docs: cut restated upstream behavior left by the stamp sweep (batch 7) #15515

Problem

The stamp sweep in #15476 (docs: drop stamps from pointer records (batch 7)) found these passages restating upstream behavior. Under contract 6.0.0 (#15468) restated behavior is not stamped; it is cut to a pointer or rewritten as this repository's own rule. The sweep left them unchanged because cutting them changes what a skill tells the agent.

Items

• plugins/actionlint/skills/setup/SKILL.md:179: says an undeclared userConfig key silently cannot be set natively
• plugins/ai-briefing/skills/generate/SKILL.md:115: says the Bash tool env lacks the plugin's CLAUDE_PLUGIN_DATA and another plugin's hook can export one
• plugins/app-control/skills/drive/reference/pid-binding.md:19: paraphrases proc(5) on who may write comm and the cmdline NUL layout
• plugins/autonomy/reference/routines.md:350: says repository settings load no plugins into a cloud session
• plugins/autonomy/reference/routines.md:477: record paraphrasing the skills page on cloud-session skill and plugin loading
• plugins/autonomy/skills/setup/SKILL.md:200: says a watched repo's settings cannot supply plugin options
• plugins/autonomy/skills/setup/templates/trigger-adapters.md:128: restates which repository role may apply labels
• plugins/context-budget/README.md:128: restates that settings hooks run after the mod chain and a mod can skip or override them
• plugins/context-budget/README.md:143: restates that ConfigChange fires after any write and cannot ask
• plugins/context-budget/skills/audit/SKILL.md:377: restates that auto mode's classifier, not a person, decides protected-path writes
• plugins/context-budget/skills/setup/SKILL.md:62: restates that a hook that cannot start is a non-blocking error
• plugins/context-budget/skills/setup/SKILL.md:87: restates that a changed option shows only in a fresh session
• plugins/context-guard/README.md:403: restates that Claude Code reloads a mod when its options change
• plugins/context-guard/skills/setup/SKILL.md:48: restates that installing Claude Code puts no node on PATH and how a hook that cannot start is treated
• plugins/context-guard/skills/setup/SKILL.md:238: restates what an uninstall removes
• plugins/context-guard/reference/cloud-headless-capture.md:12: restates what $.session.usage() reports and where mods run
• plugins/context-guard/reference/cloud-headless-capture.md:25: restates that plugins are unavailable in Desktop WSL sessions
• plugins/context-guard/reference/reader-contract.md:206: restates that used_percentage counts the input side only
• plugins/context-guard/reference/reader-contract.md:447: restates when the statusline percentage is computed and what turning auto-compact off does
• plugins/instruction-placement/README.md:174: restates that Claude loads the whole ancestor chain
• plugins/instruction-placement/skills/check/SKILL.md:33: restates that bad globs and over-budget patterns fail silently
• plugins/instruction-placement/skills/check/SKILL.md:192: restates that a symlinked rule may point outside the repository
• plugins/instruction-placement/skills/migrate/SKILL.md:457: restates that the Bash tool env has no CLAUDE_PLUGIN_ROOT
• plugins/instruction-placement/skills/realign/SKILL.md:228: same restatement as migrate SKILL.md:457
• plugins/rate-limit-guard/README.md:~302: "Claude Code reloads a module when its options change" paraphrases the Register doc comment in index.d.ts.
• plugins/plugin-quality/skills/audit/reference/component-types/config.md:46: "Every stdout line from command reaches Claude as a notification" restates the monitors section.
• plugins/plugin-quality/skills/audit/reference/evidence-packet.md:30: restates how the upstream data-dir id is formed.
• plugins/plugin-quality/skills/audit/reference/evidence-packet.md:155-157: restates the sub-agents "Available tools" section.
• plugins/skill-quality/scripts/check-skill.sh:981: restates the frontmatter-reference behavior for unparsable frontmatter (the error string at :997 repeats it).
• plugins/skill-quality/scripts/check-skill.sh:927: restates the picker alias behavior.
• plugins/skill-quality/skills/setup/SKILL.md:66: "a same-session rerun still resolves the OLD root" restates "Reference a saved value".
• plugins/skill-quality/README.md:67: "under 500 lines" restates the upstream number (the stored copy is check-skill.sh LINE_HARD_CAP).
• plugins/multi-agent/reference/config.md:60-66: paraphrases the sub-agents "choose a model" section (aliases and resolution order). Line numbers are from before the edits.
• plugins/multi-agent/reference/config.md:99-107: "Where effort applies" paraphrases the upstream effort precedence and inheritance rules.
• plugins/multi-agent/reference/config.md:118-123: "Hard cap" paraphrases how maxEffortLevel behaves.
• plugins/mcp-tools/skills/audit/reference/checklist.md:40: "Source conflict" bullet restates what the mcp/env-vars pages and changelog 2.1.295/2.1.296 say about length limits.
• plugins/mcp-tools/skills/audit-posture/reference/checklist.md:215: "Source conflict" bullet restates the mcp page against changelog 2.1.274 on sdk-entry skips.
• plugins/code-metrics/skills/principles/reference/measures.md:168: shellmetrics values depend on the parser shell (repeats the README).
• plugins/visualization/skills/visualize/context/decision-matrix.md:303: public-sharing failures for mermaid, SVG data: and AVIF pages (repeats claude-code#79824 and #81410).
• plugins/visualization/skills/visualize/context/decision-matrix.md:270: terminal Markdown rendering (repeats the interactive-mode page).
• plugins/visualization/skills/visualize/context/decision-matrix.md:281: "no public docs section covers them" note.
• plugins/discipline/skills/sweep-all/reference/batched-pass.md:61-66: what CLAUDE_CODE_MAX_TOOL_USE_CONCURRENCY caps, and the subagent limit.
• plugins/discipline/skills/use-your-skills/SKILL.md:27-28: a manual-only skill never reaches the listing.
• plugins/discipline/skills/use-your-skills/SKILL.md:87-88: descriptions are dropped on overflow.
• plugins/repo-hygiene/skills/clean/reference/invocation-forms.md:13-14: ${CLAUDE_SKILL_DIR} substitution.
• plugins/repo-hygiene/skills/setup/SKILL.md:99-100: which scopes plugin option values are read from (it had no stamp).
• plugins/writing/skills/be-concise/reference/sources.md:76: Inverted Pyramid (repeats the NN/g article).
• plugins/writing/skills/be-concise/reference/sources.md:84: F-shaped pattern (repeats the NN/g article).
• plugins/writing/skills/be-concise/reference/sources.md:122: licensing (repeats the NN/g terms page).
• plugins/writing/skills/be-concise/reference/sources.md:175: lists the Google style guide's rules.
• plugins/writing/skills/be-concise/reference/sources.md:221: paraphrases ACUS 2017-3.
• plugins/writing/skills/be-concise/reference/sources.md:230: paraphrases DOE-STD-1029.
• plugins/writing/skills/be-concise/reference/sources.md:242: paraphrases the kernel's "Describe your changes".
• plugins/writing/skills/be-concise/reference/doctrine.md:181: the Concise output style's content floor (repeats the output-styles page).
• plugins/developer-experience/skills/build-cli/SKILL.md:44: glosses of the sysexits(3) codes.
• plugins/developer-experience/skills/audit-tools/SKILL.md:42: the legacy .claude/commands status.
• plugins/fleet/skills/reach/SKILL.md:84-87: paraphrases the Login-expired entry.
• plugins/fleet/skills/reach/SKILL.md:131-133: paraphrases the workspace-not-trusted error.
• plugins/fleet/skills/reach/SKILL.md:229-232: -p reads stdin as the prompt.
• plugins/fleet/skills/reach/reference/relay.md:88-90: paraphrases Control inbound messages.
• plugins/fleet/skills/reach/reference/relay.md:157-158: restates the workspace-not-trusted error.
• plugins/fleet/skills/reach/reference/relay.md:166-167: low confidence; this may be a pointer only.
• plugins/computer-use/skills/diagnose/reference/failure-diagnostics.md:49: paraphrases the Cowork support article (machine must be awake)
• plugins/context7/skills/lookup/context/mcp.md:31: paraphrases MCP unset-variable behavior
• plugins/dometrain/skills/grounding/SKILL.md:55: describes the server's result shape
• plugins/dometrain/skills/update/SKILL.md:70: paraphrases the variable-resolution table
• plugins/playwright/skills/playwright/SKILL.md:137: same variable-resolution paraphrase
• plugins/improvement/skills/find/SKILL.md:272: same variable-resolution paraphrase (Read tool bytes, no CLAUDE_PLUGIN_ROOT in Bash)
• plugins/miro/README.md:14: restates the stderr and debug-log route
• plugins/mutation-testing/skills/principles/reference/tooling.md:46: restates what inPlace: true writes
• plugins/repo-fleet-hygiene/skills/audit/reference/official-sources.md:3: the bullets paraphrase each linked page
• plugins/speech/reference/elevenlabs.md:9: restates the models page on eleven_v4
• plugins/go-format/README.md:17 and plugins/ruff-format/README.md:18: "neither applies to PostToolBatch" restates the hooks page
• plugins/improvement/skills/find/context/ci-health.md:40,49,104: restate the /timing retirement, the per-search cap, and that actions_list has no created filter
• plugins/pixel-art/reference/backends.md:53: restates the Aseprite license shape (EULA replaced GPLv2 in August 2016)
• plugins/wizard/skills/unattended/SKILL.md:237: restates SecretManagement Get-Secret search and -AsPlainText behavior
• plugins/wizard/skills/unattended/SKILL.md:252: restates Keychain and pass/gpg-agent prompting
• plugins/tdd/skills/principles/reference/characterization-testing.md:79: the Sources list summarizes the upstream acceptance workflows (read date removed)
• plugins/core/README.md:46: restates the dependency behavior on install, disable and prune. The stamp was kept because the record quotes the error string Dependency "<name>" is disabled

Kept, unsure (53)

• Mixed As-of tables, decided per row (D26): pointer rows now leave the As of and Recheck cells empty, and value rows keep their date. 18 rows changed, in context-budget native-context/native-skill-doctor, context-guard reader-contract (the record bullets under its tunable-surfaces table), bugs native-bug, prototype bundled-design, github native-install-github-app, firecrawl native-web-tools and verification native-verify. These rows stay dated because they mix a probe with a pointer:
  • plugins/context-budget/skills/audit/reference/native-context.md:12 (rests on the probed description)
  • plugins/verification/skills/confirm/reference/native-verify.md:9, :10 (an extraction plus a pointer)
  • plugins/debugging/skills/debug/reference/native-debug.md:9, :11 (a quoted description; gated: false)
  • plugins/repo-fleet-hygiene/skills/audit/SKILL.md:197 (a fleet observation plus a pointer)
• lib/session-bridge/session_bridge.py:942: pointer list, but the code depends on the flag and policy-key strings it names; kept with README.md:235-257, which records those strings.
• lib/managed-scope.sh:29, :82: the cache path and drop-in merge key names are strings the code compares against; kept as stored values.
• lib/docs-fetcher-gate.mjs:116: comment beside the UID constant (Node's Windows chmod behavior); kept under the constant-comment rule.
• scripts/plugin-validate-report.mjs:18: pointer record, but gatingHooks and hasCatch are upstream field names the script reads.
• plugins/corrections/skills/codify/SKILL.md:129: decision with an issue trigger (fix(source-control): apply a config change on main to pr-merge's policy read for PRs opened before it #8834); dropping the date makes check-skill.sh flag the #N as a history citation, so the stamp stays until that check accepts a dateless trigger.
• plugins/actionlint/skills/setup/SKILL.md:190: "no matching upstream issue as of 2026-07-23" is an absence note (rule says drop), but it sits beside a local repro on actionlint 1.7.x and its trigger re-attributes the -shellcheck= decision
• plugins/ai-slop/skills/audit/reference/catalog.md:1229: model-era record; dated record of measured phrase lists, but its trigger is also the promotion decision's condition (could be drop-date, keep-trigger)
• plugins/autonomy/reference/routines.md:389: "Reviewed 2026-09-06" survey note; the trigger is about a re-survey, and the date is the note's subject
• plugins/autonomy/reference/trigger-dispatch.md:28: research gaps "carried open as of 2026-09-06"; trigger fires when "this review date passes six months", so the date is load-bearing
• plugins/autonomy/reference/runner/escalation.md:262: same shape as trigger-dispatch.md:28
• plugins/autonomy/reference/runner/escalation.md:254: absence record verified against release v1.44.0 (absence note vs version beside it)
• plugins/autonomy/reference/telemetry.md:137: GenAI conventions absence record (no release/tag); trigger is our migration decision's condition
• plugins/autonomy/skills/setup/context/windows-surfaces.md:24,37,51,74,83,123,147,177,199: Claim/Basis records that restate upstream behavior but each carries stored defaults, versions or command strings; kept as stored values, whole file is a conversion candidate
• plugins/autonomy/README.md:96: provenance paragraph keeps "fetched " notes (not counted by the script) and its undated trigger at :107
• plugins/context-guard/scripts/cache-line.mjs:14: header comment basis for the field names the script reads; treated as a comment beside constants
• plugins/context-guard/reference/cloud-headless-capture.md:112: channels table of dated observations holding field names
• plugins/instruction-placement/context/verified-mechanics.md:58: changelog record naming versions 2.1.288/2.1.290; kept as a version gate
• plugins/instruction-placement/skills/migrate/reference/sources.md:376: "as of 2026-09-21" names a past revision of another repository file
• plugins/{bash-format,biome-format,guardrails,rate-limit-guard}/skills/setup/SKILL.md (bash-format:34, biome-format:28, guardrails:32, rate-limit-guard:24): the placeholder pointer record. I kept it as a stored string because the line above it quotes the upstream placeholder [shell command execution disabled by policy] word for word. About 18 setup skills carry the same record, so the batch should decide it the same way everywhere.
• plugins/attribution/skills/audit/reference/rubric.md:404-423: three records that deliberately restate our own upstream-drift contract and org standard for an offline judge ("Each is pinned so it can be re-derived"). If the stamps come out, the paragraph's own premise no longer holds, and cutting the prose would break the offline design.
• plugins/attribution/skills/audit/reference/source-fetch.md:37: a record "restated deliberately and carried as a stamped record so the restatement stays honest". Same reasoning as rubric.md:404-423.
• plugins/bash-format/README.md:42: "checked 2026-10-04 against shfmt v3.14.1" is an observed upstream bug state tied to a version (Associative array key with hyphen breaks in v3.6.0 mvdan/sh#956).
• plugins/biome-format/skills/setup/SKILL.md:170: the nodeLinker pointer. The text lists the upstream value names (pnp, node-modules, pnpm), so I treated them as stored strings.
• plugins/guardrails/README.md:~311 and plugins/guardrails/hooks/run-guards.sh:~147: the timed-out-hook pointer, kept as a version gate ("on Claude Code 2.1.295 or later").
• plugins/guardrails/README.md:~334: restates the mcp__plugin_<plugin>_github__<tool> name form, a string copied from upstream.
• plugins/guardrails/README.md:~461: restates the path ~/.claude/plugins/data/<id>/, a string copied from upstream.
• plugins/guardrails/hooks/refuse.sh:25: the trigger also covers the probed CLAUDE_CODE_SESSION_ATTENDED value (2.1.301) and the permissionDecision/ask strings the script writes.
• plugins/guardrails/hooks/workflow-resilience-check.sh:118: the comment explains the FINDING constant, which names an env var, and holds an open question (feat(guardrails): evaluate CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS in workflow-resilience-check #6756) with its own trigger.
• plugins/guardrails/hooks/block-root-delete-target.sh:82,279,361,401,1587,3665: notes from reading the source of other tools (find, pwsh, watch, xargs, parallel, env/uutils 0.10.0, shadow sg). They record option sets the guard's model depends on.
• plugins/plugin-quality/skills/audit/reference/component-types/hook.md:67-71: a Claim/Basis observation "read by loading it in Claude Code 2.1.288". It has a version but no stored number.
• plugins/skill-quality/reference/invocation-probes.md:114-118: a --disallow-tools help-line observation (2.1.316).
• plugins/skill-quality/scripts/check-skill.sh:2310-2313: a pointer comment recording the default false for an absent disable-model-invocation, kept as a stored default.
• plugins/multi-agent/reference/config.md:90: the mechanical model-default record. I kept the full stamp as a model default; it could also be read as a recorded decision with a named trigger.
• plugins/implementation/skills/report/SKILL.md:89-90,123-124 and plugins/plugin-quality/skills/report/SKILL.md:89-90,123-124: not edited. These must stay byte-identical to the canonical plugins/discovery/skills/report/SKILL.md (scripts/validate-plugin-contracts.mjs:1162-1186; validate-plugin-contracts.test.sh:1219). Remove the stamps from the canonical file and then from every copy together.
• plugins/visualization/skills/visualize/context/decision-matrix.md:33: the mermaid runtime probe record. Its version is in the linked PR comment, not on this line.
• plugins/mcp-tools/skills/audit/reference/server-discovery.md:39: the only stamp on the stored SDK spellings.
• plugins/mcp-tools/skills/audit/reference/server-discovery.md:69: the only stamp on the stored _meta argument and attribute names.
• plugins/code-metrics/skills/audit-suppressions/SKILL.md:163: the only stamp on the marker catalog's directive strings.
• plugins/discipline/skills/sweep-all/reference/inheritance-preflight.md:3-10: a file-wide "verified 2026-11-04, Claude Code 2.1.318" record. It covers an error string, and batched-pass.md:30-32 cites it as the file's opening stamp.
• plugins/repo-hygiene/skills/clean/scripts/destructive-guard.sh:63,75: Claim: as of 2026-09-29 comments about our own code, not about an upstream page.
• plugins/fleet/skills/reach/SKILL.md:243: the PowerShell-to-wsl.exe quoting probe, which may be a measured string.
• plugins/user-experience/agents/evaluator.md:6: a frontmatter comment pointer next to the effort: high constant.
• plugins/core/README.md:60: a mixed record (restated behavior plus one quoted error string), kept as a stored value
• plugins/improvement/skills/find/context/ci-health.md:79: the Metrics record whose stored content is the API field names (run_attempt, run_started_at, created_at) that the commands read
• plugins/markdown-format/skills/setup/SKILL.md:89: a probe record of markdownlint-cli2 --help output (banner first, no --version), observed tool behavior
• plugins/pixel-art/reference/backends.md:59,68: records whose stored content is the Aseprite CLI flag and API names
• plugins/playwright/skills/demo-video/reference/quality-control.md:45: the record behind the stored reading-time values
• plugins/playwright/skills/playwright/reference/tracing-and-video.md:89: release-version pointers, read as version gates
• plugins/debugging/skills/debug/reference/ecosystem-debugging.md:17,22 and plugins/toolchain/skills/check/context/go.md:17: records whose stored content is flag or API names
• plugins/mutation-testing/skills/audit/context/tool-test-restriction.md:29,49,63: stored-value table and its source record
• plugins/mutation-testing/skills/principles/reference/metrics.md:10, operators-and-states.md:9, scaling-and-suppression.md:10, tooling.md:12: "Fetched … Recheck" source headers over files of copied formulas and figures
• plugins/repo-fleet-hygiene/skills/audit/SKILL.md:147: a pointer sentence that holds a measured figure

