interdev

keep priority P3 effort S wave 3

What it does: A lean 2-skill companion plugin (no agents/commands/hooks/own MCP server). Skill mcp-cli documents the third-party mcptools CLI (f/mcptools) for on-demand, out-of-context MCP discovery+invocation (mcp tools/resources/prompts/call, aliases, mcp guard allow/deny, stdio/HTTP/SSE transports, basic/bearer/env auth). Skill working-with-claude-code is a bundled offline mirror of 41 official Claude Code docs in references/ plus an update_docs.js fetcher that scrapes docs.claude.com/llms.txt. Both ship SKILL-compact.md variants for progressive disclosure. Skill-authoring and plugin-dev were deliberately extracted to interskill/interplug (2026-02-25), leaving interdev tightly scoped.

Rationale: Mature, narrowly scoped, structurally tested (tests/structural assert 2 skills + frontmatter), and squarely on a winning 2026 frontier theme. The mcp-cli skill is exactly the MCP-context-bloat antidote the SOTA findings champion: it interacts with MCP servers WITHOUT loading their tool schemas into the context window — the same goal as Tool Search Tool / RAG-MCP / Code Mode (one measurement cited 3 servers eating 72% of a 200K window before any query). It even documents mcp guard allow/deny tool patterns, which lands on the MCP-security/tool-poisoning trend (MCPTox >60% ASR, Snyk ToxicSkills). The working-with-claude-code skill operationalizes the user’s own ‘ground in live docs, not memory’ discipline and is reasonably fresh (covers marketplace plugins, subagents, skills+progressive disclosure). Nothing here is structurally outdated; the only real drift is point-in-time snapshot freshness, which is a small in-place fix, not a redesign.

SOTA gap: Bundled CC docs are a static Apr-2026 snapshot with no freshness metadata or staleness gate — they reference Tool Search Tool / defer_loading only once, lagging the progressive-disclosure / deferred-tool frontier the rest of the ecosystem (and this very harness’s ToolSearch) is built on. update_docs.js is a fire-and-forget regex scraper with no ‘last-fetched’ record or diff/age check, so the skill cannot tell the user when its mirror is stale.

Redundancy: Bounded and intentional. Sibling overlap with interplug (plugin dev) and interskill (skill authoring) is by design — those concerns were explicitly extracted from interdev in 2026-02-25, leaving clean seams; no merge warranted. The working-with-claude-code reference-library role partially overlaps interknow/qmd as a knowledge source, but offline-bundled official docs are a distinct, deliberate approach (no embedding/index dependency). mcp-cli has no overlap with any other plugin (intermux/interlock are coordination, not MCP-server interaction).

Actions:

Add freshness tracking to working-with-claude-code: have update_docs.js write a references/_meta.json with fetch timestamp + per-doc hash, and add a one-line staleness note to SKILL.md so the agent can warn when the mirror is older than N weeks (drives the only modernize lever — keeps pace with the progressive-disclosure/Tool-Search-Tool doc churn).
Re-run update_docs.js to pull current Claude Code docs (Tool Search Tool / defer_loading, code-execution-with-MCP, /effort) so the bundled mirror reflects the deferred-tool frontier rather than the Apr-2026 snapshot.
Fix likely-incorrect package names in mcp-cli SKILL.md (e.g. @anthropic/mcp-server-brave-search, @anthropic/mcp-server-puppeteer) and add a short ‘security: treat tool descriptions as untrusted input; prefer mcp guard allow-lists’ note aligned to the MCPTox/tool-poisoning findings.
Add a one-paragraph cross-reference in mcp-cli explaining when to use mcptools CLI (one-off / unconfigured servers, out-of-context) vs the harness’s native ToolSearch/deferred-tool flow (already-registered servers) so the two progressive-disclosure mechanisms don’t get confused.