delegate-evidence
Use when the delegate-routing skill seems not to work, or before stating how Claude, Codex, Kiro or Kimchi actually behave: delegation, nesting, model or effort reach, system prompts, worktrees, hooks. Answer from the delegate map or the acceptance suite, never from memory.
- License 许可
- Unlicense
- 最近更新
- 2026-11-13
- Stars
- ★ 2
- Forks
- 0
01 前言 Front Matter
- name
- delegate-evidence
- description
- Use when the delegate-routing skill seems not to work, or before stating how Claude, Codex, Kiro or Kimchi actually behave: delegation, nesting, model or effort reach, system prompts, worktrees, hooks. Answer from the delegate map or the acceptance suite, never from memory.
03 SKILL.md 正文
Delegate evidence
Two sources answer questions about harness behavior. Use them instead of memory.
The map: how each harness behaves
packages/delegate-routing/docs/delegates/ holds the map. Start at its
README.md. Every fact cell carries an evidence mark and a case id, such as
claude:depth. evidence.md lists the pinned versions and the open unknowns.
To rerun a row, give its case id to the runner. It reads the case tables in
packages/delegate-routing/probes/delegates/<harness>/README.md, runs each
command and greps its output for the expected excerpt:
run=packages/delegate-routing/probes/delegates/run.py
python3 "$run" --list # every id, runnable or why not
python3 "$run" --only=kimchi/claude:s1-fg-pins # one case: <harness>/<case id>
python3 "$run" --only='claude:dmu_*,codex/codex:R1.*' # bare id or prefix*
It prints MATCH, MISMATCH or SKIP per case and exits with the MISMATCH count.
LIVE cases run only with --live. Its docstring holds the case-table format a
new row must follow.
The acceptance suite: does the skill route correctly
packages/delegate-routing/eval/ runs real sessions on Claude, Codex, Kiro and
Kimchi and checks what the explicit suite configuration makes an agent do. Its
fixtures enable their own runtimes and routing; personal devenv.local.nix is
outside those inputs. Validate every case and print each launch without starting
anything:
python3 packages/delegate-routing/eval/suite.py --dry-run
A live run spends model turns on the operator's logins, so the operator runs it.
--list prints the case ids; --only=<id>[,<id>|<prefix>*], --case <id> or
--harness <name> narrows a run; eval/README.md lists the operator steps.
Rules
- Make no claim about harness behavior without a map row or a run. If the map lacks it, add a case to that harness's probe directory and run it.
- Run login-free cases yourself. Cases marked OFFLINE, fake-provider or capture-mock cases need no account.
- For a case that needs a login, an account or a paid turn, give the operator a numbered list of steps: the setup, the exact command, and what output to send back.
When the operator reports the skill misbehaving
- Give the operator the
suite.py --only=<id>command for the reported behavior's cases. - Run the map rows behind the failing behavior, to see what the harness really does at the pinned version.
- Fix the skill, the configuration, or the map.
- Report what the evidence showed, citing the case ids and outputs.
04 同仓库更多 Skills
stack-test
Use when you need to run tests or formatters across commits in a stack. Use INSTEAD of manual git test run or looping git checkout + test. Prevents: testing every commit when only the tip has to pass, unbounded parallel jobs exhausting memory, cache misunderstandings.
stack-summary
Use when you need to audit, analyze, or review stack quality before restructuring. Produces a structured summary with per-commit classification, philosophy audit, and violation flags. Output feeds directly into /stack-plan restructure mode. Use INSTEAD of manual git log inspection.
stack-submit
Use when you need to push, submit, or create pull requests for a commit stack. Rebases only the current stack onto upstream main, validates, and pushes INSTEAD of manual git sync + git submit + PR/MR creation. Handles branch creation, stacked PR/MR creation with correct base branches.
stack-split
Use when you need to break up, split, or decompose a large commit into smaller atomic commits. Use INSTEAD of manual git rebase -i + edit or git reset HEAD^. Prevents: non-working intermediate commits, wrong split ordering, missed downstream restack.
stack-plan
Use when you need to plan commits, restructure a stack, or commit uncommitted work as organized atomic commits. Builds a stack from a description, uncommitted work, or existing commits INSTEAD of dumping everything into one monolithic commit with git add -A && git commit, or manual git rebase -i, git reset --soft, or ad-hoc git move sequences. Prevents: wrong commit ordering, forward references, git move -F panics, untested intermediates.
stack-fix
Use when you need to fix, correct, or update content in an earlier commit — line-level edits, content moves, or structural changes. Absorbs staged changes into correct stack commits INSTEAD of running git absorb or git commit --fixup via Bash, or hand-running git prev + edit + git amend + git restack --merge. Prevents: missed dry-run preview, leftover staged changes going unnoticed, forgetting to restack. Falls back to guided manual amend when absorb cannot route hunks.