meta-apply
GitHub作为特权落地网关,仅允许人类用户授权将经跨模型陪审团审核通过的元优化补丁应用到技能库。通过独立 jury 验证和血缘检查,确保自我修改的安全性,防止未授权或错误的变更生效。
Trigger Scenarios
Install
npx skills add wanshuiyin/Auto-claude-code-research-in-sleep --skill meta-apply -g -y
SKILL.md
Frontmatter
{
"name": "meta-apply",
"description": "Privileged applier that LANDS meta-optimize \/ corpus-audit patches the user approved — the ONLY skill permitted to mutate the skill corpus from a self-modification proposal, with cross-model jury and human approval at landing. Use when the user says \"meta apply\", \"\/meta-apply\", \"land the staged patches\", \"应用优化\", after a \/meta-optimize run.",
"allowed-tools": "Bash(*), Read, Write, Edit, Grep, Glob, mcp__codex__codex, mcp__codex__codex-reply",
"argument-hint": "[patch-number-or-all]"
}
Meta-Apply: the privileged landing gate for self-modification patches
This skill exists to enforce a privilege boundary. Producers like
/meta-optimize (and, later, corpus-audit) are
read-only — no Write/Edit, no apply step; they can only stage candidate patches
under .aris/meta/pending/. This skill is the only place a staged patch becomes a real
change to the corpus. Splitting "propose" from "land" across two skills with different
tool grants is what makes "a loop cannot apply its own patch" structural rather than a
sentence the producer is asked to obey.
It is human-invoked only. It runs when the user explicitly types /meta-apply after
reading the producer's REPORT. That invocation IS the landing authorization (the locked
2026-05-30 decision: the human stays at the landing gate).
The acquittal is generated HERE, not relayed
The earlier design trusted a jury_verdict: pass field in the producer's manifest. That
is forgeable — the producer writes that field, so a buggy/misaligned producer could
stage a killed patch labelled "pass." So this skill does not trust any producer-written
verdict. It runs the cross-model jury itself, at landing time, on the actual staged
diff (fresh thread, paths-only). The verdict is produced where it is consumed, by the
privileged human-invoked skill — nothing to forge.
The non-negotiable rules (jury-at-landing, reject-default)
For each staged patch the user asks to land, in order — any failure ⇒ skip & report, never silently apply:
- The human named THIS patch. Apply only patches the user listed (
/meta-apply 1,3orall); default to applying nothing. - Fresh cross-model jury PASS, obtained now. Run
mcp__codex__codex(fresh thread, NOT codex-reply;model: gpt-5.6-sol,config: {"model_reasoning_effort": "ultra"},sandbox: read-only, paths-only perreviewer-independence.md) on the staged.diff+ its target. Ask: does this change improve the harness without regressions; PASS or KILL + one-line reason. Include the scope-limits block fromreview-scope-limits.mdin that prompt: this jury judges ARIS's own mechanism, so an over-defensive KILL permanently blocks a good patch. Note the block bans proposing new hash binding — it is not a reason to KILL a patch that touches the existing provenance stamp. KILL ⇒ refuse. The human cannot override a KILL — they may only pick among jury-PASSED survivors. (A loop can DRIVE; only the cross-model jury can ACQUIT.) - Author ≠ reviewer family. The author is the producer's executor model; the reviewer
is the codex model that just judged it. Run
provenance.py assert_cross_family— if it raises (same family / unknown), refuse. (Here it always holds: producer=Claude, jury=codex. The check is the structural backstop.)
Workflow
Step 0: Load staging + resolve the helper
PENDING=".aris/meta/pending"
[ -d "$PENDING" ] || { echo "Nothing staged. Run /meta-optimize first."; exit 0; }
echo "Staged:"; cat "$PENDING/manifest.jsonl"
Resolve provenance.py via the 4-layer chain in
integration-contract.md §2
(.aris/tools/ → tools/ → $ARIS_REPO/tools/ → $ARIS_REPO/tools/ via
~/.aris/repo).
