$temporary_dir" checkout --quiet --detach FETCH_HEAD > ``` > > Read applicable licence declarations and notices at that commit. If copying is not permitted, report the restriction and stop. > 4. Inspect source entries together, then copy the directory containing skillPath into `./skills/kev-knowledge`. Use the user's path if selected. Keep the original SKILL.md, relative links, scripts, binary assets, and executable modes. Exclude .git metadata. Reject symlinks and paths outside the Skill directory. After checking entries, use cp -a where available. A regular source directory needs no custom copy script. Do not save this page wrapper as SKILL.md. > Preserve author credit, notices, and applicable licence files from repository or parent directories. Add PROVENANCE.md with the Skill page, source URL, actual commit, original path, and licence. Retain any existing PROVENANCE.md and record new provenance separately. Batch source inspection, copying, and provenance work where practical. > 5. In the project root, run `skilld install ./skills/kev-knowledge --mode copy --plain`. If skilld is unavailable, use `npx skilld install ./skills/kev-knowledge --mode copy --plain`. This known command needs no help lookup. Install does not support --json. Use detected Agent targets, or add --agent for the targets the user selected. Install the local path, never the upstream selector. If installation fails, preserve the local copy and report the exact failure. > 6. Confirm the local lockfile source and installed Agent copies once. Report the local path, actual commit, and Agent targets. After edits, reinstall the same local path. Upstream updates must not replace it. Do not publish or push unless the user asks. # Kev knowledge graph Every Devin session on this repo since 17 Sep 2026 is distilled into the private repo `jaredpalmer/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 (`digest/`). A nightly Devin Cloud automation reads new sessions through the Devin API and writes the notes (`automation/nightly.md` there). Use it instead of re-deriving history from git or guessing why a rule exists. It is a separate private repo on purpose: transcripts hold private names and verbatim conversation; never copy `digest/` text into this repo or any public surface. ## 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. ## Get it ```sh [ -d /tmp/kev-knowledge ] || gh repo clone jaredpalmer/kev-knowledge /tmp/kev-knowledge # or ~/dev/kev-knowledge on Jared's Mac git -C /tmp/kev-knowledge pull --ff-only -q ``` If the clone is refused, the session has no access to the private repo: say so and fall back to PLAN.md, the git log and `research-archive-2026-09-24`; do not guess at what the graph would have said. ## Search it `rg` always works and is enough for most questions; qmd adds semantic search if it is installed. ```sh K=/tmp/kev-knowledge cat $K/index.md # the map: topics with session counts, sessions by date rg -n -i "" $K/topics $K/sessions # identifiers, PR numbers, flags, suite names rg -n -i "" $K/timeline.md $K/prs.md # when, and which sessions touched a PR rg -n -i "" $K/digest # quotes from transcripts (long files; read with sed -n 'a,bp') ``` ```sh # optional: npm install -g --allow-scripts=node-llama-cpp @tobilu/qmd (models download on first use) cd $K && qmd update && qmd embed -c kev # the checked-in .qmd/index.yml builds a project-local index; no trust prompt qmd search "" -c kev -n 8 # BM25, instant qmd vsearch "" -c kev -n 8 # semantic, seconds qmd query "" -c kev -n 8 # hybrid + LLM rerank, best quality, slow on CPU qmd get "kev/topics/.md" qmd search "" -c kev-transcripts -n 5 # transcripts are excluded from default search ``` Read order: `topics/` (distilled, trustworthy) -> `sessions/.md` (what happened, the exact asks and outcomes) -> `digest/.md` (transcript, for quotes). ## Map | file | use | |---|---| | `index.md`, `timeline.md` | hub; the story since 17 Sep 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, the 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, the thermonuclear standard | | `sessions/warm-lute.md` | the nine-day CLI session that ran rounds 4-29 and every 27B release | | `sessions/cloud-<12 hex>.md` | Devin Cloud sessions; each links its app.devin.ai URL | | `prs.md`, `issues.md` | every PR/issue -> the sessions that discussed it (`prs.md#pr-`) | ## Citing Name the note path (`kev-knowledge/topics/calibration.md`) or the session id. Numbers in the notes were written by hand from transcripts; for a published number prefer `docs/claims.json`, PLAN.md or the model card here, and say which you used. ## Adding to it The nightly automation picks up every Devin Cloud session on its own; nothing to do for normal work. To record something immediately, or to add a lesson the automation would not infer, open a PR on `jaredpalmer/kev-knowledge` editing `tools/session_notes.py` or `topics/*.md`, run `python3 tools/build_graph.py` there, and never edit the generated files (`sessions/*.md`, `index.md`, `prs.md`, `issues.md`) by hand. The procedure the automation follows, and the by-hand refresh (`tools/extract_cloud_sessions.py`, needs `DEVIN_API_KEY`), are in that repo's README and `automation/nightly.md`.