feat(interns-skills-mvp): ship setup-interns + using-interns v0.1.0; project-bootstrap 1.5.0 -> 1.6.0

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>
This commit is contained in:
2026-05-05 17:41:11 +03:00
parent 7464ae4bc4
commit 0eb7dd1d7c
14 changed files with 632 additions and 8 deletions

View File

@@ -1,5 +1,5 @@
# Task Board
_Updated: 2026-05-01_
_Updated: 2026-05-05_
## 🟢 [shared-wiki-skill-catalog] — публикация каталога claude-skills в shared projects-wiki
**Status:** done
@@ -164,12 +164,12 @@ _Updated: 2026-05-01_
---
## [interns-skills-mvp] — Ship `setup-interns` + `using-interns` skills (v0.1.0) and bump `project-bootstrap` 1.5.0 → 1.6.0 with `delegate to interns when allowed` CLAUDE.md trigger, manifest entry, Step 5 commentary paragraph, idempotent merge into existing CLAUDE.md files. Two-skill pair following the setup-X / using-X prior art (setup-context7/using-context7, setup-projects-meta/using-projects-meta). Per-session permission grant policy mirroring project-discipline Rule 4 (ask-mode default, conversational grant/revoke, always-ask paths for secrets even with active grant, session-end reset).
## 🟢 [interns-skills-mvp] — Ship `setup-interns` + `using-interns` skills (v0.1.0) and bump `project-bootstrap` 1.5.0 → 1.6.0 with `delegate to interns when allowed` CLAUDE.md trigger, manifest entry, Step 5 commentary paragraph, idempotent merge into existing CLAUDE.md files. Two-skill pair following the setup-X / using-X prior art (setup-context7/using-context7, setup-projects-meta/using-projects-meta). Per-session permission grant policy mirroring project-discipline Rule 4 (ask-mode default, conversational grant/revoke, always-ask paths for secrets even with active grant, session-end reset).
**Status:** ready
**Where I stopped:** (not started)
**Next action:** Read design at `claude-skills/.wiki/concepts/interns-design.md` (just ingested) — Layer 3 section. Scaffold `skills/setup-interns/SKILL.md` and `skills/using-interns/SKILL.md` following setup-context7 / using-context7 structure. Then patch `skills/project-bootstrap/SKILL.md` Step 5 + `assets/CLAUDE.md.template` + `bootstrap-manifest.md`, bump to 1.6.0.
**Branch:** n/a
**Status:** done
**Where I stopped:** `skills/setup-interns/` v0.1.0 shipped (8-phase install: detect `.common/lib/interns-mcp/` source, `pip install -e`, write `.common/secrets/interns.env` with endpoint API keys + ensure `.gitignore` covers it, register `mcpServers.interns` in `~/.claude.json` with absolute Python interpreter path + `cwd` to project root); `skills/using-interns/` v0.1.0 shipped (runtime policy mirroring project-discipline Rule 4: ask-mode default, conversational grant/revoke, always-ask paths for `.env` / `**/secrets/**` / `**/*.key` / `**/.ssh/**` / `**/.aws/credentials` etc with transitive rule + cost-cap >$0.10, session-end reset; routing hints for `bulk_text_read` + `transcript_distill`); `project-bootstrap` 1.5.0→1.6.0 with canonical CLAUDE.md trigger `delegate to interns when allowed` inserted between `follow project discipline` and `we're on Windows`, new Step 5 commentary paragraph, Step 5.5 manifest table extended with `setup-interns` + `using-interns` + `project-discipline` rows; root `CLAUDE.md` dogfood line added; both READMEs written; descriptions verified within budget (setup-interns 899 / using-interns 814 / ≤900); both archives rebuilt + installed; harness listing confirms full descriptions (no H1 fallback) for all three skills
**Next action:** (none — kept until merged)
**Branch:** master
<!-- created-by: OpeItcLoc03@DESKTOP-NSEF0UK / from: meeting-room / 2026-05-05T13:59:07.540Z -->
---