Step 1: Jury-at-landing for each requested patch
For every patch the user asked to land, read its staged .diff and target, then run the
fresh codex jury (Rule 2) — paths-only, no producer reasoning, no prior-round context.
Record {patch, jury_verdict, jury_thread_id, one_line_reason}. Print a one-line result
per patch (PASS → eligible / KILL → refused: <reason>).
The producer may have written an advisory pre-screen into the manifest to help the human read the REPORT — ignore it for the landing decision. Only this fresh verdict counts.
Step 2: Land the survivors (Write/Edit only — never Bash)
For each patch that PASSED Step 1 and was named by the user:
- Back up the target to
.aris/meta/backups/<date>/<target>(use the Write tool to copy contents; corpus paths are not Bash-writable whencorpus_write_guardis active — and the applier should use Write/Edit for corpus mutation anyway). - Apply the diff by Edit/Write on the target corpus file.
- Stamp provenance on the changed file:
python3 "$PROVENANCE" stamp "$TARGET" --author "$AUTHOR" \ --reviewer "$JURY_MODEL" --verdict-id "$JURY_THREAD_ID"stamp()re-asserts cross-family and refuses on same-family — the structural backstop at the moment the authorization record is written. The stamp is a process receipt (who authored, who acquitted-at-landing, content hash) — NOT a claim the change is correct. - Log to
.aris/meta/optimizations.jsonl:{ts, patch, target, author_model, reviewer_model, jury_thread_id, applied: true}.
Step 3: Report
Per patch: LANDED <target> (+ backup path + provenance sidecar) or
REFUSED <patch>: <reason>. Remove landed patches from .aris/meta/pending/. Remind the
user a landed patch is revertable from its backup, and to test the changed skill next run.
Provenance is a receipt, not an acquittal of correctness
A stamp records that a change passed a process (cross-model jury at landing + human landing), not that it is correct. To prevent "approved-but-wrong with a stamp that vouches for it" (false-authority laundering — worse than no stamp, because a later auto-curator reads it as evidence):
- The stamp carries
verdict_id(auditable review) +content_hash(a later hand-edit invalidates it). - Recommended (not yet built): a TTL forcing re-review of long-lived auto-authored artifacts, and a behavioral auditor that REVOKES a stamp when a landed skill misbehaves. Track as follow-up; never treat a stamp as permanent truth.
Key Rules
- Human-invoked only. Never run as a side-effect of another skill or a hook.
- Jury-at-landing, reject-default, no override. The binding verdict is produced HERE on the staged diff; never trust a producer-written verdict; the human picks among survivors, never resurrects a KILL.
- Cross-family or refuse.
assert_cross_familymust not raise. Adeterministic:<verifier>reviewer is valid per skill-governance.md. - Corpus mutation goes through Write/Edit (reviewable, attributable), not Bash. The
corpus_write_guardhook (if installed) additionally denies Bash corpus writes — it does NOT gate Write/Edit, so it does not by itself stop this skill from editing the corpus; the jury-at-landing + stamp discipline above is what governs Write/Edit mutations (that discipline is procedure, not a hook-enforced mechanism). - Back up before every mutation. Reversible by construction.
- Only land staged patches. Applies what producers staged in
.aris/meta/pending/; invents nothing of its own.
Review Tracing
Save each landing-jury codex call's trace per
review-tracing.md to
.aris/traces/meta-apply/<date>_run<NN>/ — the acquittal that landed a corpus change must
be forensically recoverable.
Version History
-
f4f20f9
Current 2026-08-20 05:00
新增 review-scope-limits.md 共享参考文件,限制评审者提议范围(如禁止提议新的哈希绑定),并在 meta-apply 等9个提示点插入该限制,以解决 gpt-5.6-sol 过度防御问题。
- 53562a7 2026-07-25 10:41


