agent-wiki-compare-out…
Compare successful and failed normalized agent trajectories to derive evidence-backed agent-wiki guidelines. Use when Codex has multiple runs for the same or…
Publish a private guideline to a configured write-scope repo.
$ npx -y skills add AgentToolkit/altk-evolve --skill evolve-lite-publish --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/evolve-lite-publishContext preview
The summary Claude sees to decide when to auto-load this skill.
Publish a private guideline to a configured write-scope repo.
name: evolve-lite:publish description: Publish a private guideline to a configured write-scope repo.
Publish one or more private guidelines from `.evolve/entities/guideline/` into a configured **write-scope** repo. The entity is stamped with `visibility: public`, `owner`, `published_at`, and `source`, moved into the local clone of the write repo, and committed / pushed to the remote.
The same local clone is also what `evolve-lite:sync` pulls from — so you and anyone else publishing to the same repo stay in sync.
Read `evolve.config.yaml`. If no entry has `scope: write`, tell the user:
> "You need at least one write-scope repo to publish to. Run evolve-lite:subscribe with --scope write to set one up, then come back."
Then stop.
If `identity.user` is missing, ask for it and add it to the config.
Ensure `.evolve/` is gitignored at the project root:
grep -qxF '.evolve/' .gitignore 2>/dev/null || echo '.evolve/' >> .gitignore
Filter `repos:` to entries with `scope: write` (Step 1 already aborted if there were zero, so at least one exists here).
Let `{repo}` be the chosen repo name and `{branch}` its configured branch (default `main`).
List files in `.evolve/entities/guideline/` and ask the user which to publish.
For each selected file, run:
python3 .bob/skills/evolve-lite-publish/scripts/publish.py --entity "{filename}" --repo "{repo}" --user "{identity.user}"Build `{names}` as a comma-joined list of selected filenames, and `{guideline_paths}` as a space-joined list of the corresponding `guideline/{filename}` paths inside the clone (the files the publish script just wrote).
git -C ".evolve/entities/subscribed/{repo}" add -- {guideline_paths}
git -C ".evolve/entities/subscribed/{repo}" commit -m "[evolve] publish: {names}"
git -C ".evolve/entities/subscribed/{repo}" push origin "{branch}"On push success, continue to Step 7.
If the push fails and stderr mentions `rejected` / `non-fast-forward` / `fetch first`, another writer pushed to `{branch}` in between. Rebase the local commit and push once more:
git -C ".evolve/entities/subscribed/{repo}" fetch origin "{branch}"
git -C ".evolve/entities/subscribed/{repo}" rebase "origin/{branch}"review. Do not `git rebase --continue` or `git push` without an explicit user confirmation.
1. `git -C ".evolve/entities/subscribed/{repo}" status --porcelain` lists the conflicted paths. If any are `UD`, `DU`, or binary, skip to the abort step — those aren't safe to auto-resolve. 2. For each `UU`/`AA` file, read the conflict markers. During a rebase, `<<<<<<< HEAD` is the **remote's** version and the section under the commit sha is the **publish change** being replayed (opposite of a regular merge). Write an intent-preserving resolution; don't `git add` yet. 3. Show the user the diff (`git -C ".evolve/entities/subscribed/{repo}" diff HEAD -- {file}`) per resolved file with a one-line strategy summary, and ask whether to **continue** (stage + `rebase --continue` + push) or **abort** (roll back for manual resolution). 4. On **continue**:
git -C ".evolve/entities/subscribed/{repo}" add {resolved-files}
git -C ".evolve/entities/subscribed/{repo}" rebase --continue
git -C ".evolve/entities/subscribed/{repo}" push origin "{branch}"Then Step 7. If `rebase --continue` surfaces a new conflict, loop from step 1. 5. On **abort** — user declined, conflict isn't safely resolvable, or the proposed merge feels unsafe:
git -C ".evolve/entities/subscribed/{repo}" rebase --abortThe local publish commit is preserved at `.evolve/entities/subscribed/{repo}` but not on the remote. Tell the user to either (a) resolve manually in that directory (`git fetch origin {branch} && git rebase origin/{branch}`, fix conflicts, `git add` + `git rebase --continue`, `git push origin {branch}`) or (b) re-run `evolve-lite:publish` with a different filename if the conflict is a shared name.
If the push fails for any other reason (auth, network, missing remote ref), surface git's error and stop — rebase will not help.
Tell the user what was published and to which repo.
the write-scope clone at `.evolve/entities/subscribed/{repo}/guideline/`, with `visibility: public`, `owner: {user}`, `published_at`, and `source` stamped in frontmatter
Coding agents repeat the same mistakes because they start fresh every session. Evolve gives agents memory — they learn from what worked and what didn't, so each session is better than the last.
Repo: AgentToolkit/altk-evolve
Compare successful and failed normalized agent trajectories to derive evidence-backed agent-wiki guidelines. Use when Codex has multiple runs for the same or…
Read all atomic guidelines in wiki-twobatch/guidelines/ and propose themed clusters that group near-duplicates. Writes cluster pages and updates _config.yaml;…
Consult an agent-wiki for guidelines relevant to the task at hand. The wiki itself documents how to retrieve from it (AGENTS.md). Use this skill once you know…
Read a normalized Claude Code trajectory JSON and extract reusable guidelines into wiki-twobatch/guidelines/. Use when mining saved trajectories for reusable…
Ingest one or more agent trajectories (raw bob/claude traces or normalized JSON) into an agent-wiki end-to-end — convert, summarize, extract guidelines,…
Read a normalized Claude Code trajectory JSON and write an episodic summary page to wiki-twobatch/summaries/. Use when summarizing one or more saved…