View File

@@ -0,0 +1,38 @@
# interns-skills-mvp
## Goal
Ship the first version of the `setup-interns` + `using-interns` skill pair (v0.1.0 each) and roll the canonical `delegate to interns when allowed` CLAUDE.md trigger into `project-bootstrap` (1.5.0 → 1.6.0). The pair exposes the local `interns` MCP server (Layer 2 of `.wiki/concepts/interns-design.md`) to Claude Code under a per-session permission grant that mirrors `project-discipline` Rule 4 — ask-mode by default, conversational grant/revoke, always-ask paths for secrets even with an active grant, session-end reset. The runtime server itself is out of scope for this task; the skills land their full lifecycle (install procedure, runtime policy, bootstrap integration) so that on every machine where the MCP server is later installed, Claude already knows the policy.
## Key files
- `.wiki/concepts/interns-design.md` — Layer 3 Skills section is the primary spec
- `skills/setup-interns/SKILL.md` — to be created (mirror `skills/setup-projects-meta/SKILL.md` 8-phase shape)
- `skills/setup-interns/README.md` — to be created (mirror `skills/setup-projects-meta/README.md`)
- `skills/using-interns/SKILL.md` — to be created (mirror `skills/using-projects-meta/SKILL.md` two-class shape; grant mechanism mirrors `skills/project-discipline/SKILL.md` Rule 4)
- `skills/using-interns/README.md` — to be created
- `skills/project-bootstrap/SKILL.md` — Step 5 commentary + Step 5.5 manifest table; bump `version:` 1.5.0 → 1.6.0
- `skills/project-bootstrap/assets/CLAUDE.md.template` — append `delegate to interns when allowed` between `follow project discipline` and `we're on Windows`
- `CLAUDE.md` (root, dogfood) — same line addition
- `.wiki/log.md``## [2026-05-05] decision | interns-skills-mvp shipped` entry
- `dist/setup-interns.skill`, `dist/using-interns.skill`, `dist/project-bootstrap.skill` — rebuilt archives
## Decisions log
- 2026-05-05: ship as a pair (setup-X + using-X) following the validated prior-art pattern — same shape as setup-context7/using-context7 and setup-projects-meta/using-projects-meta. Rationale: clean separation of one-time install (mutates user-level config; needs confirmation gates) from runtime policy (read every session; no mutations). Cross-reference between the two via the `Prerequisites` section.
- 2026-05-05: setup-interns assumes inline `.common/lib/interns-mcp/` layout per design's "current склонение". Open question (Gitea repo + subtree) deferred — v0.1.0 documents the inline path; if a separate repo lands later, a 0.2.0 minor adds the alternate clone branch.
- 2026-05-05: grant mechanism mirrors project-discipline Rule 4 verbatim — ask-mode default per session, conversational grant ("разреши интернов" / "allow interns"), conversational revoke, always-ask paths even under an active grant, session-end automatic reset. Rationale: same security posture, same UX vocabulary the user already trained on.
- 2026-05-05: always-ask path list copied from design Layer 3 verbatim plus the transitive rule (Claude reading a matched file and forwarding content also triggers the gate). Plus cost-based ask: any call with estimated cost > $0.10 (sanity bound).
- 2026-05-05: `description:` budget — keep ≤ 900 chars per `feedback_skill_description_length_limit` memory. Verify char count before shipping; if over, trim trigger phrases first, never the policy summary.
## Open questions
- [ ] How does the user's actual `.common/` layout look on disk? Setup skill currently assumes `<project-root>/.common/lib/interns-mcp/` and `<project-root>/.common/secrets/interns.env`. If the convention is `~/projects/.common/`, the discovery phase needs to widen.
- [ ] Should `setup-interns` ship a scaffolding sub-step that writes a placeholder `.common/lib/interns-mcp/` from a template if the source isn't there? Probably not — kicks the can to a separate "build the runtime" task.
## Completed steps
- [x] Read design `.wiki/concepts/interns-design.md` Layer 3 section
- [x] Survey prior art: setup-context7, setup-projects-meta, using-context7, using-projects-meta, project-bootstrap@1.5.0, project-discipline (Rule 4 grant mechanism)
- [x] Mark 🔴 active in `.tasks/STATUS.md`, create this file
## Notes
- Char-count check after writing each `description:` — paste into a quick `wc -c` PowerShell snippet (or count manually). 1024 is the hard ceiling, 900 is the safety margin.
- Restart-Claude-Code caveat goes in setup-interns Phase 8, identical to setup-projects-meta and setup-context7.
- `using-interns` does NOT mention restart — it's a policy skill, no MCP install side.
- Bootstrap-manifest table gets `setup-interns` and `using-interns` rows.

