Skip to content

feat(performance): hand protect's fragile-optimization guardrails to /corrections:codify #17172

Description

@kyle-sexton

Parent

Refs #15703
Source: #16029, part 2 of #15729. The list recorded in #16602 gives /performance:protect the route "call /corrections:codify, presence-gated, with the strongest-first fallback".

Depends on

Outcome

Protect's deliverable, a counter ceiling checked in CI, stays as it is, and it already hands the ceiling to /toolchain:ratchet when that is available. Its section 5 then picks the guardrail for a fragile optimization from its own list (generate the copy and add a drift test; keep the differential as a standing test; a handoff test). After this change section 5 hands each guardrail's mechanism to /corrections:codify when it is available, with the shared strongest-first fallback when it is not, and still says the ratchet alone passes a subject that got cheaper by breaking.

Demo: with a corrections stand-in plugin loaded, run /performance:protect on a verified win that came from a precomputed table. The output proposes the ceiling as today and names /corrections:codify for the guardrail that keeps the table and its source in step, instead of choosing a test itself.

Where the text changes

Line numbers as read on main at a483a2c1; they may move. Owner decision of 2026-11-18 on #16029 (option A): the codify call covers only the choice the skill makes beyond its fixed deliverable, which stays as it is.

File and lines Present wording Change
plugins/performance/skills/protect/SKILL.md:189-200 (section 5, "Guardrails for fragile optimizations") "A fast path that copies a slow one ...: generate the copy from the source, and add a test that fails when the two drift. A fast path that must match the real one: keep verify's differential as a standing test ... A handoff between two implementations: a test that drives input through the switch ... Propose these beside the ratchet." Keep the three fragility cases as what to guard; the mechanism for each goes to /corrections:codify (gate plus shared fallback). Keep "The ratchet alone passes a subject that got cheaper by breaking."
same file, :219 (Output, "Guardrails proposed") "Guardrails proposed: <tests, or none needed and why>" "Guardrails proposed" names the codify hand-off or the fallback proposal

Call form: /corrections:codify "<what must not drift>".

Decisions

Key interfaces

  • "A skill that carries no ladder copy calls /corrections:codify, presence-gated, with a two-sentence strongest-first fallback. Only a skill that applies its own selection rule holds a synced copy and gets a row in the ladder's Selection rules table." (List enforcement-placing skills and give each one route to the ladder #15729)
  • The seam-phrasing convention: an availability gate, a stated fallback beside it, the plugin named and never the marketplace.
  • /toolchain:ratchet keeps owning the ceiling, as the "Who writes the ceiling" section says.

Done when

  • Section 5 hands each guardrail's mechanism to /corrections:codify behind the availability gate, followed by the shared fallback, and names no mechanism of its own.
  • Sections 1 to 4, the ceiling hand-off to /toolchain:ratchet and the Boundary list are unchanged.
  • The skill gains no ladder copy or Selection rules row, and no rung list is restated.
  • The pull request body quotes the sentence that picked a mechanism from a private list before the change. If the recheck finds none, the skill is left unchanged and the finding is reported on List enforcement-placing skills and give each one route to the ladder - part 2 of 2 #16029.
  • One new eval case loads a corrections stand-in plugin, in the shape of session-flow's re-mechanical-finding-check-proposal case, and grades that a fragile optimization's guardrail is handed to /corrections:codify while the ceiling is still proposed.
  • The pull request body names the existing case whose run reads the changed text (candidate: pr-ratchets-verified-spawn-count), or says none does. The lane runs no eval.
  • A changelog fragment under .changes/performance/ describes the change.
  • AC10: no new enforcement-placing skill is added here.
  • Open the pull request with Closes this issue in its body, then stop: the lane adds no label.

After the pull request opens (host session, not the lane)

The pull request stays held with do-not-merge for the eval; a host session runs only the eval and posts its record; the merge lane releases the hold when the record matches the head.

  • Hold line: the hub's sweep (or a host session) adds do-not-merge and posts, from a listed human's account, a comment with the line <!-- pr-hold reason=eval --> on a line of its own.
  • Run budget. List enforcement-placing skills and give each one route to the ladder #15729 asks for "that skill's cases on the opus and sonnet aliases, three runs per side, and no more than 240 runs per plugin". The owner decision of 2026-11-16 in the owner-decisions record (hop 4 Q15, which names List enforcement-placing skills and give each one route to the ladder #15729, amended by hop 7 Q3) replaces the run counts: per changed skill, one targeted before-and-after case on sonnet at 3 runs a side; opus runs only to confirm a sonnet loss; a new case gets a 1 to 3 run smoke first. Select cases with --case, use --runs 3 and a --max-cost-usd ceiling, and start no batch after a usage-limit error. Ceiling for this plugin: 1 skill x (3 smoke + 6 sonnet + 6 opus confirm) = 15 runs, under the parent's 240.
  • Control: the replaced list is a partial cut, so the pull request body carries a detection-control result. No paired no-loss compare (#15703 comment 6906800940).
  • On a LOSS reading: a trace read of each dropped case and an opus confirm; only a loss traced to the changed text counts.
  • Record: the verdict and the control result go into the pull request body. When the run is VALID and shows no loss, the host session posts, from a listed human's account, a comment with the line <!-- eval-pair verdict=no-loss validity=VALID head=<40-hex head SHA> --> naming the current head, keeping "do not merge" and do-not-merge out of that comment (docs/contracts/loop-lane/README.md, "Cross-lane PR hold"). A loss or a run that is not VALID leaves the hold in place; say so on the pull request.
  • Release: pr-merge's scheduled step "Release eval holds whose record matches the head" removes do-not-merge once that record names the current head and no other hold stands. A push after the record needs a new record.

