slbug · Nov 13, 2026 · Nov 13, 2026 · Nov 13, 2026 · Nov 13, 2026
diff --git a/.claude-plugin/marketplace.json b/.claude-plugin/marketplace.json
@@ -5,7 +5,7 @@
   },
   "metadata": {
     "description": "Claude Code plugins by Alexander Grebennik",
-    "version": "1.16.19"
+    "version": "1.16.20"
   },
   "plugins": [
     {
@@ -19,7 +19,7 @@
         "source": "git-subdir",
         "url": "slbug/claude-ruby-grape-rails",
         "path": "plugins/ruby-grape-rails",
-        "ref": "v1.16.19"
+        "ref": "v1.16.20"
       },
       "relevance": {
         "topic": "Rails",

diff --git a/.claude/rules/agent-development.md b/.claude/rules/agent-development.md
@@ -67,12 +67,22 @@ one model response count as one turn. The `Complete analysis by turn
 ~M` rule is satisfied when the agent's M-th model invocation has
 completed.
 
-## Foreground Dispatch
-
-Plugin agents are spawned foreground only. No skill body, fanout
-template, or example MAY pass `run_in_background: true` on an
-`Agent(...)` call. Parallel = multiple Agent tool calls in one
-message; never the background flag.
+## Agent Dispatch
+
+No skill body, fanout template, or example MAY pass
+`run_in_background: true` on an `Agent(...)` call. Parallel = multiple
+Agent tool calls in one message; never the background flag.
+
+| Session | Runtime behavior |
+|---|---|
+| Interactive terminal, VS Code, desktop (fork mode on, the default) | Main-conversation subagents start in the background. The Agent tool exposes no `run_in_background` parameter. Completion arrives as a notification |
+| `claude -p`, Agent SDK, or `CLAUDE_CODE_FORK_SUBAGENT=0` | Claude chooses foreground or background. The flag exists — the prohibition above applies |
+
+Every skill that spawns an agent and consumes its result — fanout or
+single agent — MUST carry an explicit "wait for the agent(s) to
+complete" step between the spawn step and the first artifact read,
+Artifact Recovery step, or use of the return text. Without it, a
+background agent's result is consumed before it exists.
 
 ## Memory
 

diff --git a/.claude/rules/hook-development.md b/.claude/rules/hook-development.md
@@ -32,6 +32,19 @@ Add or update a short policy comment near the top of the hook when behavior chan
   ignored on `clear`/`compact`). Use the JSON form to combine it with
   `additionalContext`. Read input `session_title` first to avoid clobbering
   a user `/rename` (see `check-scratchpad.sh`)
+- `UserPromptSubmit` receives a typed skill or slash command as the
+  line the user typed (`/rb:plan ...`, `/ruby-grape-rails:plan ...`),
+  NOT the expanded skill body, and runs before the expansion. Match
+  both prefixes on `.prompt` (see `session-title.sh`). Use
+  `UserPromptExpansion` to match on the resolved command name
+- `PostToolUse` command hooks get 10 seconds, whatever their `timeout`,
+  when the user ends the turn with `Ctrl+Enter` while the tool runs.
+  CC kills a hook still running at that point. Only `async: true` hooks
+  outlive it. Keep every blocking `PostToolUse` hook safe to kill
+  mid-run — no partial writes, temp files cleaned by `trap <fn> EXIT`.
+  Register cleanup on `EXIT` only: bash runs the `EXIT` trap when a
+  signal terminates it, while a cleanup-only handler on `HUP`/`INT`/`TERM`
+  swallows the signal and the script continues to its normal exit
 - `SubagentStart` uses `hookSpecificOutput.additionalContext` to inject context into subagents
 - `PostToolUseFailure` uses `hookSpecificOutput.additionalContext` for debugging hints
 