View File

@@ -42,3 +42,4 @@ Parseable: `grep "^## \[" .wiki/log.md | tail -20`.
## [2026-05-01] ingest | shared-wiki packages/claude-skills — каталог всех 20 скиллов опубликован в projects-wiki (3 commits: page + index + log on Gitea, ae2cc9a..001cdd0); группировка bootstrap / wiki+tasks / MCP / caveman / discovery+platform; cross-link с concepts/setup-using-skill-pair и packages/projects-meta-mcp
## [2026-05-05] ingest | concepts/interns-design
## [2026-05-05] decision | interns-skills-mvp — shipped `setup-interns` v0.1.0 (8-phase install: detect `.common/lib/interns-mcp/`, `pip install -e`, `.common/secrets/interns.env` write, `mcpServers.interns` registration with absolute Python interpreter + `cwd`) and `using-interns` v0.1.0 (runtime policy mirroring project-discipline Rule 4: ask-mode default, conversational grant/revoke, always-ask paths for `.env`/secrets/keys/SSH/credentials with transitive rule, cost-cap >$0.10, session-end reset; routing hints for `bulk_text_read` + `transcript_distill`); `project-bootstrap` 1.5.0→1.6.0 with canonical CLAUDE.md trigger `delegate to interns when allowed` between `follow project discipline` and `we're on Windows`, Step 5 commentary paragraph, manifest table extended with both new skills + `project-discipline` row; root `CLAUDE.md` dogfood updated; both READMEs written; descriptions verified (setup-interns 899 chars, using-interns 814 chars, both under 900 budget); all three rebuilt + installed + listed by harness with full descriptions (no H1 fallback)

View File

@@ -8,4 +8,5 @@ use task management system
check across all projects
pull remote before work
follow project discipline
delegate to interns when allowed
we're on Windows

Binary file not shown.

BIN
dist/setup-interns.skill vendored Normal file

Binary file not shown.

BIN
dist/using-interns.skill vendored Normal file

Binary file not shown.

View File

@@ -57,7 +57,8 @@ projects bootstrapped at different times debuggable.
3. **Steps 15.** Create or skip each piece in order — git, README, `.wiki/`,
`.tasks/`, `CLAUDE.md`. Steps 3 and 4 delegate to the setup-skills.
4. **Step 5.5.** Write `bootstrap-manifest.md` recording the versions of
`project-bootstrap`, `setup-wiki`, and `setup-tasks` used.
`project-bootstrap`, `setup-wiki`, `setup-tasks`, `project-discipline`,
`setup-interns`, and `using-interns` used.
5. **Step 5.6.** Detect the official `superpowers@claude-plugins-official`
plugin via `~/.claude/plugins/installed_plugins.json`. If absent, print
a one-time chat recommendation with the install command and upstream
@@ -104,5 +105,10 @@ target with `CLAUDE_SKILLS_DIR=/path bash scripts/install.sh …`.
- [`setup-tasks`](../setup-tasks/) — companion, owns `.tasks/` layout.
- [`using-wiki`](../using-wiki/) — runtime policy for working with `.wiki/`.
- [`using-tasks`](../using-tasks/) — runtime policy for working with `.tasks/`.
- [`project-discipline`](../project-discipline/) — cross-project rules
activated by the `follow project discipline` trigger.
- [`setup-interns`](../setup-interns/), [`using-interns`](../using-interns/) —
pair behind the `delegate to interns when allowed` trigger; cheap-LLM
delegation under a per-session permission grant.
- Karpathy's LLM Wiki gist:
<https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f>

