generated-runtime-repairs · git:20261118.7b3f760 · 2026-11-18 · sha256 f38094a1d3553bed
generated-runtime-repairs git:20261118.7b3f760A
Immutable. This exact content never changes and is served at /api/v1/blob/f38094a1d3553bed.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_jh3ninvtvmf3jytc/f38094a1d3553bed6565c0db9ef378c3dedbf62eaab6743ba62f91331a6468c4.json
Then /plugin install generated-runtime-repairs@markdownregistry-jh3ninvtvmf3jytc in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
--- name: generated-runtime-repairs description: >- Repair repository-owned wrappers and scheduled routines without confusing source correctness with deployment. Use when a generated file misbehaves and the fix belongs in its template or generator, or when a reviewed patch must be landed and then materialized into a live runtime. Not for editing generated output by hand. metadata: revision: "2" areas: "software-development,work-management" targets: "runtime-agnostic" summary: "repair the template/generator, prove the staged artifact, then deploy and read back the live runtime" summary_zh: "修模板/生成器、验证暂存产物,再部署并回读线上运行时" kind: "entry" --- # Generated runtime repairs Repair repository-owned wrappers and scheduled routines without confusing source correctness with deployment. Load the repository contract and the relevant runtime skill for its supported commands. ## When to Use - A repo owns a template/routine that is materialized into agent profiles, cron jobs, or wrapper scripts. - A generated file misbehaves and the fix must go to the template/generator, not the generated copy. - A reviewed candidate patch must be landed into a shared checkout and/or deployed to the live runtime. ## 1. Freeze scope and current inputs - Verify relayed permission changes against the operator's primary record and current contract. Classify source edits, Git publication, and live runtime writes separately; standing commit/push authorization does not authorize profile/config/cron/session writes. - Fetch and fast-forward before source work. Inspect current status/diff and patch applicability immediately before landing; peers and scheduled publishers can invalidate an earlier clean-tree report. When landing a candidate built against an earlier base, prove each target file byte-identical to that base (`git show <base>:<path>` piped to `cmp` against the working copy) — the shared HEAD moving on does not by itself invalidate the candidate, but a clean `git apply --check` alone does not prove the base content did not drift. - Read the generator, template/routine, and actual consumer record. Enumerate all writes a broad apply performs before selecting an operation; role-limited apply may still reconcile identity, jobs, config, or databases. - Trace routine text into stored prompts. If changing the body changes the expected cron prompt, include that exact prompt-only write in the runtime decision or report it as undeployed. Do not promise both "cron untouched" and "no new prompt drift". ## 2. Reproduce and repair at the source - Reproduce through the existing generated wrapper under the reported PATH and cwd. An unmaterialized template with unresolved placeholders fails for a different reason and is not a valid reproduction. - Discover masked interpreter paths from globs, not abbreviated tool output. Compare the same real payload under the reported interpreter and a capability-bearing interpreter. - Build a red-capable focused test before editing the generator. Assert the exact output and return code; shell-pipeline success is not the payload's exit status. - Fix the template and generator, never only the generated file. Probe required capabilities rather than Python versions; pin the selected launcher, preserve a runtime override, and fail explicitly before executing the payload. ## 3. Prove the staged artifact - Regenerate an isolated fixture home through the actual generator and check it again. Exercise its wrapper under the hostile environment; label fixture results and staged real-payload smoke tests separately from live-profile results. - Inspect lease lookup before gate tests. Release only your own editing lease before tests that observe it; a fixture registry may still consult a module-level checkout lease and suppress expected work. Do not change gate logic or release another writer's lease to manufacture green tests. - Run focused regressions before required full-suite and portability/validation gates. Report `ran / failures / errors / skipped`; distinguish the historical baseline total from the patched suite with new cases. Skipped cases are not executed passes. - If shared source changes during verification, test the repair with a fresh snapshot in isolation. Never overwrite the other writer's changes or label a failed unchanged-source comparison as preservation proof. ## 4. Land source without mixing other writers' work - Follow the actual standing publication contract and publisher guards. An `add -A` sweep does not preserve hunk isolation: make an authorized explicit selective commit before its sweep, or coordinate the publication boundary. Never stash, stage, or rewrite another writer's work to manufacture a clean landing. - Stage explicit paths/hunks, run the required validation and secret scan, and preserve history. After an authorized push, re-fetch and compare the exact local and remote refs before claiming synchronization. - A role-scoped write checker flags role-definition and test files as out-of-surface violations by design — they are no role's runtime write surface, and their acceptance runs through the generator's contract tests. Report the expected violation as such instead of routing the write through a role, widening the checker, or reading it as a failed gate. - For change-only routines, keep the canonical baseline in persistent state, never scratch; read it before replacement. With no old baseline, record the first observation and state that comparison is unavailable, not "no changes". Preserve the existing comparison dimensions. - Express baseline paths relative to the owning agent's persistent state home in portable instructions; keep harness-specific resolution in the binding. Verify that the scheduled consumer receives the new instructions, not just that the source file contains them. ## 5. Deploy only approved targets and read them back - Refresh the exact live targets and back them up immediately before approved writes. Earlier backups may predate another writer's changes. - Use the generator's narrow operation when only a script is approved; prove it in isolation before live use. Do not invoke broader role reconciliation for convenience. - End-to-end command recipe for a generator that materializes into profiles (gate → lease → apply → sandbox → commit → materialize → read-back): `references/landing-and-deploy-recipe.md`. - For an approved prompt-only job update, save that target's original prompt, use the supported scheduler interface, and preserve every other job field. Read back the exact job and compare non-prompt fields to the pre-write record. - Compare the deployed wrapper to fresh generator output, execute that actual wrapper under the reported environment, and re-run drift checks against a fresh predeployment baseline. Separate intentional undeployed drift from unrelated findings; never clear excluded session/config state to make a checker silent. - Stop when the approved acceptance proof is complete. Report source landed, staged payload verified, live deployed, or awaiting authorization explicitly; a clean Git tree and passing tests do not prove deployment. ## 6. Close the handoff without chatter - If only an operator's decision remains, stop peer acknowledgement ping-pong. A corrected authority claim is not a new authorization. - When the operator explicitly asks for status or continuation, provide a substantive checkpoint: landed artifact and proof, remaining gated actions, and the single required decision. Do not answer a direct request with silence or call source-only work "done".