fix(jobs): kill/reaper lifecycle correctness + tests

.coderabbit.yaml

@@ -0,0 +1,162 @@
+# yaml-language-server: $schema=https://coderabbit.ai/integrations/schema.v2.json
+
+# mcp-ssh — CodeRabbit configuration
+#
+# Single Rust binary: an MCP server giving an AI agent remote shell + file
+# access to ONE host, over authenticated MCP Streamable HTTP at /mcp.
+# Authority: CLAUDE.md (Conventions + NEVER) is absolute. Cite the rule.
+
+language: en-US
+early_access: true
+enable_free_tier: true
+
+tone_instructions: "mcp-ssh: Rust 2024, tokio, axum, rmcp. CLAUDE.md NEVER rules are absolute — no password/token in logs/errors/responses, never run as root, never weaken auth, never add a 4th MCP tool (parametrize bash/job/file). Concrete diffs, skip nits."
+
+reviews:
+  profile: assertive
+  request_changes_workflow: true
+  high_level_summary: true
+  high_level_summary_placeholder: "@coderabbitai summary"
+  auto_title_placeholder: "@coderabbitai"
+  poem: false
+  review_status: true
+  collapse_walkthrough: false
+  sequence_diagrams: true
+  estimate_code_review_effort: true
+  suggested_reviewers: true
+  auto_assign_reviewers: false
+  abort_on_close: true
+
+  auto_review:
+    enabled: true
+    auto_incremental_review: true
+    ignore_title_keywords:
+      - "WIP"
+      - "DRAFT"
+      - "DO NOT REVIEW"
+    drafts: false
+    base_branches:
+      - main
+
+  tools:
+    shellcheck: { enabled: true }
+    markdownlint: { enabled: true }
+    yamllint: { enabled: true }
+    actionlint: { enabled: true }
+    hadolint: { enabled: true }
+    gitleaks: { enabled: true }
+    github-checks: { enabled: true, timeout_ms: 900000 }
+    ruff: { enabled: false }
+    biome: { enabled: false }
+    eslint: { enabled: false }
+    rubocop: { enabled: false }
+
+  path_instructions:
+    - path: "**/*"
+      instructions: |
+        Review only changed files and code directly affected. Behaviour
+        changes must update docs/ in the same PR. Flag missing tests for new
+        public behaviour.
+
+    # Rust source — CLAUDE.md Conventions
+    - path: "src/**/*.rs"
+      instructions: |
+        Enforce CLAUDE.md Conventions:
+        - No unwrap/expect outside main.rs and #[cfg(test)]; panic in the
+          request path crashes every client.
+        - Errors typed with thiserror; anyhow only at the main.rs boundary.
+        - Async end-to-end. No std::sync::Mutex on the request path (use
+          tokio::sync), never held across .await. No block_on; spawn_blocking
+          for blocking I/O.
+        - Newtype over bare primitives with meaning (JobId, not String).
+          Make illegal states unrepresentable.
+        - tracing spans (request_id, tool), never println!.
+        - Files <=300 LOC, one responsibility. clippy -D warnings + rustfmt
+          clean is the floor.
+
+    # Auth + OAuth — security review required
+    - path: "src/auth.rs"
+      instructions: |
+        SECURITY REVIEW. HTTP Basic middleware. Never log/return the
+        password. Flag any path that bypasses or weakens auth.
+    - path: "src/oauth/**"
+      instructions: |
+        SECURITY REVIEW. OAuth 2.1: discovery, dynamic registration, PKCE on
+        authorize+token, bearer validation. No token material in logs/errors.
+
+    # Config — fail fast, secrets
+    - path: "src/config.rs"
+      instructions: |
+        Validated at boot — missing auth creds fail at startup, never at
+        request time. MCP_SSH_ALLOWED_HOSTS must be required. Secrets never
+        logged.
+
+    # Tools — constant 3-tool surface
+    - path: "src/tools/**"
+      instructions: |
+        Exactly 3 tools (bash/job/file) dispatching on action. NEW capability
+        = a new param/action, never a 4th tool. Thin adapters — no business
+        logic in the dispatch arm.
+
+    # Jobs — backgrounding, log pagination, process-group kill
+    - path: "src/jobs/**"
+      instructions: |
+        Output streams to a per-job log file; poll paginates (cursor/limit)
+        so chatty commands never flood agent context. Jobs spawn in their own
+        process group so kills reach descendants. Kill escalates TERM->KILL.
+        Reaper drops jobs >24h, killing any still-Running group first.
+
+    # Integration tests — real server over HTTP
+    - path: "tests/**/*.rs"
+      instructions: |
+        Boot the server and issue real MCP requests over HTTP. New behaviour
+        ships with the test that proves it.
+
+    # CI
+    - path: ".github/workflows/**"
+      instructions: |
+        Runner: blacksmith-2vcpu-ubuntu-2404. Pin actions to major versions.
+        Required jobs: fmt, clippy (-D warnings), test.
+
+    # Dockerfile
+    - path: "Dockerfile"
+      instructions: |
+        Multi-stage build. Non-root user — never run as root. No build
+        secrets baked in.
+
+    - path: "**/*.md"
+      instructions: |
+        Lead with the rule, tables over prose, file paths over descriptions,
+        no filler. Valid links, accurate code examples.
+
+  path_filters:
+    - "!target/**"
+    - "!**/target/**"
+    - "!**/*.log"
+    - "!.env*"
+    - "!**/.DS_Store"
+
+chat:
+  auto_reply: true
+  art: false
+
+knowledge_base:
+  opt_out: false
+
+  web_search:
+    enabled: true
+
+  code_guidelines:
+    enabled: true
+    filePatterns:
+      - "**/CLAUDE.md"
+      - "docs/**"
+
+  learnings:
+    scope: auto
+
+  issues:
+    scope: auto
+
+  pull_requests:
+    scope: auto