diff --git a/.claude/rules/skill-development.md b/.claude/rules/skill-development.md
@@ -83,9 +83,14 @@ skills/{name}/
   subdir-per-topic (research per-aspect). See
   `plugins/ruby-grape-rails/references/run-manifest.md` for schema,
   staleness rules, and write protocol.
-- Agent dispatch: foreground only. Skill bodies MUST NOT pass
+- Agent dispatch: skill bodies MUST NOT pass
   `run_in_background: true` on any `Agent(...)` call. Parallel = many
   Agent tool calls in a single message, NOT the background flag.
+  Interactive sessions start main-conversation subagents in the
+  background regardless, so every skill that spawns an agent and
+  consumes its result — fanout or single agent — MUST state an explicit
+  wait-for-completion step before its first artifact read. See
+  `.claude/rules/agent-development.md` § "Agent Dispatch".
 
 ## Colon Naming
 

diff --git a/.claude/skills/cc-changelog/references/analysis-rules.md b/.claude/skills/cc-changelog/references/analysis-rules.md
@@ -50,7 +50,7 @@ Imperative-only. Map CC changelog entries to specific plugin components.
 | `marketplace.json` schema | `.claude-plugin/marketplace.json` |
 | `${CLAUDE_PLUGIN_DATA}` | `plugins/ruby-grape-rails/hooks/scripts/setup-dirs.sh`, `plugins/ruby-grape-rails/hooks/scripts/log-progress.sh` |
 | `${CLAUDE_PLUGIN_ROOT}` | All hooks.json paths |
-| `settings.json` keys | `plugins/ruby-grape-rails/settings.json` ships only `subagentStatusLine`. Documented supported keys per CC docs: `agent` + `subagentStatusLine`. New documented keys → INFO unless plugin adopts. |
+| `settings.json` keys | Plugin ships no `settings.json`. `subagentStatusLine` is removed from CC (loads, no effect). A plugin's `agent` setting is deprecated. New documented plugin-level keys → INFO unless plugin adopts. |
 | Plugin validation changes | CI workflow |
 
 ### Tool System

diff --git a/.claude/skills/docs-check/SKILL.md b/.claude/skills/docs-check/SKILL.md
@@ -116,6 +116,10 @@ Write findings to .claude/docs-check/reports/{surface}-report.md
 Agent body owns BLOCKER/WARNING/INFO/PASS classification, "no large
 docs paste" rule, "stop after returning". Call site stays minimal.
 
+Wait for every `docs-surface-validator` to complete before step 7 —
+interactive sessions run them in the background and notify on
+completion. A report is MISSING only after its worker has completed.
+
 ### 6. Structural Baseline (Always)
 
 Keep these results in view while synthesizing:

diff --git a/.claude/skills/docs-check/references/doc-pages.md b/.claude/skills/docs-check/references/doc-pages.md
@@ -37,7 +37,7 @@ Primary cached pages (plugin-critical):
 | `env-vars.md` | Environment variable contract: `CLAUDE_PLUGIN_ROOT`, `CLAUDE_PLUGIN_DATA`, `CLAUDE_PROJECT_DIR`, etc. |
 | `errors.md` | Error taxonomy driving hook failure classification and `StopFailure` matchers |
 | `cli-reference.md` | `claude plugin validate`, `--plugin-dir`, flags used by eval/doctor scripts |
-| `statusline.md` | Plugin-level `subagentStatusLine` setting schema |
+| `statusline.md` | Status line surface — confirms no plugin-level status line setting has returned |
 | `discover-plugins.md` | Marketplace install flow, `git-subdir` path resolution |
 | `sandboxing.md` | Bash sandboxing semantics affecting hook side-effect guarantees |
 | `context-window.md` | PreCompact/PostCompact hook timing and payload |
@@ -70,7 +70,7 @@ Additional cached pages (context/reference):
 | `permission-modes.md` | Permission modes and implications for plugins |
 | `permissions.md` | Comprehensive permissions guide |
 | `fast-mode.md` | `/fast` semantics and eligible models — relates to `effort: max` skill settings |
-| `output-styles.md` | Output-style surface adjacent to statusline + `subagentStatusLine` |
+| `output-styles.md` | Output-style surface adjacent to the status line |
 | `troubleshooting.md` | Hook failure diagnosis patterns |
 | `common-workflows.md` | Plan/Work/Review lifecycle comparison with `/rb:*` workflows |
 | `whats-new/index.md` | Weekly release notes index — `cc-changelog` + docs-check feed |
@@ -194,7 +194,7 @@ Use:
 Questions this answers:
 
 - Is the `claude plugin ...` subcommand used by doctor/eval still documented?
-- Does plugin-level `subagentStatusLine` match the documented schema?
+- Do docs reintroduce a plugin-level status line or plugin `settings.json` key? The plugin ships none.
 
 ### Built-in Feature Overlap
 

diff --git a/.claude/skills/docs-check/references/validation-rules.md b/.claude/skills/docs-check/references/validation-rules.md
@@ -313,8 +313,10 @@ Checks for `.claude-plugin/plugin.json`:
    - metadata: `$schema`, `name`, `displayName`, `version`,
      `description`, `author`, `homepage`, `repository`, `license`,
      `keywords`, `metadata`, `defaultEnabled`, `dependencies`
-   - behavior: `settings` (only `agent` and `subagentStatusLine` take
-     effect), `userConfig`, `channels`
+   - behavior: `userConfig`, `channels`. The plugin ships no
+     `settings.json` — `subagentStatusLine` is removed (loads, no
+     effect) and a plugin's `agent` setting is deprecated (warns in
+     `claude plugin validate`)
    - components: `skills`, `commands`, `agents`, `hooks`, `mcpServers`,
      `lspServers`, `outputStyles`, `workflows`
    - `experimental.themes`, `experimental.monitors`, `experimental.evals`

diff --git a/.claude/skills/session-deep-dive/SKILL.md b/.claude/skills/session-deep-dive/SKILL.md
@@ -81,6 +81,7 @@ Spawn one analysis subagent per transcript. Include:
 - reminder that metrics are heuristic
 
 Write each report to `.claude/session-analysis/{short_id}-report.md`.
+Wait for every subagent to complete before step 5.
 
 ### 5. Compress When Needed
 

diff --git a/.github/copilot-instructions.md b/.github/copilot-instructions.md
@@ -34,15 +34,13 @@ PR review: flag prose that violates this rule.
 | Hooks | `plugins/ruby-grape-rails/hooks/hooks.json` + `scripts/*` | Shell + Ruby triggered by CC events (PostToolUse, PostToolUseFailure, SessionStart, etc) |
 | Executables | `plugins/ruby-grape-rails/bin/*` | Plugin CLIs added to Bash PATH; no extension; chmod +x; mixed bash + Ruby |
 | Library | `plugins/ruby-grape-rails/lib/*.rb` | End-user runtime modules (Ruby ≥ 3.4, stdlib only) |
-| Plugin settings | `plugins/ruby-grape-rails/settings.json` | Only `agent` + `subagentStatusLine` keys supported per CC docs |
 | Eval | `lab/eval/` | Deterministic Python eval framework for plugin quality |
 | Contributor tooling | `.claude/` | NOT shipped; `.claude/rules/` (auto-loaded path-scoped rules), `.claude/skills/` (contributor-only skills) |
 
 ### Currently-shipped binaries
 
 | Binary | Purpose |
 |---|---|
-| `subagent-statusline` (bash) | advisory statusline renderer; fail-open |
 | `detect-stack` (Ruby) | `/rb:init` stack detection |
 | `extract-permissions` (Ruby) | transcript-based Bash permission recommender |
 | `resolve-base-ref` (bash) | eval-able BASE_REF resolver for diff comparisons |
@@ -151,10 +149,8 @@ the PR.
   `hooks/hooks.json` references, sourcing in other `*.sh` files
   (`source ".../LIB_NAME.sh"`), matching tests under
   `lab/eval/tests/test_runtime_scripts.py`.
-- **`bin/<exec>` renamed / added / removed** → also check
-  `plugins/ruby-grape-rails/settings.json` (statusline command path),
-  `plugins/ruby-grape-rails/hooks/scripts/install-statusline-wrapper.sh`,
-  any skill/agent/script that shells out, `CLAUDE.md` "Executables"
+- **`bin/<exec>` renamed / added / removed** → also check any
+  skill/agent/script that shells out, `CLAUDE.md` "Executables"
   enumeration.
 - **New skill / agent / hook added** → also check corresponding
   `lab/eval/evals/`, `lab/eval/triggers/`, `lab/eval/dimensions/`

diff --git a/.github/instructions/plugin-review.instructions.md b/.github/instructions/plugin-review.instructions.md
@@ -121,17 +121,14 @@ Exception files (human-facing, narrative OK): `README.md`,
   the shebang; rescue at process boundary, never use `Kernel#eval` /
   `Kernel#system` with unvalidated input (Iron Law 12)
 - Advisory fail-open pattern (empty stdout, exit 0 on any error) is
-  intentional for statusline and similar advisory executables — do NOT
-  flag as "missing error handling"
+  intentional for advisory executables — do NOT flag as "missing error
+  handling"
 - From `hooks.json`, `.mcp.json`, `monitors/monitors.json`, template
   expansion is supported; reference bundled scripts via
   `${CLAUDE_PLUGIN_ROOT}/...`
 
 ### Currently shipped binaries (plugins/ruby-grape-rails/bin/)
 
-- `subagent-statusline` (bash) — advisory, fail-open. Empty stdout +
-  exit 0 on any error. Referenced from `settings.json` indirectly via
-  `~/.claude/ruby-grape-rails-subagent-statusline` wrapper
 - `detect-stack` (Ruby) — `/rb:init` stack detection. Outputs
   `key=value` pairs; treats unreadable manifests as absent
 - `extract-permissions` (Ruby) — analyzes session transcripts to
@@ -288,32 +285,12 @@ Notes:
 
 ## Plugin Settings (plugins/**/settings.json)
 
-- Only `agent` and `subagentStatusLine` keys are supported per CC docs
-  (`plugins-reference.md` standard plugin layout)
-- Unknown keys are silently ignored by CC — do NOT flag partial coverage
-  of other settings fields
-- `subagentStatusLine.command` does NOT expand `${CLAUDE_PLUGIN_ROOT}`
-  and CC does NOT export `CLAUDE_PLUGIN_ROOT` to the statusline
-  subprocess nor add plugin `bin/` to its PATH. Plugin-bundled
-  statusline scripts therefore require a SessionStart hook that writes
-  a small wrapper at `~/.claude/<plugin-id>-subagent-statusline`
-  pointing at the current absolute plugin path. The plugin
-  `settings.json` then references that stable user-home path. The
-  wrapper must be rewritten only when its content differs from the
-  desired content (plugin version changes change the absolute path).
-  Do NOT flag this indirection as unnecessary — it is required by the
-  documented CC substitution scope
-
-### subagentStatusLine payload schema (as observed, not fully documented)
-
-The statusline subprocess receives base hook fields plus `columns` and a
-`tasks[]` array. Each task provides `id`, `type` (e.g. `local_agent`),
-`status`, `description`, `label` (usually the same text as `description`),
-`startTime` as epoch milliseconds (13-digit integer), `tokenCount`,
-`tokenSamples` as a number array, and `cwd`. The docs also list `.name`
-but current CC payloads do NOT include it — match emoji/label from
-`.label` first, falling back to `.description` and finally `.name`. Do
-NOT flag the fallback chain as over-engineered.
+- The plugin ships NO `settings.json`. Flag any PR that adds one
+- `subagentStatusLine` is removed from CC — the key loads and has no
+  effect. Flag any reintroduced statusline script, wrapper, or setting
+- A plugin's `agent` setting is deprecated in CC — it still applies and
+  warns at startup, in `claude plugin list`, and in
+  `claude plugin validate`. Flag any PR that adds it
 
