setup-interns (v0.1.0) — 8-phase install of the local `interns` MCP server:
detect `.common/lib/interns-mcp/` source, `pip install -e`, write
`.common/secrets/interns.env` with endpoint API keys (gitignored), register
`mcpServers.interns` in `~/.claude.json` with absolute Python interpreter
path + `cwd` so the runtime resolves config relative to project root.
Mirrors setup-projects-meta / setup-context7 confirmation-gate shape.
Description 899 chars (under the 900 budget per
feedback_skill_description_length_limit).
using-interns (v0.1.0) — runtime policy for `mcp__interns__*`. Per-session
permission grant mirrors project-discipline Rule 4: ask-mode default,
conversational grant ("разреши интернов" / "allow interns"), conversational
revoke, always-ask paths for `**/.env`, `**/secrets/**`, `**/*.key`,
`**/.ssh/**`, `**/.aws/credentials` etc with transitive rule (Claude can't
bypass by reading the file with the local Read tool and forwarding content),
cost-cap >$0.10 always asks, session-end automatic reset. Routing hints for
`bulk_text_read` (3+ files or one file >400 lines) and `transcript_distill`
(before .wiki/log.md updates / session summaries). Description 814 chars.
project-bootstrap (1.5.0 -> 1.6.0, MINOR — capability added):
* assets/CLAUDE.md.template: insert `delegate to interns when allowed`
between `follow project discipline` and `we're on Windows`.
* SKILL.md Step 5: same insertion in inline template + new commentary
paragraph explaining the trigger, no-op semantics, install pointer.
* SKILL.md Step 5.5: bootstrap-manifest table extended with
`setup-interns`, `using-interns`, `project-discipline` rows.
* README.md: Workflow Step 5.5 description + See also section both pick up
the new skills.
Root CLAUDE.md dogfood: `delegate to interns when allowed` line added.
.wiki/log.md decision entry. .tasks/STATUS.md task moved to 🟢 done.
.tasks/interns-skills-mvp.md per-task file with goal / key-files /
decisions log / open questions.
Build + install verified: `dist/{setup-interns,using-interns,
project-bootstrap}.skill` rebuilt; `bash scripts/install.sh ...` succeeded;
harness skill listing shows full descriptions for all three (no H1
fallback) — confirms the description budget held.
Versioning per project-discipline Rule 3:
* setup-interns: 0.1.0 (first edit of unversioned artifact, MAJOR=0).
* using-interns: 0.1.0 (first edit of unversioned artifact, MAJOR=0).
* project-bootstrap: 1.5.0 -> 1.6.0 (MINOR — adds capability without
breaking existing CLAUDE.md merge or manifest consumers).
Server runtime itself (`.common/lib/interns-mcp/`) is out of scope for this
task — the skills land lifecycle, policy, and bootstrap integration so any
machine where the MCP server is later installed already has Claude's
policy in place.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
3.9 KiB
3.9 KiB
using-interns
Runtime policy for the local interns MCP server. Delegates predictable bulk I/O and summarization tasks to cheap intern LLMs (DeepSeek / Kimi / Ollama-hosted) so Claude saves Anthropic quota.
using-interns governs usage of an installed server. Initial setup (pip install, .env write, MCP registration) is owned by setup-interns.
Full design: .wiki/concepts/interns-design.md (in this repo) — Layer 1 / 2 / 3 architecture, MVP catalog, always-ask paths, routing hints.
When it triggers
- CLAUDE.md contains
delegate to interns when allowed(added byproject-bootstrap1.6.0+). - User says: "use interns", "delegate to an intern", "разреши интернов", "allow interns", "go ahead with interns".
- Claude is about to read 3+ files, or one file > 400 lines, just to build context.
- Claude is about to update
.wiki/log.mdor write a session summary from a long transcript. - If
mcp__interns__*tools are missing, this skill delegates tosetup-internsbefore doing anything else.
What it does
- Picks the right intern for the task from a routing table (
bulk_text_readfor reads,transcript_distillfor summaries; future interns added in.common/config/interns/config.yaml). - Guards every call with a per-session permission grant that mirrors
project-disciplineRule 4 — ask-mode default, conversational grant / revoke, session-end reset. - Enforces always-ask paths —
.env,**/secrets/**,**/*.key,**/*.pem,**/.ssh/**,**/.aws/credentials,**/.netrcetc. — even when delegation is granted. Server-side enforcement plus a transitive rule: Claude can't bypass by reading the file with the localReadtool and forwarding the content. - Caps cost — calls with estimated cost > $0.10 always go to ask-mode regardless of grant.
- Tells Claude what NOT to delegate — architecture, debugging, auth / payments, final commit messages, final user-facing answers.
Permission grant cheat-sheet
| Phrase | Effect |
|---|---|
| (default) | Ask before every mcp__interns__* call. |
| "разреши интернов" / "allow interns" | Free delegation until session end (subject to always-ask paths + cost cap). |
| "отзови интернов" / "revoke interns" | Back to ask-mode. |
| (session ends) | Reset to ask-mode. The next session asks again. |
Tool quick reference
| Tool | Required args | Optional | Purpose |
|---|---|---|---|
mcp__interns__bulk_text_read |
paths: list[str], question: str |
max_tokens: int |
Read N files, answer with file:line citations. |
mcp__interns__transcript_distill |
paths: list[str], question: str |
max_tokens: int |
Distill transcript / log into structured action list. |
Pass paths, not file contents. The MCP server reads and applies the always-ask matcher server-side; forwarding content as a string arg bypasses the gate.
When NOT to use
- Architecture / design decisions — Claude's job.
- Debugging — cheap models miss subtle bugs.
- Auth / payments / PII / production data — recommendation, not enforced, but obvious.
- Final commit messages, PR descriptions, user-facing answers — don't sub-delegate the deliverable.
- A single 50-line file you can read yourself for free.
Full policy and examples in SKILL.md.
Install
From the repo root:
bash scripts/install.sh using-interns
Works on Windows under git-bash, Linux, macOS.
See also
setup-interns— companion, owns the server install +mcpServers.internsregistration.project-discipline— Rule 4 (per-session push grant) is the prior art for the grant mechanism here.using-projects-meta,using-context7— sisterusing-Xskills for other MCP servers (same shape)..wiki/concepts/interns-design.md— full architecture spec.