Meta-Apply: the privileged landing gate for self-modification patches
Before each reviewer slot, follow Selectable Reviewers:
resolve invocation → .aris/reviewer.json → Claude Code default (Codex MCP)
with reviewer_route.py through the canonical helper resolver, preserving this
slot's tier, fresh contexts, source/image inputs and deterministic gates.
Codex examples/pins below apply only to backend=codex; other routes use their
own first-call/reply contract. Save the route and actual returned identity;
unavailable selections are REVIEW_UNAVAILABLE. Classify acceptance from actual
families/completion/scope. Forward invocation reviewer/model/effort overrides
to reviewer-bearing children. Generation workers are unchanged.
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 1,3
or all); default to applying nothing.
- Fresh cross-model jury PASS, obtained now. Run
mcp__codex__codex (fresh thread,
NOT codex-reply; model: gpt-6-astra, config: {"model_reasoning_effort": "ultra"}, sandbox: read-only,
paths-only per reviewer-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 from
review-scope-limits.md in 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 actual returned model that just judged it. Run
provenance.py assert_cross_family — if it
raises (same family / unknown), refuse. Use actual author/reviewer identities, never infer the family from
the host or transport. This strict landing gate remains mandatory.
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/ → $ARIS_REPO/tools/ via
~/.aris/repo).
Step 1: Jury-at-landing for each requested patch
For every patch the user asked to land, read its staged .diff and target, then run the
fresh codex jury (Rule 2) — paths-only, no producer reasoning, no prior-round context.
Record {patch, jury_verdict, jury_thread_id, one_line_reason}. Print a one-line result
per patch (PASS → eligible / KILL → refused: <reason>).
The producer may have written an advisory pre-screen into the manifest to help the
human read the REPORT — ignore it for the landing decision. Only this fresh verdict
counts.
Step 2: Land the survivors (Write/Edit only — never Bash)
For each patch that PASSED Step 1 and was named by the user:
- Back up the target to
.aris/meta/backups/<date>/<target> (use the Write tool
to copy contents; corpus paths are not Bash-writable when corpus_write_guard is
active — and the applier should use Write/Edit for corpus mutation anyway).
- Apply the diff by Edit/Write on the target corpus file.
- Stamp provenance on the changed file:
python3 "$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_family must not raise. A
deterministic:<verifier> reviewer is valid per skill-governance.md.
- Corpus mutation goes through Write/Edit (reviewable, attributable), not Bash. The
corpus_write_guard hook (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.
1---2name: meta-apply-23description: "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."4argument-hint: "[patch-number-or-all]"5allowed-tools: Bash(*), Read, Write, Edit, Grep, Glob, mcp__codex__codex, mcp__codex__codex-reply, mcp__antigravity__antigravity, mcp__antigravity__antigravity-reply, mcp__grok__grok, mcp__grok__grok-reply, mcp__claude-review__review, mcp__claude-review__review_reply, mcp__claude-review__review_start, mcp__claude-review__review_reply_start, mcp__claude-review__review_status6metadata:7 source: https://skillmd.com/skills/wanshuiyin/meta-apply-28---910# Meta-Apply: the privileged landing gate for self-modification patches1112> **Before each reviewer slot**, follow [Selectable Reviewers](../shared-references/reviewer-backends.md):13> resolve invocation → `.aris/reviewer.json` → Claude Code default (Codex MCP)14> with `reviewer_route.py` through the canonical helper resolver, preserving this15> slot's tier, fresh contexts, source/image inputs and deterministic gates.16> Codex examples/pins below apply only to `backend=codex`; other routes use their17> own first-call/reply contract. Save the route and actual returned identity;18> unavailable selections are `REVIEW_UNAVAILABLE`. Classify acceptance from actual19> families/completion/scope. Forward invocation reviewer/model/effort overrides20> to reviewer-bearing children. Generation workers are unchanged.212223This skill exists to enforce a **privilege boundary**. Producers like24[`/meta-optimize`](../meta-optimize/SKILL.md) (and, later, `corpus-audit`) are25**read-only** — no `Write`/`Edit`, no apply step; they can only *stage* candidate patches26under `.aris/meta/pending/`. **This skill is the only place a staged patch becomes a real27change to the corpus.** Splitting "propose" from "land" across two skills with different28tool grants is what makes "a loop cannot apply its own patch" structural rather than a29sentence the producer is asked to obey.3031It is **human-invoked only.** It runs when the user explicitly types `/meta-apply` after32reading the producer's REPORT. That invocation IS the landing authorization (the locked332026-05-30 decision: the human stays at the landing gate).3435## The acquittal is generated HERE, not relayed3637The earlier design trusted a `jury_verdict: pass` field in the producer's manifest. That38is **forgeable** — the producer writes that field, so a buggy/misaligned producer could39stage a killed patch labelled "pass." **So this skill does not trust any producer-written40verdict.** It runs the cross-model jury **itself, at landing time, on the actual staged41diff** (fresh thread, paths-only). The verdict is produced where it is consumed, by the42privileged human-invoked skill — nothing to forge.4344## The non-negotiable rules (jury-at-landing, reject-default)4546For each staged patch the user asks to land, in order — any failure ⇒ skip & report,47never silently apply:48491. **The human named THIS patch.