CLAUDE.md

@@ -53,7 +53,8 @@ Keep this accurate — it's the navigation aid.
 | `src/config.rs` | env + TOML file config; fails fast if auth creds missing |
 | `src/auth.rs` | HTTP Basic auth middleware |
 | `src/oauth/` | minimal OAuth 2.1 server: discovery metadata, dynamic client registration, authorize + token with PKCE, bearer validation |
-| `src/jobs.rs` | job engine: run a command, return inline if fast (<2s) else a job id (or immediately when `bg`); output streams to a per-job log file, polled paginated; hourly reaper drops jobs >24h old |
+| `src/jobs/mod.rs` | job engine: run a command, return inline if fast (<2s) else a job id (or immediately when `bg`); output streams to a per-job log file, polled paginated |
+| `src/jobs/reaper.rs` | hourly reaper drops jobs >24h old (killing any still-`Running` group first); process-group kill helpers (TERM→KILL escalation), shared with `job(action="kill")` |
 | `src/tools/mod.rs` | MCP tool surface (`#[tool_router]`/`#[tool]` from rmcp): 3 tools (`bash`/`job`/`file`) dispatching on `action`. Thin adapters over jobs + files |
 | `src/tools/files.rs` | file operations (`tokio::fs`; `ls`/`find`/`grep` shelled out) |
 
@@ -125,7 +126,8 @@ Non-negotiable: SOLID, SRP, tested code. The bar: idiomatic, boring, readable Ru
 | Config / env / required creds | `src/config.rs` |
 | HTTP Basic auth | `src/auth.rs` |
 | OAuth 2.1 (discovery, registration, PKCE, bearer) | `src/oauth/` |
-| Running commands, backgrounding, job logs | `src/jobs.rs` |
+| Running commands, backgrounding, job logs | `src/jobs/mod.rs` |
+| Reaper eviction + process-group kill helpers | `src/jobs/reaper.rs` |
 | Tool definitions / MCP surface | `src/tools/mod.rs` |
 | File operations | `src/tools/files.rs` |