View File

@@ -1,6 +1,6 @@
---
name: project-bootstrap
version: 1.5.0
version: 1.6.0
description: >
Initializes or upgrades a project in the current folder: git, .gitignore, README.md,
.wiki/ using Karpathy's method, .tasks/ for task tracking, CLAUDE.md with skill triggers.
@@ -265,6 +265,7 @@ use task management system
check across all projects
pull remote before work
follow project discipline
delegate to interns when allowed
we're on Windows
```
@@ -291,6 +292,19 @@ push only after explicit per-session approval. Install the skill on the host
if `project-discipline` is not in `~/.claude/skills/`; otherwise the trigger is
silently dead like any other absent skill.
The `delegate to interns when allowed` line activates the `using-interns` skill,
which lets Claude offload predictable bulk I/O and summarization tasks
(reading 3+ files, distilling long transcripts) to cheap intern LLMs via the
local `interns` MCP server (`mcp__interns__bulk_text_read`,
`mcp__interns__transcript_distill`, etc.) — saves Anthropic quota at ~125× the
per-call cost reduction on bulk reads. Per-session permission grant mirrors
`project-discipline` Rule 4: ask-mode default, conversational grant / revoke,
always-ask paths for `.env` / secrets / keys / SSH credentials even with an
active grant, session-end reset. The skill is a no-op until the `interns` MCP
server is registered — install via `setup-interns` on a fresh machine if
`mcp__interns__*` tools are missing. Full design at
`.wiki/concepts/interns-design.md` in the `claude-skills` repo.
The `we're on Windows` line activates the `active-platform` skill and pins the
project's default platform to Windows / PowerShell — so generated commands and
README quick-starts use PS-native syntax. Bootstrapping on a Linux or macOS
@@ -322,6 +336,8 @@ Skills used to initialize this project's `.wiki/` and `.tasks/` layout, with the
| `setup-wiki` | <version> | wiki canonical layout |
| `setup-tasks` | <version> | tasks canonical layout |
| `project-discipline` | <version> | cross-project policy |
| `setup-interns` | <version> | interns MCP server install (one-time, per machine) |
| `using-interns` | <version> | interns runtime policy + per-session permission grant |
This file is overwritten if `project-bootstrap` is re-run on the same project. For history, use `git log .wiki/concepts/bootstrap-manifest.md`.
```

View File

@@ -8,4 +8,5 @@ use task management system
check across all projects
pull remote before work
follow project discipline
delegate to interns when allowed
we're on Windows

View File

@@ -0,0 +1,74 @@
# setup-interns
One-time skill that brings up the local `interns` MCP server on a new (or freshly broken) machine. Detects the runtime source at `<project-root>/.common/lib/interns-mcp/`, runs `pip install -e`, writes `.common/secrets/interns.env` with the endpoint API keys, and registers `mcpServers.interns` in `~/.claude.json`.
The runtime policy for *using* the resulting tools lives in [`using-interns`](../using-interns/) — `setup-interns` is the only place that touches user-level config or the project's secrets directory.
Full design (Layer 1 / 2 / 3, MVP catalog, always-ask paths, routing hints):
`.wiki/concepts/interns-design.md` (in this repo).
## When it triggers
- User says: "install interns", "set up interns", "configure interns",
"настрой интернов", "установи интернов", "interns не работает",
"interns isn't working".
- [`using-interns`](../using-interns/) detects missing `mcp__interns__*` tools and delegates here via its Prerequisites section.
- A new machine in the multi-machine fleet — install once per machine that wants delegation.
## What it touches
| Path | Role |
|---|---|
| `<project-root>/.common/lib/interns-mcp/` | server source (must already exist; not created here) |
| `<project-root>/.common/config/interns/config.yaml` | catalog of interns + endpoints (read-only here) |
| `<project-root>/.common/secrets/interns.env` | endpoint API keys (gitignored) |
| `~/.claude.json` (`mcpServers.interns`) | MCP registration |
| `pip` site-packages | editable install of `interns_mcp` |
## Hard rules
- **Never auto-mutate.** Phase 1 (discovery) and Phase 2 (plan) always pause for explicit confirmation. A trigger phrase grants permission to inspect, not to run `pip install` or write secrets.
- **Never echo endpoint API keys in chat.** Edit / Write tool calls inevitably contain them (that's how they land in `.env`); chat output must not.
- **Never run on absent source.** If `.common/lib/interns-mcp/pyproject.toml` is missing, stop and tell the user — initializing a fresh runtime is a separate task.
- **Always `chmod 600` `.env` on Linux / macOS.** Token leak otherwise.
- **Pin the absolute Python interpreter path** in `mcpServers.interns.command`. A bare `python` resolves to whatever shadows it on `PATH` at session start.
## Procedure (high-level)
1. **Phase 0** — environment sanity (`python` ≥ 3.11, `pip`, network to endpoints).
2. **Phase 1** — discovery (source / module / config / keys / MCP registration / `.gitignore`).
3. **Phase 2** — plan + confirm. Wait for explicit "ok" / "go" / "поехали".
4. **Phase 3** — backup (`~/.claude.json`, existing `.common/secrets/interns.env`).
5. **Phase 4**`pip install -e .common/lib/interns-mcp/`.
6. **Phase 5** — write `.common/secrets/interns.env` (per-key, missing-only) + ensure `.gitignore` rule.
7. **Phase 6** — register `mcpServers.interns` in `~/.claude.json` with absolute interpreter path + `cwd`.
8. **Phase 7** — best-effort smoke test (in-session caveat: real verification is after restart).
9. **Phase 8** — restart guidance + final report.
Full procedure with shell snippets and templates lives in [`SKILL.md`](SKILL.md).
## Rollback
1. Stop. Don't fix forward.
2. `cp <file>.bak-<ts> <file>` for `~/.claude.json` and `.common/secrets/interns.env`.
3. Optional: `pip uninstall interns-mcp` to undo the editable install.
4. Restart Claude Code.
5. Confirm `mcp__interns__*` tools are gone (or back to the prior version).
## Install
From the repo root:
```bash
bash scripts/install.sh setup-interns
```
Works on Windows under git-bash, Linux, macOS.
## See also
- [`using-interns`](../using-interns/) — runtime policy + per-session permission grant for the `mcp__interns__*` tools.
- [`setup-projects-meta`](../setup-projects-meta/) — companion pattern, similar 8-phase shape.
- [`setup-context7`](../setup-context7/) — sister setup-skill for the context7 MCP plugin.
- [`project-discipline`](../project-discipline/) — Rule 4 (per-session push grant) is the prior art for the grant mechanism in `using-interns`.
- `.wiki/concepts/interns-design.md` — full architecture spec.

View File

@@ -0,0 +1,228 @@
---
name: setup-interns
version: 0.1.0
description: Installs and configures the local `interns` MCP server — a FastMCP stdio app at `<project-root>/.common/lib/interns-mcp/` that delegates bulk reads, transcript distillation, and other predictable I/O to cheap intern LLMs (DeepSeek / Kimi / Ollama) so Claude saves Anthropic quota. Procedure: detect the server source, `pip install -e` it, write `.common/secrets/interns.env` with the endpoint API keys, register `mcpServers.interns` in `~/.claude.json`. Use this skill when the user says "install interns", "set up interns", "configure interns", "настрой интернов", "установи интернов", "interns не работает", "interns isn't working", or whenever the `mcp__interns__*` tools are missing in a session that needs delegation. Cross-platform — Windows / Linux / macOS. Mutates user-level config and writes secrets; pauses for confirmation before every write.
---
# setup-interns
> One-time skill that gets the local `interns` MCP server running with the user's endpoint API keys and a registered `mcpServers.interns` entry in `~/.claude.json`. Stops at confirmation gates because the procedure runs `pip install`, writes a secret-bearing `.env` file, and edits user-level config.
Reference: full design lives in this repo at `.wiki/concepts/interns-design.md` (Layer 1/2/3 architecture, MVP catalog, always-ask paths, routing hints).
## When to use
- User explicitly asks: install / set up / configure interns.
- A `using-interns`-driven task fails because `mcp__interns__*` tools aren't available in the session.
- New machine in the user's multi-machine fleet — install once per machine that wants delegation.
- Migrating a stale install (endpoint key rotated, server source moved, broken `dist/` artifact) — same procedure, Phase 1 detects what's already in place.
## Out of scope
- Building or scaffolding the `.common/lib/interns-mcp/` source tree itself. The skill expects the source already in place per the inline `.common/` convention from the design (or a future Gitea repo when that branch lands). If the source is absent, Phase 1 stops with a clear message — initializing a fresh runtime is a separate task.
- Issuing or rotating endpoint API keys (Ollama Cloud, OpenRouter, etc.). The skill *uses* keys the user already has; if there is none, it points at the provider's settings page and stops.
- Running `interns-mcp` itself — the Claude Code harness spawns it on session start.
- Authoring new interns or editing `.common/config/interns/config.yaml` — that's a content task, not a setup task.
- Any other MCP server.
## Hard rule: don't auto-mutate config
The procedure runs `pip install`, writes `.common/secrets/interns.env` (carries endpoint API keys), and edits `~/.claude.json`. **Always pause for explicit confirmation between Phase 1 (discovery, read-only) and Phase 2 (plan), and again before Phase 3 (backup + writes).** A trigger phrase grants permission to inspect, not to install or write secrets.
## Procedure
### Phase 0 — Environment sanity
- Confirm Claude Code is the current harness (need `mcpServers` registration in `~/.claude.json`).
- Confirm `python` ≥ 3.11 and `pip` are on `PATH`. Report the resolved interpreter path (`python -c "import sys; print(sys.executable)"`); the skill will pin this exact path in the MCP registration so a later `python` shadowed by another env doesn't silently take over.
- Confirm network reachability to the configured endpoints (default: `https://ollama.com/v1`). On HTTP 401 / 403 later, the API key is dead — stop and ask for a new one.
- Pick paths: `<project-root>/.common/lib/interns-mcp/` (source), `<project-root>/.common/config/interns/config.yaml` (catalog), `<project-root>/.common/secrets/interns.env` (keys, gitignored), `~/.claude.json` (MCP registration). POSIX-style paths resolve correctly under git-bash on Windows.