Out of scope

Related

Activity

  1. added
    agent-readyFully specified and briefed; eligible for autonomous pickup from the frontier.
    on Nov 18, 2026
  2. added
    needs-triageNot yet classified. Floor until a type and one priority tier are set.
    on Nov 18, 2026
  3. github-actions commented on Nov 18, 2026

    @github-actions
    Contributor

    Size: Small
    Complexity: Low
    Uncertainty: Known
    Flags: none
    Blocked by: none
    Blast radius: section 5 and the Output line of the performance plugin's protect skill, plus one changelog fragment; people and agents who run /performance:protect read it.
    Reversible: yes; revert the skill-text edit and the fragment.
    Open questions: none
    I understand the request this way. Section 5 of /performance:protect should hand the mechanism for each fragile-optimization guardrail to /corrections:codify when that skill is available. It should use the shared strongest-first fallback otherwise. The ratchet sentence and the ceiling hand-off to /toolchain:ratchet stay as they are, and the Output line and a changelog fragment are updated. The issue text already settles intent, so nothing more is needed from you. The two host-run done-when items (the pull request quote and the new eval case) are split out by the refinement comment. I found no duplicate or merged fix. #17163, #17162, #17160, #17159, #17158 and #17174 are sibling follow-ons for other skills, not duplicates.

    Automated triage. A maintainer reviews it.

  4. removed
    needs-triageNot yet classified. Floor until a type and one priority tier are set.
    on Nov 18, 2026
  5. melodic-automation-lanes commented on Nov 18, 2026

    @melodic-automation-lanes
    Contributor

    Not admitted to the issue lane: its done-when asks for what an unattended lane run cannot perform: item 4 (tracker-output: "The pull request body quotes the sentence that picked a mechanism from a priv..."), item 5 (eval: "One new eval case loads a corrections stand-in plugin, in the shape of sessio..."); recommended split: keep the other items here as the admissible issue and move these to a child issue for the host queue (host-run). Admission takes the issue once each of those is met, when intake triages it again (.github/actions/resolve-config/admission-rules.json).

  6. melodic-automation-lanes commented on Nov 18, 2026

    @melodic-automation-lanes
    Contributor

    Refinement. A read-only run explored this repository, searched its issues and pull requests, and read the allowed docs to draft what follows. On an issue a listed human filed, it counts as the issue's spec once posted, while unedited; it is still model text, so check it.

    Intended outcome: Section 5 of /performance:protect hands each fragile-optimization guardrail's mechanism to /corrections:codify behind an availability gate with the shared fallback, keeping the ratchet sentence, with the Output line and a changelog fragment updated.

    Split. Done-when items 4, 5 of the issue body need a person: they are checks a person runs, listed below, and the lane does not do them. This issue keeps the rest of its done-when.

    Not done by the lane: checks a person runs

    External state: none

    Related work

    Sources read

    • plugins/performance/skills/protect/SKILL.md:189-200: section 5 lists three fragility cases (generate the copy plus drift test; verify's differential as standing test; a handoff test) and ends 'Propose these beside the ratchet. The ratchet alone passes a subject that got cheaper by breaking.'
    • plugins/performance/skills/protect/SKILL.md:219: Output line reads 'Guardrails proposed: <tests, or none needed and why>'
    • plugins/performance/evals/pr-ratchets-verified-spawn-count/case.yaml: existing performance case that may read the changed text
    • plugins/session-flow/evals/re-mechanical-finding-check-proposal/case.yaml: the model shape for the new stand-in eval case

    Automated refinement by the lanes App. Intake triages the issue again with it.

  7. melodic-automation-lanes commented on Nov 18, 2026

    @melodic-automation-lanes
    Contributor

    {"v":1,"actor_id":337595936,"title_sha256":"6928aa459a7b381bc45572d882897b48b99e5f38c30c89719779565ff0217095","body_sha256":"2888319875cc59e55e29b20a064e734d9aaed09544380b6126788f21b3add6e2"}

  8. melodic-automation-lanes commented on Nov 18, 2026

    @melodic-automation-lanes
    Contributor

    {"v":1,"actor_id":337595936,"title_sha256":"6928aa459a7b381bc45572d882897b48b99e5f38c30c89719779565ff0217095","body_sha256":"2888319875cc59e55e29b20a064e734d9aaed09544380b6126788f21b3add6e2"}

  9. github-actions commented on Nov 18, 2026

    @github-actions
    Contributor
    No description provided.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    agent-readyFully specified and briefed; eligible for autonomous pickup from the frontier.

    Type

    Fields

    Work status

    needs-triage

    Priority

    medium

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions