Skip to content

feat(plugin-profiles)!: switch plugin sets by profile and remove the core bundle - #16483

Merged
kyle-sexton merged 25 commits into
mainfrom
feat/plugin-profiles
Nov 16, 2026
Merged

kyle-sexton merged 25 commits into
mainfrom
feat/plugin-profiles

Conversation

@kyle-sexton

@kyle-sexton kyle-sexton commented Nov 16, 2026 •

Copy link
Copy Markdown
Member

Closes #16429
Closes #16432
Closes #16435
Closes #16436
Refs: #16428
Refs: #16433
Refs: #16434

Summary

Adds the plugin-profiles plugin, which turns whole sets of plugins on or off for one session, for a folder, or as the initial profile everywhere, so a session lists only the skills its workflow phase needs. It ships seven profiles and removes the core bundle plugin in the same release. This is container #16428's single PR: its six slices landed on the integration branch feat/plugin-profiles, so the #16433 and #16434 work rides here on purpose (owner decision: one pull request for the container's slices).

Fix

  • Engine and CLI (Apply a plugin profile to a folder from the command line #16429): plugins/plugin-profiles/scripts/profiles.mjs (apply <name> --for session|folder|initial [--dry-run] [--yes], status, clear [--dry-run] [--yes], check, inventory, list, usage, save, install) over one-concern modules in lib/: a pure planner (plan.mjs), profile files on three levels (profile-files.mjs), the plugin inventory (inventory.mjs), settings reads and all-or-nothing writes (settings-io.mjs, atomic-file.mjs), the applied-profile record (record.mjs), the prerequisite gate (prerequisite-gate.mjs). It writes only enabledPlugins and skillOverrides entries (plus the one-session file's env.PLUGIN_PROFILES_SESSION), never a repository's committed settings, and shows every change before it writes, clear included. apply always-on applies the always-on profile alone. In a cached install, check reads the catalog of the marketplace the plugin was installed from (plugins/known_marketplaces.json).
  • Profiles and check (Ship the seven plugin profiles with a CI profile check #16435): seven files in profiles/ (always-on, discovery, planning, interface-design, build, review, harness-upkeep), the only home of each profile's members. scripts/check-plugin-profiles.sh fails CI on a shipped id missing from the catalog, a wrong marketplace, or an interface-design profile missing a planning member; wired into the Check Catalog job of .github/workflows/pr-require-checks.yml.
  • One-session start and session-start notice (Start one session in a profile and report a folder's profile at session start #16432): --for session writes a file for claude --settings; a SessionStart hook (node, exec form, about 30 ms) names the folder's profile, entries that differ, plugins installed since, an unfinished switch and the surface, without a nested claude. userConfig.remember_folder_profile turns folder remembering off.
  • /plugin-profile mod (Switch plugin profiles in a running session with /plugin-profile #16433): picker, preview, confirm; an in-session switch writes the on entries, reloads plugins once at the end of the turn, then writes the off entries; refuses when remembering is off or the session runs a one-session profile; asks before enabling a plugin with an MCP or language server. Only the mod uses the mods API; the CLI and skill apply without it.
  • Guided skill (Guide a user to their plugin profiles and install only on yes #16434): /plugin-profiles:profile with apply, status, clear, check and setup; setup inventories plugins and skills, reads local use counts, asks about the user's work and about each hook-shipping plugin, and installs a missing plugin only on an explicit yes from a known marketplace, then switches it off at user level unless the initial profile includes it.
  • core removed (Remove the core bundle plugin from the catalog #16436): plugin, catalog entry, validator block and its tests, fragment-plugins row, fixture and doc mentions.

Retired plugin

core is retired (docs/migration-playbook.md "Retiring a published plugin"). Its users drop the core@melodic-software entry from enabledPlugins; the always-on profile (apply always-on --for initial) replaces it.

Verification

  • The seven Node suites (lib/{check,inventory,plan,profile-files,surface}.test.mjs, scripts/{profiles,install-real-claude}.test.mjs): 110 pass, 0 fail. CI runs each through its sibling .test.sh in the Test Bash job (test-bash in .github/workflows/pr-require-checks.yml); bash scripts/plan-test-lanes.sh --base origin/main exits 0 and plans all seven.
  • Install with the real claude (scripts/install-real-claude.test.mjs, Guide a user to their plugin profiles and install only on yes #16434 AC4 and AC20): against the fixture marketplace in a throwaway HOME and config, install draw@fixture-market --dry-run installs nothing, no flag exits 1 and installs nothing, --yes installs at user scope and leaves enabledPlugins["draw@fixture-market"] false. It ran, not skipped, on the build host with Claude Code 2.1.325 (1 pass, about 3 s); it skips where claude is not on PATH, so a green Test Bash job on a runner without claude does not show it ran.
  • Guided setup (Guide a user to their plugin profiles and install only on yes #16434), three headless claude -p sessions on this head in a throwaway HOME and config with the plugin loaded from the branch, one per setup case, graded by the cases' own graders, all passing with no permission denial:
    • No install without a yes: the session ran inventory --json and install draw@fixture-market --dry-run despite being told to skip the other questions, asked "Install draw@fixture-market?", and installed and saved nothing.
    • Install only after a yes: preview, then one install draw@fixture-market --yes; the plugin ended installed at user scope with its enabledPlugins entry false, and the profile was saved as a user file by the tool.
    • Inventory, hook-shipping plugins and a saved profile: inventory and usage ran, the reply named which plugin ships hooks and kept it per profile from the user's stated answer (the prompt gives that answer up front, so the question was not asked live), and the profile was previewed, then saved as a user file by the tool.
    • The status eval case ran valid in claude plugin eval and passed.
  • Cloud and Desktop WSL sessions: status, apply, clear and install say profiles do not apply and write nothing, and status --session-start prints nothing (tests in scripts/profiles.test.mjs and lib/surface.test.mjs). apply --for initial writes user settings and clear --for initial restores them; status names the session's surface.
  • End to end from a copy laid out like a cached install (temp config, HOME and project, stub claude): apply build --for folder --dry-run, apply build --for folder --yes, status, clear --dry-run, clear --yes and check each exit 0; the dry runs write nothing, and clear --yes removes the file the apply created.
  • bash scripts/test-plugin-mods.sh: all mod tests pass, plugin-profiles included, no skip line.
  • bash scripts/check-plugin-profiles.sh: exit 0 on the tree; its wrapper test exits 1 on a fixture naming nope@melodic-software (3/3).
  • Phase 2 checks in temp HOME and project with a stub claude: only enabledPlugins/skillOverrides change (jq -S 'del(...)' identical); clear restores the file byte for byte (cmp 0) and removes a file the tool created; a missing required prerequisite makes apply --yes exit 1 with the file unchanged; status --json 0.03 s.
  • scripts/sync-shared-copies.sh --check, scripts/generate-catalog.test.sh (16/16), validate-plugin-contracts and its tests (230/230), admit.test.mjs (710/710), the hook validators, /skill-quality:check (0 FAIL), and pre-push check-affected: pass.
  • jq -e '[.plugins[].name] | index("core") == null' .claude-plugin/marketplace.json: true.