 ## Do NOT Flag
 
@@ -322,8 +299,6 @@ NOT flag the fallback chain as over-engineered.
 - `omitClaudeMd: true` on specialist agents (intentional context savings)
 - Colon-namespaced skill names like `rb:plan` (known compatibility item)
 - Advisory fail-open (empty stdout, exit 0) in `bin/` and advisory hooks
-- Partial coverage in plugin `settings.json` (only `agent` and
-  `subagentStatusLine` are documented as supported)
 - Pointer-only links to `references/preferences/tool-batching.md`
   from skill or agent bodies (canonical examples doc for the
   tool-batching preference). Pointer = the link without restating
@@ -399,7 +374,12 @@ NOT flag the fallback chain as over-engineered.
   one consolidated verdict per artifact.
 - `run_in_background: true` on any `Agent(...)` call in shipped
   skill bodies, fanout templates, or example snippets — flag as
-  BLOCKER. Plugin agents dispatch foreground only.
+  BLOCKER.
+- A skill — fanout or single agent — whose spawn step is followed by an
+  artifact read, Artifact Recovery step, or use of the return text with
+  no explicit wait-for-completion step between them. Interactive
+  sessions run main-conversation subagents in the background, so the
+  read would precede the result.
 - Raw `mv`, `cp`, `jq -i`, `rm`, or improvised shell against any
   `RUN-CURRENT.json` path or per-agent artifact path in shipped skill
   bodies, hook scripts, or examples — flag as BLOCKER. All manifest

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -7,6 +7,35 @@ Versioning: [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
 ## [Unreleased]
 
+## [1.16.20] - 2026-11-13
+
+### Removed
+
+- Custom subagent status line. Claude Code removed the
+  `subagentStatusLine` setting, so subagent rows always use the default
+  look. The plugin no longer ships `settings.json`,
+  `bin/subagent-statusline`, or the `install-statusline-wrapper.sh`
+  `SessionStart` hook. A leftover
+  `~/.claude/ruby-grape-rails-subagent-statusline` wrapper is unused and
+  safe to delete by hand.
+
+### Changed
+
+- `/rb:review`, `/rb:brainstorm`, `/rb:investigate`, and the
+  `/rb:research` provenance step state that spawned agents run in the
+  background in interactive sessions and that results are read only
+  after every agent completes. The "foreground" wording is dropped from
+  the `/rb:plan` and `/rb:brainstorm` spawn steps.
+
+### Fixed
+
+- `format-ruby.sh` removes its `ruby-format.*` stderr capture when the
+  hook is killed or exits mid-formatter, instead of only after the
+  formatter returns.
+- Contributor tooling: `markdownlint-cli` raised to `^0.50.0`, with
+  `overrides` pinning `minimatch` and `katex` to patched releases.
+  `npm audit` reports 0 vulnerabilities.
+
 ## [1.16.19] - 2026-10-01
 
 ### Changed
@@ -3150,7 +3179,8 @@ Prevents context exhaustion with 3 compression strategies
 - 100+ reference documents across all skill domains
 - Plugin development guide with size guidelines and checklists
 
-[Unreleased]: https://github.com/slbug/claude-ruby-grape-rails/compare/v1.16.19...HEAD
+[Unreleased]: https://github.com/slbug/claude-ruby-grape-rails/compare/v1.16.20...HEAD
+[1.16.20]: https://github.com/slbug/claude-ruby-grape-rails/compare/v1.16.19...v1.16.20
 [1.16.19]: https://github.com/slbug/claude-ruby-grape-rails/compare/v1.16.18...v1.16.19
 [1.16.18]: https://github.com/slbug/claude-ruby-grape-rails/compare/v1.16.17...v1.16.18
 [1.16.17]: https://github.com/slbug/claude-ruby-grape-rails/compare/v1.16.16...v1.16.17