** Apply only patches the user listed (`/meta-apply 1,3`50 or `all`); default to applying nothing.512. **Fresh cross-model jury PASS, obtained now.** Run `mcp__codex__codex` (fresh thread,52 NOT codex-reply; `model: gpt-6-astra`, `config: {"model_reasoning_effort": "ultra"}`, `sandbox: read-only`,53 paths-only per [`reviewer-independence.md`](../shared-references/reviewer-independence.md))54 on the staged `.diff` + its target. Ask: *does this change improve the harness without55 regressions; PASS or KILL + one-line reason.* Include the scope-limits block from56 [`review-scope-limits.md`](../shared-references/review-scope-limits.md) in that prompt:57 this jury judges ARIS's own mechanism, so an over-defensive KILL permanently blocks a58 good patch. Note the block bans *proposing new* hash binding — it is not a reason to59 KILL a patch that touches the existing provenance stamp. **KILL ⇒ refuse.** The human60 cannot override a KILL — they may only pick among jury-PASSED survivors. (A loop can61 DRIVE; only the cross-model jury can ACQUIT.)623. **Author ≠ reviewer family.** The author is the producer's executor model; the reviewer63 is the actual returned model that just judged it. Run `provenance.py assert_cross_family` — if it64 raises (same family / unknown), refuse. Use actual author/reviewer identities, never infer the family from65 the host or transport. This strict landing gate remains mandatory.6667## Workflow6869### Step 0: Load staging + resolve the helper7071```bash72PENDING=".aris/meta/pending"73[ -d "$PENDING" ] || { echo "Nothing staged. Run /meta-optimize first."; exit 0; }74echo "Staged:"; cat "$PENDING/manifest.jsonl"75```7677Resolve `provenance.py` via the 4-layer chain in78[`integration-contract.md`](../shared-references/integration-contract.md) §279(`.aris/tools/` → `tools/` → `$ARIS_REPO/tools/` → `$ARIS_REPO/tools/` via80`~/.aris/repo`).8182### Step 1: Jury-at-landing for each requested patch8384For every patch the user asked to land, read its staged `.diff` and target, then run the85fresh codex jury (Rule 2) — paths-only, no producer reasoning, no prior-round context.86Record `{patch, jury_verdict, jury_thread_id, one_line_reason}`. Print a one-line result87per patch (`PASS → eligible` / `KILL → refused: <reason>`).8889> The producer may have written an *advisory* pre-screen into the manifest to help the90> human read the REPORT — **ignore it for the landing decision.** Only this fresh verdict91> counts.9293### Step 2: Land the survivors (Write/Edit only — never Bash)9495For each patch that PASSED Step 1 **and** was named by the user:96971. **Back up** the target to `.aris/meta/backups/<date>/<target>` (use the **Write** tool98 to copy contents; corpus paths are not Bash-writable when `corpus_write_guard` is99 active — and the applier should use Write/Edit for corpus mutation anyway).1002. **Apply** the diff by **Edit/Write** on the target corpus file.1013. **Stamp provenance** on the changed file:102 ```bash103 python3 "$PROVENANCE" stamp "$TARGET" --author "$AUTHOR" \104 --reviewer "$JURY_MODEL" --verdict-id "$JURY_THREAD_ID"105 ```106 `stamp()` re-asserts cross-family and refuses on same-family — the structural backstop107 at the moment the authorization record is written. The stamp is a **process receipt**108 (who authored, who acquitted-at-landing, content hash) — NOT a claim the change is109 correct.1104. **Log** to `.aris/meta/optimizations.jsonl`:111 `{ts, patch, target, author_model, reviewer_model, jury_thread_id, applied: true}`.112113### Step 3: Report114115Per patch: `LANDED <target>` (+ backup path + provenance sidecar) or116`REFUSED <patch>: <reason>`. Remove landed patches from `.aris/meta/pending/`. Remind the117user a landed patch is revertable from its backup, and to test the changed skill next run.118119## Provenance is a receipt, not an acquittal of correctness120121A stamp records that a change passed *a process* (cross-model jury at landing + human122landing), not that it is *correct*. To prevent "approved-but-wrong with a stamp that123vouches for it" (false-authority laundering — worse than no stamp, because a later124auto-curator reads it as evidence):125126- The stamp carries `verdict_id` (auditable review) + `content_hash` (a later hand-edit127 invalidates it).128- **Recommended (not yet built):** a TTL forcing re-review of long-lived auto-authored129 artifacts, and a behavioral auditor that REVOKES a stamp when a landed skill misbehaves.130 Track as follow-up; never treat a stamp as permanent truth.131132## Key Rules133134- **Human-invoked only.** Never run as a side-effect of another skill or a hook.135- **Jury-at-landing, reject-default, no override.** The binding verdict is produced HERE136 on the staged diff; never trust a producer-written verdict; the human picks among137 survivors, never resurrects a KILL.138- **Cross-family or refuse.** `assert_cross_family` must not raise. A139 `deterministic:<verifier>` reviewer is valid per skill-governance.md.140- **Corpus mutation goes through Write/Edit** (reviewable, attributable), not Bash. The141 `corpus_write_guard` hook (if installed) additionally denies Bash corpus writes — it142 does NOT gate Write/Edit, so it does not by itself stop this skill from editing the143 corpus; the jury-at-landing + stamp discipline above is what governs Write/Edit144 mutations (that discipline is procedure, not a hook-enforced mechanism).145- **Back up before every mutation.** Reversible by construction.146- **Only land staged patches.** Applies what producers staged in `.aris/meta/pending/`;147 invents nothing of its own.148149## Review Tracing150151Save each landing-jury codex call's trace per152[`review-tracing.md`](../shared-references/review-tracing.md) to153`.aris/traces/meta-apply/<date>_run<NN>/` — the acquittal that landed a corpus change must154be forensically recoverable.