Kept, no anchor checked (36)

• plugins/app-control/reference/control-skill.md:174: memray site root and py-spy repo root, no section anchor
• plugins/app-control/reference/control-skill.md:187: pprof repo root
• plugins/app-control/reference/control-skill.md:193: perf-record man page (has sections, no anchor)
• plugins/app-control/reference/control-skill.md:194: heaptrack repo root
• plugins/autonomy/skills/setup/context/guardrail-slice.md:131: tools-reference table (text states it has no anchor)
• plugins/autonomy/skills/setup/templates/trigger-adapters.md:136: fine-grained PAT permissions page, "Metadata" section named in prose
• plugins/autonomy/skills/setup/templates/trigger-adapters.md:141: GitHub App on-behalf-of-a-user page
• plugins/context-budget/README.md:77: env-vars page for SLASH_COMMAND_TOOL_CHAR_BUDGET (stored-value record, kept stamp)
• plugins/context-guard/reference/reader-contract.md:586: mods reference page root for what a mod receives
• plugins/guardrails/hooks/block-credential-read.sh:39: links https://code.claude.com/docs/en/permissions with no anchor. The likely anchor is #what-read-and-edit-rules-cover, which guardrails setup SKILL.md already uses.
• plugins/implementation/README.md:133-134: the sub-agents and permissions page links have no anchors.
• plugins/implementation/skills/implement-dispatch/SKILL.md:19: https://git-scm.com/docs/git-worktree.
• plugins/implementation/skills/implement/context/replace.md:70-74: the Scientist README, Jest snapshot testing and Istio traffic mirroring links.
• plugins/implementation/skills/implement-dispatch/reference/holdout-tests.md:98-99: the Vitest CLI and Jest CLI pages.
• plugins/multi-agent/scripts/allowed-tools-pairing.test.sh:7: the permissions page, no anchor.
• plugins/multi-agent/scripts/allowed-tools-pairing.test.sh:17: the skills page, no anchor.
• plugins/mcp-tools/skills/audit-posture/reference/checklist.md: the Registry moderation, Docker sandboxes and Sandbox coverage (page intro) pointers.
• plugins/code-metrics/skills/audit-suppressions/SKILL.md:141-162: several entries in the Upstream syntax list (flake8, mypy, pylint, ShellCheck, Stryker, Semgrep, RuboCop and others).
• plugins/developer-experience/skills/build-cli/SKILL.md:46: the sysexits(3) man page.
• plugins/developer-experience/skills/build-cli/reference/cli-contract.md:22: clig.dev "Arguments and flags", cited by name.
• plugins/developer-experience/skills/build-cli/reference/cli-contract.md:23,26: the code.claude.com best-practices page.
• plugins/developer-experience/skills/build-cli/reference/cli-contract.md:24,29,61: the writing-tools-for-agents article, cited by section name.
• plugins/writing/skills/be-concise/reference/sources.md:182,247: the Google style highlights page and the kernel submitting-patches page.
• plugins/user-experience/reference/ai-features.md:36-39: the Apple HIG Generative AI page and the PAIR guidebook.