Remaining

  • Switch plugin profiles in a running session with /plugin-profile #16433: the live demo of an in-session switch (build then /plugin-profile review, one reload at the end of the turn, no /reload-skills) needs an interactive session on the installed release; covered here only by claude plugin test cases.
  • Guide a user to their plugin profiles and install only on yes #16434's three setup eval cases cannot run in claude plugin eval today: its sandbox write-denies the run's config directory and the nested claude that inventory and install call is not reachable, so their first runs were invalid. They stay in the repository, validated (0 FAIL), for when the harness allows nesting; the headless sessions under Verification prove the behavior instead (owner decision, 2026-11-15).
  • Guide a user to their plugin profiles and install only on yes #16434 stays open; this pull request delivers its skill, cases and install proof, and a small follow-up pull request closes it with the two things it still owes:
    1. The per-hook case's prompt without the up-front answer, and graders/per-hook-question.md requiring the question to be asked live.
    2. One headless session showing that question asked.
  • Unproven by a live session: the skill body now says a "Not installed" line does not stop an apply, only a missing required prerequisite stops --yes. In one headless session, before that line, the skill told the user it could not run an apply with --yes because always-on members were not installed. No session has run since the line was added.
  • Unverified: whether a --settings file's env value reaches SessionStart hooks; the CLAUDE_CODE_ENTRYPOINT values the surface detection reads are undocumented (read from the 2.1.325 binary).

Owner decisions

Related

Follow-ups

  • Pre-existing (advisory): docs/migration-playbook.md "Retiring a published plugin" does not say to drop the plugin from scripts/fragment-plugins.txt, without which the changelog-fragment check fails.

🤖 Generated with Claude Code

https://claude.ai/code/session_015XWwmd6RuLMb2eXqV5uoEJ

kyle-sexton and others added 12 commits November 15, 2026 19:47
…mmand line

Adds the plugin-profiles plugin (on by default) with its engine and the
always-on and build profiles. scripts/profiles.mjs applies a profile to
a folder's local settings, user settings or a one-session --settings
file after a preview and a confirm, checks the required prerequisites of
each plugin it switches on, writes only enabledPlugins and
skillOverrides through a hash-checked temp-file rename, records the
values it replaced, and clear restores them.

Refs #16429

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmd6RuLMb2eXqV5uoEJ
…ep repository profiles out of user settings

Declares the claude CLI in prerequisites.json, registers the
lib/plugin-options.mjs cluster now carried at the same path by two
plugins, separates version from description in plugin.json, and reads
no repository profile when applying the initial profile, since that
writes user settings.

Refs #16429

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmd6RuLMb2eXqV5uoEJ
Profiles from plugin-profiles replace the fixed core bundle, so core
leaves the catalog in the same release that brings them.

- Delete plugins/core/ and its marketplace.json entry; regenerate
  docs/catalog.md.
- Drop the validator's core bundle check and its test cases.
- Drop core from scripts/fragment-plugins.txt: a removed plugin has no
  releases left, and the fragment gate otherwise asks for a core
  fragment.
- Point the intake-triage install fixture at plugin-profiles.
- Correct the dependencies row of docs/plugin-philosophy.md.
- Note the removal in the plugin-profiles changelog fragment.

BREAKING CHANGE: core@melodic-software is no longer published. Installed
copies keep working until uninstalled; drop its enabledPlugins entry
and apply a profile instead.

Refs: #16436

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmd6RuLMb2eXqV5uoEJ
Add discovery, planning, interface-design, review and harness-upkeep
beside always-on and build, each with a description line, members from
the hub lifecycle scan.

The check action reads every profile level through the one profile
parser: it fails on a shipped id missing from the catalog, an id under
another marketplace (naming the fix), a missing description, an exclude
naming an always-on member, and interface-design lacking a planning
member; it reports user and repository ids that are not installed and
installed plugins no profile names. scripts/check-plugin-profiles.sh
runs it in the Check Catalog job.

Refs #16435

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmd6RuLMb2eXqV5uoEJ
…mbering

A SessionStart command hook (exec form, node) runs status --session-start:
it names the folder's profile, entries that differ, plugins installed since
the apply, an unfinished switch and the surface, reads installed plugins
from disk without a nested claude, and prints nothing in a folder with no
applied profile.

lib/surface.mjs detects the surface from CLAUDE_CODE_REMOTE and
CLAUDE_CODE_ENTRYPOINT: VS Code and Desktop refuse a one-session apply with
the folder offer, cloud sessions report that profiles do not apply and write
nothing, and an undetected surface allows the folder level only. A
one-session apply and status say to pass --settings again on resume.

The remember_folder_profile option (default true): when false a folder apply
is refused and status reports the initial profile, naming a folder record as
not in force.

Refs #16432

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmd6RuLMb2eXqV5uoEJ
…gin-profile

The mod answers /plugin-profile with a picker, the apply --dry-run preview
and a confirm, all without a model turn. It writes the on entries, runs
/reload-plugins once at the end of the next turn, then writes the off
entries, completing the off-pending record; a switch enabling an MCP or
language server plugin names it before asking. It runs the CLI as a process
and imports no lib/ code. The CLI gains list [--json], apply --dry-run
--json, and refuses --defer-off in a one-session profile.

Refs: #16433

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmd6RuLMb2eXqV5uoEJ
…nonical tsconfig

Moves the /plugin-profile entry into the branch's one changelog fragment,
copies the registered tsconfig.json byte for byte, and fixes a spelling the
typos gate refuses.

Refs: #16433

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmd6RuLMb2eXqV5uoEJ
… only on yes

Add /plugin-profiles:profile with apply, status, clear, check and a guided
setup that inventories plugins and skills, reads local usage history, asks
about the user's work and each hook-shipping plugin, and proposes profiles
to save. Every write is previewed with --dry-run and runs with --yes only
after the user's yes.

The CLI gains usage (Claude Code's own skill and plugin counts, or no
history), save (a user or repository profile file, printed first) and
install (user scope, known marketplaces only, never for a plugin only a
repository profile names, then off at user level unless the initial
profile includes it); inventory lists each plugin's skills. Evals for
setup and status are added, validated, not run.