• plugins/user-experience/reference/first-run.md:54-66: the Apple HIG Privacy page, Android "Request runtime permissions" and the Apple HIG Onboarding page.
• plugins/user-experience/reference/flows-ia.md:105-107: the Next.js project-structure page.
• plugins/user-experience/reference/flows-ia.md:212-215 and measurement.md:81-83: GOV.UK Service Manual pages.
• plugins/user-experience/reference/deliverable.md:25-27: the ICC/ESOMAR Code PDF.
• plugins/user-experience/reference/research-methods.md:111-113: the W3C WAI involving-users page.
• plugins/github/reference/browser-automation.md:42: https://code.claude.com/docs/en/chrome platform note; the record says it sits above the first heading with no anchor
• plugins/explainer-video/skills/produce/SKILL.md:121: EBU R 128 and R 128 s2 are PDFs cited by section number with no anchor
• plugins/explainer-video/skills/produce/SKILL.md:133: the Netflix timed-text article has no section anchor
• plugins/playwright/skills/demo-video/reference/posting.md:73, quality-control.md:42, playwright/reference/storage-and-auth.md:156, test-generation.md:108: see b7-helper-D-1.md
• plugins/context7/skills/lookup/context/mcp.md:6: context7.com/docs (stored-value record)
• plugins/debugging/skills/analyze-profile/reference/formats/chrome-trace.md:10, heapsnapshot.md:7: docs roots
• plugins/speech/reference/elevenlabs.md:13,24: models and best-practices pages

Restored value stamps

Per D26, a record that holds a version, count, byte size, hash, timing or exit code read from upstream keeps its stamp. The third commit restores 33 stamps (in 23 files) that the sweep had removed: the visualization decision-matrix Mermaid probe counts, plus version-bearing records in autonomy, context-budget, context-guard, explainer-video, firecrawl, fleet, instruction-placement, mcp-tools, playwright, rate-limit-guard, skill-quality and toolchain. The other 61 regex candidates turned out to be behavior-only or noise (IPs, exhibit and arXiv ids, our own constants). These were left removed:

• plugins/developer-experience/skills/build-cli/reference/cli-contract.md: the Count column is our tally of upstream sources, not an upstream value.
• plugins/playwright/README.md:96, plugins/guardrails/README.md:1568, plugins/visualization/skills/visualize/SKILL.md:409, decision-matrix.md:212: each record holds a value, but the removed stamp covered a separate pointer sentence.

Withheld

• skills/report/SKILL.md in implementation and plugin-quality must stay byte-identical to plugins/discovery/skills/report/SKILL.md (scripts/validate-plugin-contracts.mjs). Discovery belongs to another batch, so these copies keep their stamps here.

Do

Per item: cut to the anchored pointer, keep as our own rule with its reason, or record why it stays. A cut that changes agent behavior needs an eval or a stated basis.

Related

Refs: #15476, #15468