Refs: #16434

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmd6RuLMb2eXqV5uoEJ
@kyle-sexton
kyle-sexton marked this pull request as ready for review November 16, 2026 01:57
@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Codex Review Summary: ✅ Completed

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-11-16T02:01:18.900905Z 574b7c8 Draft marked ready
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 574b7c86c5

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

kyle-sexton and others added 2 commits November 15, 2026 21:02
Give each Node suite a sibling .test.sh wrapper, the convention
scripts/plan-test-lanes.sh routes to test-bash, and drop the single
tests/profiles.test.sh that ran them all, so no suite runs twice.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
@github-actions

Copy link
Copy Markdown
Contributor

Lane run report

Field Value
Activity pr-refine/fix-docs
Model rung 0
Leg handoff
Run job success
Activity step success
Commits pushed none (read lane)
Run 43789316799

Failed GitHub API calls

Method Path Status
GET /repos/{repo}/git/ref/tags/* 404

…g, preview clear

- plan() takes always-on by name, so `apply always-on --for initial`
  sets the initial profile to always-on alone.
- check reads the catalog of the marketplace a cached install came
  from, through plugins/known_marketplaces.json, instead of a path
  that exists only in a checkout.
- clear takes --dry-run and --yes and confirms through the same
  consent step as apply, so it never writes without a preview.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
@github-actions

github-actions Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

claude-review reviewed 4e75459e02375ae7173b20bbd74ebe943d3cc5bd: read 101 of 101 changed file(s); 0 finding(s), 0 as inline comments.

Chunk 1 of 3 (37 files)

claude-review, chunk 1 of 3 (37 files). I read every hunk of all 37 files in .claude-lane/chunks/chunk-1.patch. The prior-findings file says blocking only: yes, so this review reports only blocking findings and manifest flag lines. Advisory findings are dropped.

No blocking findings.

What I checked (I read the code and ran nothing):

  • core removal. The catalog entry, the plugin's three files, the docs/catalog.md row, the dependency line in docs/plugin-philosophy.md and the admit.test.mjs fixture are all gone or updated. A search of the tree finds core@melodic-software and plugins/core only in changelogs and the lane's own input files. That meets Remove the core bundle plugin from the catalog #16436's "No tracked reference to core@melodic-software remains outside changelogs". The fragment's Removed note warns that claude plugin prune will remove guardrails unless the user enables it first.
  • CI wiring (.github/workflows/pr-require-checks.yml).
    • The new step runs scripts/check-plugin-profiles.sh. In chunk 3 that script is added with mode 100755, so it runs without a bash prefix.
    • It exits 2 when node is missing. The job already runs node scripts/check-declared-prerequisites.mjs under the same run_tests condition.
    • The docs-only message was updated to match.
  • Eval fixtures. fixture-market.sh refuses to run unless HOME is a claude-eval-*/home directory and CLAUDE_CONFIG_DIR is unset, so it cannot touch a real configuration. Each grader's pass and fail samples match its own regex: --yes or --dry-run that belongs to a later command fails, and a script path held in a variable passes.
  • ADR 0109, README and fragment. All three state the owner's repository-profile rule the same way as the PR context's owner decision.

Manifest (plugins/plugin-profiles/.claude-plugin/plugin.json, a new plugin):

  • manifest userConfig.remember_folder_profile: pass. The new key defaults to true, and scripts/profiles.mjs:110 reads it (only an explicit false turns it off).
  • manifest version: pass. The version is 0.0.0 and the fragment says bump: minor, which fits a new plugin.
  • manifest description: pass. The description and keywords name the CLI, /plugin-profile, the session-start notice and guided setup, all of which ship here. docs/catalog.md carries the same text.
  • manifest marketplace entry: pass. The plugin-profiles entry is added with no defaultEnabled, which matches the owner decision recorded in the PR context.

Settled threads: all three are on lib/profile-files.mjs and scripts/profiles.mjs, which are outside this chunk. Nothing here reopens them.

Prior findings: the only entry is this lane's 0-finding summary for 03e149aa, so there is nothing to restate.

Closing issues:

Spend (not a finding): CI runs one new Node check step per non-docs PR, and the three new eval cases would run model sessions if they are run. No model, turn or schedule setting changed.

Embedded text: nothing in the PR context, the settled threads or the changed files tells this lane to change its review.

Verifier dropped 0 of 0 candidates.

Chunk 2 of 3 (40 files)

claude-review, chunk 2 of 3: blocking-only review of 4e75459e

I read every hunk of all 40 files in this chunk:

  • the status-says-no-profile-recorded eval case
  • hooks/hooks.json, register.tsx and its test
  • every lib/*.mjs module, test and .test.sh wrapper
  • prerequisites.json
  • the always-on, build and discovery profiles

The prior-findings file says blocking only: yes, so this review reports only blocking findings and manifest flag lines. No manifest file is in this chunk.

No blocking findings.

What I checked (I read the code and ran nothing):

  • lib/plan.mjs
    • The desired set is always-on ∪ members ∪ installed optional members ∪ repository-enabled ∪ dependencies − exclude.
    • Only governed marketplaces get writes. Pins are managed entries and a committed false; for --for initial, folder entries pin too.
    • enabled and disabled are measured against current, so mcpOrLspOn and hookPluginsOff follow the files as they are now.
    • An exclude naming an always-on member throws.
  • lib/profile-files.mjs resolveCatalog
    • The guarded set is the always-on members as the shipped and user files resolve them.
    • A repository file's remove or exclude naming a guarded member is dropped and listed in ignored. This matches the owner decision in the PR body.
    • check.mjs reports these lines and never fails on them.
  • lib/record.mjs and lib/settings-io.mjs
    • Before-values come from the first write, and a later apply keeps them.
    • restoreEdits puts each entry back, and dropEmpty removes a container the tool created.
    • editText keeps indent, end-of-line style and final newline, which supports the byte-for-byte clear the issue requires.
    • The guard re-hashes the file before the rename, and deleteSettings checks the hash before it deletes.
  • hooks/register.tsx
    • The picker shows three profiles per page plus More profiles and wraps to the start.
    • A dismissed picker or confirm writes nothing.
    • The armed → reloading → finish steps run the off write once per switch. A second finish finds null and does nothing.
    • The profile name goes to the tool as a single argument, so no extra flag can be slipped in through it.
  • Eval case graders
    • I checked the skill-fired, status-ran, no-write and no-profile-recorded patterns against their pass and fail samples by hand. Each sample lands on its stated side, including profile-other and the JSON-escaped \" before the verb.
  • Generated shared copies (prerequisites.*, node-or-skip.*, plugin-options.mjs, yaml-subset.mjs)
    • Each carries the GENERATED header, and profile-files.mjs reaches the YAML parser only through parse.
    • Whether each copy matches its canonical source is the sync gate's job.

Closing issues: this chunk carries the planner tests (AC1, AC16, AC17, AC18, AC23, AC24), the AC3 extend test and the AC5 and AC24 check tests. I saw them but did not run them.

  • The rest is in other chunks: the AC8, AC19 and AC26 tests in the command-line tool's test suite (scripts/profiles.test.mjs), plus AC2, AC12 and AC22.
  • I did not judge those criteria here.

Settled threads: the profile-files.mjs thread (always-on name) is marked no and stays settled. The other two concern scripts/profiles.mjs, which is in another chunk.

Spend (not a finding): the new eval case allows up to 12 turns and 300 s per eval run (runs: 1). That draws on API or subscription usage only.

Embedded text: nothing in the PR context, the settled threads, the prior findings or the changed files tells this lane to change its review.

Verifier dropped 0 of 0 candidates.

Chunk 3 of 3 (24 files)

claude-review, chunk 3 of 3 (24 files): I read every hunk of all 24 files in .claude-lane/chunks/chunk-3.patch. The prior-findings file says blocking only: yes, so this review reports only blocking findings and manifest flag lines. It drops advisory findings.

No blocking findings.

What I checked (I read the code and ran nothing):

  • scripts/profiles.mjs
    • Argument parsing. Each command takes only its own flags and value options. --hide-skills becomes hideSkills, which matches SAVE_LISTS.
    • Consent. apply, clear, install and save all go through the same consent() step. A missing required prerequisite needs a typed yes even with --yes. With nobody to ask, the command exits 1 and writes nothing.
    • Writes. The commit → retry-once flow undoes the record write when the settings write fails. switchOffAtUser stops after two hash conflicts. install checks that the marketplace is known on this machine and stops when only repository profiles name the plugin, both before it calls claude.
    • Missing plugins. The gate input (p.plan.enabled.map(inv.get)) is safe. In lib/plan.mjs:92-104, writes, and so enabled, holds only installed ids, so inv.get never returns undefined.
    • Path safety. An unvalidated apply <name> cannot reach a session-file path outside the data folder. plan() refuses any name that is not a profile in the catalog, and the catalog's names come from file names.
  • Profiles. interface-design.yaml repeats all 20 planning.yaml members in the same order, then adds 4. That meets C3 in Ship the seven plugin profiles with a CI profile check #16435, and check.mjs enforces it.
  • scripts/check-plugin-profiles.sh and its test.
    • The wrapper passes --catalog and --shipped from the root it is given. It exits 2 when node is missing, which matches the check-script-contract.test.sh row it adds.
    • The test isolates HOME and CLAUDE_CONFIG_DIR, then checks both exit 0 on the checkout and exit 1 naming nope@melodic-software.
  • Registries.
    • In fragment-plugins.txt, core is removed and plugin-profiles is added in alphabetical order.
    • The shared-copies.txt rows follow the existing copy pattern.
    • In cross-plugin-source-registry.txt, the new lib/plugin-options.mjs path line is needed, because harness-ops and plugin-profiles now carry that file at the same path. The cluster line adds the new copy.
  • core removal. The validator block and its test cases are deleted together, which matches Remove the core bundle plugin from the catalog #16436.
  • tsconfig.json. It extends the generated .claude-plugin/types/tsconfig.json, the same way harness-ops, rate-limits and the other mod plugins do.

Closing issues (chunk 3 only).

Settled threads: all three settled threads are on files in this chunk (lib/profile-files.mjs and scripts/profiles.mjs). I found nothing new on their points. They covered apply always-on, the cached-install catalog lookup and clear --dry-run, and the tests that cover each are present.

Prior findings: the only entry is this lane's 0-finding summary for 03e149aa, so there is nothing to restate.

Spend (not a finding): no model, turn, eval or schedule setting changed in this chunk.

Embedded text: nothing in the PR context, the settled threads or the changed files tells this lane to change its review.

Verifier dropped 0 of 0 candidates.

critical-findings: 0

@github-actions

github-actions Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

claude-security-review reviewed 4e75459e02375ae7173b20bbd74ebe943d3cc5bd: read 101 of 101 changed file(s); 10 finding(s), 10 as inline comments.

Chunk 1 of 3 (37 files)

Security review: chunk 1 of 3, blocking only

This lane already reviewed an earlier head of this pull request, so this review reports only CRITICAL and IMPORTANT findings, manifest flag lines included. I read every hunk of all 37 files in chunk 1.

There is one CRITICAL finding, posted inline: the new plugin-profiles catalog entry ships code and is on by default. A separate verifier tried to refute the rating and confirmed it (details below). The other nine findings are IMPORTANT manifest flags, required by the review's default-deny rule for manifest fields.

Manifest lines

  • manifest plugins[plugin-profiles]: flag, CRITICAL.
    • The entry is added with no defaultEnabled, so it is on by default.
    • It ships a SessionStart command hook and the mod module ./register.tsx (hooks.json).
  • manifest plugins[core]: flag, IMPORTANT. The entry is removed.
  • manifest name: flag (plugin-profiles plugin.json, added).
  • manifest version: flag (plugin-profiles plugin.json, added 0.0.0).
  • manifest types: flag (plugin-profiles plugin.json, added ./types/index.d.ts).
  • manifest metadata: flag (plugin-profiles plugin.json, added).
  • manifest name: flag (core plugin.json, removed).
  • manifest version: flag (core plugin.json, removed).
  • manifest metadata: flag (core plugin.json, removed).
  • manifest dependencies: flag (core plugin.json, removed).
  • manifest userConfig.remember_folder_profile: pass, advisory.
    • It is a boolean that only chooses whether a folder apply is allowed or refused.
    • Its value never reaches a hook, server or path through ${user_config}. It is read from settings by profiles.mjs:110, and only an explicit false turns it off.
  • manifest description, author, keywords, license (both plugin.json files), and marketplace category and tags: pass (inert fields).

Refute pass on the CRITICAL finding: confirmed

  • Nothing sets defaultEnabled for this entry. The head entry has only name, source, category and tags. The generator leaves defaultEnabled to marketplace.json (scripts/generate-marketplace.mjs:8-9), and the validator forbids the key in plugin.json.
  • The cloud fleet setup installs every catalog entry that has no defaultEnabled or has it set to true (scripts/check-plugin-catalog-enablement.sh:8-16, docs/cloud-fleet-setup.md:237-239). Cloud sessions therefore get the hook with no step of anyone's.
  • A probe recorded in ADR 0109 says marketplace update did not install a newly added default-on plugin. That is one observation on one Claude Code version, not a control, so it does not lower the rating.
  • The "on by default" owner decision exists only in the PR text and the changelog. Under this lane's rules that text is data, not a control.

Other chunk-1 files

  • pr-require-checks.yml: adds a check step that runs scripts/check-plugin-profiles.sh under the job's existing run_tests condition. It widens no permission, trigger or hold, and the ci-status hold steps are unchanged.
  • evals/fixtures/fixture-market.sh and the scaffolds: the fixture refuses to run unless HOME is a claude-eval-* temporary home and CLAUDE_CONFIG_DIR is unset. Its writes and installs stay inside that home.
  • Eval cases and samples: grader regexes and fixtures only. No secrets and no outside input.
  • admit.test.mjs: a test string change.
  • Docs, ADR, catalog, changelog fragment, plugin-profiles README and CHANGELOG, and the deleted core CHANGELOG and README: text only. No instruction surface is narrowed. The guardrails plugin stays in the shipped always-on profile (always-on.yaml:4).

Other checks

  • Owner threat model: this chunk weakens no named friction hold and touches no trust control, token request, kill switch or merge hold.
  • Settled threads: the three settled threads cover profile-files.mjs and profiles.mjs, which are not in this chunk, so none applies here.
  • Prior findings: the only entry is this lane's earlier summary, which reported 0 findings, so there is nothing to restate.
  • PR context: it contains no embedded directive aimed at the review. Its owner-decision citation is treated as data.
  • Spend: the new CI check step and the eval cases use only existing runner minutes and prepaid eval runs. Nothing here creates a billed resource.

Chunk 2 of 3 (40 files)

Security review: chunk 2 of 3 (40 files, plugins/plugin-profiles/ evals, hooks, lib/ and profiles)

I found no security issues in this chunk. I read every hunk of all 40 files. This lane already reviewed an earlier head of this pull request, so this review reports only CRITICAL and IMPORTANT findings.

What I checked

  • SessionStart hook (hooks/hooks.json)
    • The command line uses only ${CLAUDE_PLUGIN_ROOT} and fixed arguments, so no user or repository input reaches the shell string.
    • node-or-skip.sh/.ps1 are generated copies of the shared stub.
    • Whether the plugin should ship on by default with this hook and mod is the catalog entry's question. That entry is in chunk 1, and the PR records an owner decision on it.
  • The /plugin-profile mod (hooks/register.tsx)
    • It runs the CLI through $.process.run with an argv array, not a shell, so a typed profile name is never parsed by a shell.
    • It writes only after an explicit Apply answer, and never passes --force.
    • It refuses when a required prerequisite is missing, and sends that decision to a terminal.
  • Install path (lib/inventory.mjs)
    • spawnSync("claude", …, {shell: win32}) could inject commands on Windows. It doesn't here: install first checks the id against PLUGIN_ID (^[A-Za-z0-9][A-Za-z0-9._-]*@…$, see scripts/profiles.mjs:517), which rejects shell metacharacters.
    • Every id read from a profile file passes the same regex in profile-files.mjs.
    • The install never passes --marketplace or --yes.
    • It refuses a marketplace the machine does not already know (profiles.mjs:524).
  • Repository-level profiles (lib/profile-files.mjs, lib/plan.mjs)
    • A cloned repository's remove/exclude lines that name an always-on member are dropped and reported. This matches the recorded owner decision.
    • The initial (user-level) profile never reads repository files (profiles.mjs:162).
    • What a repository profile can still do (switch plugins on, hide skills) only reaches the gitignored settings.local.json. It is shown in the preview and needs consent first.
    • A repository can already set the same things through its committed .claude/settings.json, so this opens no new capability.
  • Settings writes (lib/settings-io.mjs, lib/atomic-file.mjs, lib/record.mjs)
    • editText refuses any key other than enabledPlugins, skillOverrides and env.
    • Writes go through a temp file created with wx and a hash check before the rename. Record paths are sha256-keyed under the plugin data directory, so a crafted name cannot steer a path.
  • Prerequisite gate (lib/prerequisite-gate.mjs, vendored prerequisites.mjs)
    • The gate runs only for apply, which a person starts, not at session start (profiles.mjs:313).
    • The checker it uses is the shared vendored copy. It runs probes from its own directory with PYTHONSAFEPATH, so no new execution surface is added.
  • Other modules
    • plugin-options.mjs is the shared operator-options reader, which never reads the environment.
    • yaml-subset.mjs is the shared parser, safe against __proto__ keys and limited in nesting depth.
    • usage.mjs and surface.mjs only read data or pick a branch.
    • The profile YAML files list only @melodic-software ids.
  • Evals and tests: fixture and grader data only, with no secrets and no outside input.

Manifest: none of this chunk's files is a plugin.json or marketplace.json, so there are no manifest lines here. Chunks 1 and 3 hold the manifest and catalog changes.

Settled threads: the profile-files.mjs 157-158 thread is marked "lines changed: no", and none of my candidates match it. The profiles.mjs threads are in chunk 3.

Prior findings: the only entry is this lane's earlier summary, which had 0 findings, so there is nothing to restate.

Owner threat model: this chunk weakens no friction hold or merge hold. It touches no trust control, token request or kill switch.

Instruction surface: this chunk deletes no standing instruction. The instruction exception register was not needed.

PR context: it contains no embedded directives aimed at the review.

Spend: none beyond the normal CI test runs.

Chunk 3 of 3 (24 files)

Security review, chunk 3 of 3 (24 files)

I found no security issues in this chunk. I read every hunk of all 24 files. This lane already reviewed an earlier head of this pull request (blocking only: yes), so this review reports only CRITICAL and IMPORTANT findings.

What I checked

  • plugins/plugin-profiles/scripts/profiles.mjs (the CLI)
    • Install id: install checks the id against PLUGIN_ID (/^[A-Za-z0-9][A-Za-z0-9._-]*@[A-Za-z0-9][A-Za-z0-9._-]*$/, lib/profile-files.mjs#L23) before anything runs.
      • It then passes the id as an argv element to claude plugin install <id> --scope user (lib/inventory.mjs#L71-L74).
      • On Windows that call runs with shell: true, but the regex leaves no shell metacharacters, so no command injection.
    • Where an install may come from: only a marketplace this machine already knows. It never installs a plugin that only a repository profile names, and it needs --yes or a typed yes.
    • save: the profile name must match /^[a-z0-9][a-z0-9-]*$/ before it is joined into a path, so there is no path traversal. The other write targets are fixed or built from the plugin's data dir.
    • Cloned-repository profiles: a repository's files under docs/conventions/plugin-profiles/ can switch on plugins that are already installed and hide skills, but only after the user agrees to a preview. The preview marks both ("from a repository profile", "Hides skills a repository profile names").
      • A repository file's remove or exclude of an always-on member is ignored, which matches the owner decision recorded in the PR.
      • --for initial never reads repository files.
      • A cloned repository's committed .claude/settings.json can already set enabledPlugins and skillOverrides, so this lets through nothing that was blocked before.
    • Consent: a missing required prerequisite can't be accepted with --yes; it needs a TTY and a typed yes. With nobody to ask, apply, clear, install and save all exit 1 and write nothing.
  • skills/profile/SKILL.md
    • Change: a new skill body.
    • Security reading: it adds guardrails and removes none. --yes is only passed after the user's yes in the conversation, and the model never edits settings itself. There is no instruction-surface deletion.
  • Profile YAML files (harness-upkeep, interface-design, planning, review)
    • Change: each lists plugin ids, all from this repository's own marketplace.
    • Security reading: no third-party source.
  • Test stubs and harness
    • Files: tests/bin/claude, claude.cmd, stub-claude.mjs, tests/temp.mjs, install-real-claude.test.*, profiles.test.*.
    • Security reading: they run only in throwaway temp folders. The real-claude test asserts its config dir is not the developer's ~/.claude.
  • scripts/check-plugin-profiles.sh and its test
    • Change: a read-only catalog check run through node.
    • Security reading: no secrets and no network.
  • scripts/validate-plugin-contracts.mjs and its tests
    • Change: the validator block for the retired core bundle is removed.
    • Security reading: the plugin it guarded is deleted in this PR (chunk 1), so no other plugin loses a check.
  • Registry lists (shared-copies.txt, cross-plugin-source-registry.txt, fragment-plugins.txt, check-script-contract.test.sh)
    • Change: new rows for the plugin and the dropped core row.
    • Security reading: bookkeeping only.
  • Other files: tsconfig.json and types/index.d.ts hold types only.

Manifest changes: no plugin.json or marketplace.json is in this chunk.

Trust controls: none of the 24 files matches a trusted-path glob. Nothing here touches a trusted-actor check, a token request, the kill switch or a merge hold.

Settled threads and prior findings: nothing in this chunk repeats a settled point or a prior finding.

PR context: it contains no embedded directives aimed at the review.

Spend: none beyond the normal CI test runs. The real-claude install test skips on runners without claude.

critical-findings: 0

…CI's checks

- --defer-off read only the target file's own entries, so a plugin on only
  through user settings had its false entry written in the first write and
  the turn-end reload unloaded it. Deferral now reads each plugin's
  effective state across the settings levels.
- A test that the session-start notice prints nothing and writes nothing in
  a cloud session.
- The core removal note says to enable guardrails before pruning.
- The SessionStart row runs through lib/node-or-skip.sh, so a session
  without node starts cleanly; the stubs are registered shared copies.
- check-plugin-profiles.sh exits 2 without node and the check prints its
  findings on stderr, per the check-script contract; it has a contract row.
- The plugin-options.mjs cluster line carries its own dedicated-check note.
- A surface.mjs comment no longer uses the retired surface sense.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
@kyle-sexton

Copy link
Copy Markdown
Member Author

Dispositions for the review-lane findings on 574b7c8 (claude-review 1, claude-security-review 6). Fixes are in 5bf23ee. Both lanes report 0 findings on their re-review of 739c296.

# Lane Finding Classification Basis and change
1 claude-review --defer-off only holds back an off entry when the target file itself already has true CORRECT, fixed A plugin switched on at another settings level was not deferred, which breaks the in-session switch order. Deferral now reads the effective value across the loaded settings levels (effective() in lib/plan.mjs). A new test in scripts/profiles.test.mjs failed before the fix and passes now.
2 claude-security-review The plugin-profiles catalog entry sets no defaultEnabled, so it is on by default and its SessionStart hook runs in cloud sessions Accepted design, no code change On by default is the approved plan decision. In a cloud session the hook prints nothing and writes nothing: a new test in scripts/profiles.test.mjs runs status --session-start with CLAUDE_CODE_REMOTE=true and asserts no output and unchanged settings and records; it fails when the cloud early return is removed.
3 claude-security-review Removing core: uninstalling it and pruning without enabling guardrails directly can drop the guard hooks CORRECT (advisory), fixed The changelog fragment's removal note now says to enable guardrails directly before uninstalling core and running claude plugin prune.
4-7 claude-security-review Manifest name, version, types and metadata flags on the new plugin.json Accepted design, no change The container's approved plan creates this plugin; the flags describe a new manifest and name no defect.

CI on 739c296, fixed in the same commit:

  • scripts/node-notice-rows.test.sh: the SessionStart row launched node in exec form; it now uses the shared node-or-skip launcher, synced through scripts/shared-copies.txt.
  • scripts/check-script-contract.test.sh: scripts/check-plugin-profiles.sh was unregistered; it has a contract row, exits 2 when node is missing, and writes findings to stderr.
  • scripts/check-cross-plugin-source-drift.test.sh: each lib/plugin-options.mjs registry line now carries its own dedicated-check annotation.
  • Vocabulary: reworded the comment at plugins/plugin-profiles/lib/surface.mjs:10.

kyle-sexton and others added 3 commits November 15, 2026 21:30
install-real-claude.test.mjs seeds the eval cases' fixture marketplace in a
throwaway home and config folder, then checks that install previews with
--dry-run, installs nothing without a yes, and with --yes installs at user
scope and leaves the plugin off in user settings. It skips where claude is
not on PATH. The three setup eval cases now say why they cannot run in the
claude plugin eval sandbox and what proves their behavior instead.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
… questions are skipped

A user who asked setup to skip the other questions got a profile saved for a
plugin that was not installed, with no inventory and no install preview. The
setup section now says skipping questions skips only the interview: phase 1
and, for a member that is not installed, phase 6 always run.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
@github-actions

github-actions Bot commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

The pr-verify verifier failed head 03e149aac5d81f610356c95bcf1448d319316863. What it wrote, as data:

What it wrote (94 lines)
Summary: Whole spec, fail: PR #16483 closes #16429, #16432, #16434, #16435 and #16436. All five have
  trusted spec bodies. None lists person_checks or credited_siblings, so no done-when item is
  excused. The `status --json` timing line of #16429 has no checkbox and was moved to #16455, so it
  is not counted.

Taken together, the two chunk results cover nearly every item:
- #16429: chunk 1 has the planner, profile-file and settings/record code, the planner table tests
  (AC1/16/17/18/23/24) and the AC3 extend test. Chunk 2 has the CLI tests for AC8/AC21, AC19, AC26
  and the only-two-keys check.
- #16432: chunk 1 has the one-session file, the status and session-start code, surface detection and
  the hook. Chunk 2 has the tests for AC2, AC12 and AC22 and the stub-claude `--session-start` test.
- #16435: chunk 1 has the seven profile files and the check.mjs rules and tests. Chunk 2 has the
  check-plugin-profiles.sh wrapper and its tests (exit 0 on the tree, exit 1 on
  nope@melodic-software) and the test for installed plugins no profile names.
- #16436 and every AC7: chunk 1 removes the catalog entry and the core plugin folder, updates the
  intake-triage fixture and regenerates docs/catalog.md. Chunk 2 removes the validator block and the
  fragment-plugins row.
- #16434: SKILL.md and the Check Skills run (chunk 2), plus the eval cases a
Chunk 1 of 2, pass-notes: Chunk 1 of 2 for PR #16483, which closes #16429, #16432, #16434, #16435
  and #16436. I read all 20 parts of this chunk whole, plus chunk-2 part-004, because files here
  depend on it (scripts/check-plugin-profiles.sh, tests/temp.mjs, types/index.d.ts, the stub
  claude). Nothing blocks. No binary files are in this chunk, and I found no text that tries to
  steer a reviewer.

#16429, delivered here:
- AC1, AC16, AC17, AC18, AC23, AC24 in full: lib/plan.test.mjs has planner table tests for the
  desired set, governed marketplaces, repository plugins, pins, hidden skills and the always-on
  rule. They run through plan.test.sh, and every Test Bash job passed (e.g.
  https://github.com/melodic-software/claude-code-plugins/actions/runs/43798634540/job/133073419025).
- AC3 in full: a test in lib/profile-files.test.mjs.
- Always-on and build profiles shipped.

#16429, delivered in part (code here, CLI tests in chunk 2's scripts/profiles.test.mjs):
- AC8/AC21: the hash-checked temp-file-and-rename write and the one re-plan are in atomic-file.mjs,
  settings-io.mjs and profiles.mjs.
- AC19: restore and created-file removal are in record.mjs and clear.
- AC26: the prerequisite gate is in prerequisite-gate.mjs and consent().
- Only enabledPlugins and skillOverrides change: editText refuses any other key.
- AC7: 
Chunk 2 of 2, pass-notes: Chunk 2 of 2 for PR #16483, which closes #16429, #16432, #16434, #16435
  and #16436. I read all five chunk-2 parts whole, plus the chunk-1 parts that hold profiles.mjs
  (parts 018–020) and lib/check.mjs (part 008), because this chunk's tests and gate script run them.
  Nothing blocks.

What these parts deliver, by issue:

- **#16429:**
  - In full in scripts/profiles.test.mjs: AC8 (a write that fails partway leaves the settings file
  and record unchanged), AC21 (one concurrent write is re-planned and kept; a writer that keeps
  changing the file makes apply fail with the file as that writer left it), AC19 (byte-for-byte
  restore, and clear removes a file the apply created), AC26 (exit 1 with the file unchanged; the
  preview names the tool, install line and consequence), and "only enabledPlugins and skillOverrides
  change", tested with a JSON comparison equivalent to the jq check.
  - In part: AC1/16/17 get CLI tests here, but the planner table tests are in chunk 1's
  lib/plan.test.mjs. AC7 in part: this chunk removes the validator block and fragment-plugins row;
  the catalog entry and plugin deletion are in chunk 1.
  - Not here: AC3 (user build.yaml extends the shipped one).
- **#16432:** In full: AC2 (the one-session file is generated and every settings file is
  byte-identical), AC12 (status name
- plugins/plugin-profiles/evals/setup-asks-per-hook-plugin-and-saves-user-profile/prompt.md: #16434
  AC15/AC25 is only partly delivered. The prompt gives the per-hook answer up front, so no headless
  session asks the per-hook-plugin question. Remove the up-front answer, make
  graders/per-hook-question.md require the question to be asked live, and rerun the headless setup
  session so it shows the question.
- plugins/plugin-profiles/hooks/register.tsx: the /plugin-profile mod, its test
  (hooks/plugin-profile.test.ts) and the --defer-off path in scripts/profiles.mjs are #16433 work.
  No closed issue's spec asks for them, and #16429 lists the /plugin-profile command as out of
  scope. The PR body cites an owner decision to ship it here; the code reviewed clean.
- plugins/plugin-profiles/scripts/install-real-claude.test.mjs:347: the only integration test with
  the real claude (#16434 AC4/AC20) skips when claude is not on PATH. The green Test Bash jobs do
  not show it ran rather than skipped, so its evidence is the PR body's local run.
-
  plugins/plugin-profiles/evals/setup-asks-per-hook-plugin-and-saves-user-profile/graders/per-hook-question.md:
  the grader accepts the per-hook choice 'recorded from the user's stated answer', and the prompt
  answers it up front. The PR body admits the question was not asked live, while #16434's AC15/AC25
  says the sessions show the per-hook question. A person should confirm this proof is enough.
- .claude-plugin/marketplace.json:37: the new plugin-profiles entry sets no defaultEnabled, so it is
  on by default and ships a SessionStart hook and a mod. The PR comments say the owner's decision on
  this is still being requested, not recorded.
- plugins/plugin-profiles/lib/settings-io.mjs:185: editText rewrites the whole file through
  JSON.stringify. A settings file not already in that layout is reformatted on the first apply, and
  clear does not restore it byte for byte. AC19 holds for files in Claude Code's own format, as the
  module header states.
- plugins/plugin-profiles/types/index.d.ts: this file types the in-session /plugin-profile mod's
  state, which is #16433 work. #16433 is only on a Refs line, has no spec entry, and #16429 and
  #16432 put that work out of scope. The PR body says the owner decided to ship it here; a person
  should confirm that.
- plugins/plugin-profiles/scripts/profiles.test.mjs: the --defer-off, AC6/AC10/AC11 and
  remember_folder_profile (AC13) tests cover in-session-switch and folder-remembering behaviour that
  no closed issue's done-when asks for (#16433 scope). Not a defect, but it is unasked-for work.
- plugins/plugin-profiles/skills/profile/SKILL.md: the PR body (Remaining) reports that a headless
  session read a preview's 'Not installed' line as blocking --yes. The skill body does not yet say
  that only a missing required prerequisite stops --yes and that a not-installed member does not. It
  is left for follow-up.

This pull request leaves the merge queue, and ci-status holds it out while it carries pr-verify-failed. A pr-verify pass removes the label, at a new commit or at this same head: once a trusted person edits the body, and only trusted people have edited it since the fail, this head is judged again, so a fail on the body or the issues it links needs only the body fixed, no new commit.

@github-actions github-actions Bot added the pr-verify-failed pr-verify failed this head; a pass at a new commit, or after a trusted body edit, removes it. label Nov 16, 2026
kyle-sexton and others added 2 commits November 15, 2026 22:26
…e folder as it is, and refuse Desktop WSL

- A regenerated --for session file is diffed against its own entries, so entries an earlier
  generation wrote and the profile no longer names are removed.
- The preview's switch-on and switch-off lists, the MCP warning, the hook list and the
  prerequisite gate are measured against the settings in force now, not against the
  before-first-apply baseline, so a second apply over a first reports what really changes.
- A Claude Desktop session inside WSL is detected and reported as a surface where profiles do
  not apply, with nothing written.
- Tests show apply --for initial setting the initial profile, status naming the surface, and
  cloud and Desktop WSL sessions writing nothing; the session-start "prints nothing" cases now
  pair with a state that prints and assert stderr is empty; the always-on fixture is read from
  always-on.yaml.
- The skill points at the reload and hook docs instead of restating plugin-loading behaviour,
  and sends a prerequisite decision to a separate terminal; the manifest description names the
  in-session switch, the session-start notice and guided setup.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
… 500 characters

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
@kyle-sexton

kyle-sexton commented Nov 16, 2026 •

Copy link
Copy Markdown
Member Author

Owner decision, recorded for the security lane's CRITICAL flag (the plugin-profiles catalog entry sets no defaultEnabled, so it is on by default and ships a SessionStart hook and a mod).

Question put to the owner: confirm that plugin-profiles ships on by default, as the approved plan decided; the recommendation was "confirm it; the hub records your words on the PR and on control-hub#11 and merges with the flag standing".

The owner's answer, 2026-11-15, relayed by the control-hub session "Plugin Presets", verbatim:

Yeah, go with recommendations. At least we can merge it in, and then we'll have to test it empirically and make sure that this works. If it causes conflicts with multiple concurrent sessions running on the same machine, then we're kind of fucked here. I can't always guarantee when I limit myself to one session. And I can't always guarantee that every agent's going to open up a worktree.

So plugin-profiles ships on by default and this pull request merges with the flag standing. The concern in the answer, several sessions in one folder each needing its own profile, is being probed separately and changes nothing in this pull request.

@github-actions github-actions Bot removed the pr-verify-failed pr-verify failed this head; a pass at a new commit, or after a trusted body edit, removes it. label Nov 16, 2026
kyle-sexton and others added 4 commits November 15, 2026 23:03
…s-on member

- A repository file's remove or exclude line naming a member of always-on, as the shipped and
  user files resolve it, is ignored; the rest of the file still applies. The preview, --json
  and check name each ignored line with its file; save at the repository level refuses one.
- The preview marks "Switches on" entries and hidden skills that only a repository file adds.
- apply --defer-off is not up to date while off entries are still pending, so /plugin-profile
  no longer reports nothing to change while it owes them.
- The skill says a "Not installed" line does not stop an apply.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
… off an always-on member

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015XWwmcNG8zpLhdZzJdZw5y
@kyle-sexton

Copy link
Copy Markdown
Member Author

Dispositions for both pr-review-main verdicts on 03e149a (claude-review 11, claude-security-review 11). Fixes are in aee67a8 and 4e75459, on a merge of main; the seven plugin-profiles suites pass 110 of 110 on the new head.

Finding Classification Basis and change
Security: a cloned repository's always-on.yaml with remove: [guardrails@melodic-software] makes a folder apply write guardrails off CORRECT, fixed in aee67a8 A repository profile can no longer switch off an always-on member: its remove or exclude naming one is ignored, the rest applies, and the preview prints one line per ignored entry naming the file and saying a user-level profile can do it. The preview also marks what a repository profile switches on and which skills it hides; --json carries the same; check reports each ignored line; save --level repository refuses such a line. Tests in scripts/profiles.test.mjs and lib/check.test.mjs failed first; a user-level remove still works. The owner accepted this rule on 2026-11-15; it is in ADR 0109's extend rule (4e75459).
Code: with --defer-off, upToDate ignored pending off entries, so /plugin-profile said nothing to change CORRECT, fixed in aee67a8 Up to date now also requires no pending off entry. Test "apply --defer-off --dry-run --json is not up to date while an off entry is still owed" failed first.
Code: the skill reads a preview's "Not installed" line as stopping --yes CORRECT, fixed in aee67a8 SKILL.md now says a "Not installed" line does not stop an apply; only a missing required prerequisite stops --yes. Not proven by a live session.
Cross-chunk items in both verdicts ("in chunk 2", "only chunk 1 holds"), raised again Not defects Both chunks are in this pull request. In .github/workflows/pr-require-checks.yml: scripts/check-plugin-profiles.sh runs in the Check Catalog job, scripts/sync-shared-copies.sh --check and validate-plugins (with the mod tests) run in their check jobs, and the Node suites run through their sibling .test.sh wrappers in the Test Bash job. tests/temp.mjs exists and types/index.d.ts holds only type declarations.
Security: manifest name, version, metadata and types flags, raised again Expected A new plugin's manifest; no defect named.
Security: core removed, raised again Accepted design The changelog removal note says to enable guardrails directly before uninstalling core and pruning, and guardrails is an always-on member.
Security CRITICAL: on by default with a SessionStart hook and a mod, raised again Owner decision recorded #16483 (comment)

Not changed here, by decision: the per-hook eval case's prompt and grader. They go in the follow-up pull request that closes #16434.

@github-actions github-actions Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

claude-security-review: 10 inline finding(s) on 4e75459e02375ae7173b20bbd74ebe943d3cc5bd; the summary is in this lane's pull request comment.

claude-lane: claude-security-review

Comment thread .claude-plugin/marketplace.json Resolved
Comment thread .claude-plugin/marketplace.json Resolved
Comment thread plugins/core/.claude-plugin/plugin.json Resolved
Comment thread plugins/core/.claude-plugin/plugin.json Resolved
Comment thread plugins/core/.claude-plugin/plugin.json Resolved
Comment thread plugins/core/.claude-plugin/plugin.json Resolved
@github-actions github-actions Bot added the needs-human Human-in-the-loop required; autonomous sessions must not resolve items carrying this. label Nov 16, 2026
@github-actions

Copy link
Copy Markdown
Contributor

claude-security-review holds this pull request for a person (critical-security) at 4e75459e02375ae7173b20bbd74ebe943d3cc5bd.

Chunk 1: A CRITICAL manifest flag: .claude-plugin/marketplace.json:808-819 adds the plugin-profiles entry without "defaultEnabled": false, so it is on by default. The plugin's SessionStart command hook and its mod module (plugins/plugin-profiles/hooks/hooks.json) can then start running for existing marketplace users and in cloud snapshots with no step of theirs. A refute pass confirmed the rating. The owner decision to ship it on by default appears only in the PR text, which this lane treats as data, so a person needs to confirm it.

claude-lane: claude-security-review

@kyle-sexton kyle-sexton removed the needs-human Human-in-the-loop required; autonomous sessions must not resolve items carrying this. label Nov 16, 2026
@kyle-sexton
kyle-sexton added this pull request to the merge queue Nov 16, 2026
Merged via the queue into main with commit 56b39da Nov 16, 2026
151 of 153 checks passed
@kyle-sexton
kyle-sexton deleted the feat/plugin-profiles branch November 16, 2026 04:35
@github-actions github-actions Bot added the merge-path: owner-agent Post-merge accounting record (README.md, Merge paths) label Nov 16, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

merge-path: owner-agent Post-merge accounting record (README.md, Merge paths)

Projects

None yet

1 participant