### Phase 1 — Discovery (read-only)
Search, in order. Report only "found at <path>", never echo key values.
**Server source.** Check whether `<project-root>/.common/lib/interns-mcp/pyproject.toml` exists. If absent, **stop** — the runtime source must be in place before this skill runs. Report:
```
.common/lib/interns-mcp/ not found.
This skill expects the inline interns-mcp source per the
.common/ layout (see .wiki/concepts/interns-design.md). Initialize
the runtime first, then re-run setup-interns.
```
**Build artifact.** Run `python -c "import interns_mcp" 2>&1` against the candidate interpreter. If it fails with `ModuleNotFoundError`, Phase 4 will run `pip install -e .common/lib/interns-mcp/`. If it succeeds, capture the installed location and skip the install in Phase 4.
**Config catalog.** Read `<project-root>/.common/config/interns/config.yaml`. Extract the unique set of `endpoints.<name>.api_key_env` values — these are the env var names the runtime expects to find. Capture for Phase 2.
**Existing endpoint keys.** Look in priority order, per `api_key_env` name from the config:
1. `<project-root>/.common/secrets/interns.env` (`<NAME>=...` lines).
2. Process env (`os.environ[<NAME>]`).
3. `~/.config/projects-mcp/auth.toml` — only if the user has explicitly noted the key is shared with another local MCP server (rare).
The first hit wins per key. **Never echo key values in chat.**
**MCP registration.** Read `~/.claude.json` and check `mcpServers.interns`. Note the `command` and `args`. If args point at a stale interpreter, Phase 6 will fix it.
**Gitignore sanity.** Check `.gitignore` (project root). If `.common/secrets/` (or `.common/secrets/*.env`) is not listed, flag for Phase 2 — the skill will offer to add it before writing the file.
### Phase 2 — Plan + confirm
Present a single-block plan to the user:
```
Source: <found at .common/lib/interns-mcp/ | NOT FOUND — STOP>
Module: <interns_mcp importable | will pip install -e>
Config: <found at .common/config/interns/config.yaml> — endpoints: <names>
Missing keys: <list of <NAME> not yet present in interns.env | none>
Gitignore: <covers .common/secrets/ | will add ".common/secrets/*.env">
MCP entry: <present in ~/.claude.json | will add | will fix interpreter path>
Backups: ~/.claude.json.bak-<ts>, .common/secrets/interns.env.bak-<ts> (if exists)
```
Wait for explicit confirmation ("ok", "go", "поехали"). Anything else → stop.
For each missing key, ask: "Paste a value for `<NAME>` (endpoint: `<endpoint_name>` — provider settings page: `<url-from-config-or-known-list>`), or skip and we'll leave the entry blank for you to fill later." Don't proceed past Phase 2 without resolving every required key — the runtime won't start with a missing API key.
### Phase 3 — Backup
Copy each file we will modify to `<file>.bak-YYYYMMDD-HHMMSS`:
```bash
TS=$(date +%Y%m%d-%H%M%S)
[ -f ~/.claude.json ] && cp ~/.claude.json ~/.claude.json.bak-$TS
[ -f .common/secrets/interns.env ] && cp .common/secrets/interns.env .common/secrets/interns.env.bak-$TS
```
Confirm both backups exist (when their source existed) before any further edit.
### Phase 4 — Install Python module
```bash
# from project root
python -m pip install -e .common/lib/interns-mcp/
```
Capture the resolved `python` from Phase 0; use the same interpreter for both `pip install` and the later MCP `command:` field. Verify post-install:
```bash
python -c "import interns_mcp; print(interns_mcp.__file__)"
```
If import fails — abort. Ask the user to paste the `pip install` output so the failure mode is visible.
### Phase 5 — Write `interns.env` + gitignore
If Phase 1 flagged a missing `.gitignore` rule, append it first:
```
# interns endpoint API keys
.common/secrets/*.env
```
Then write `.common/secrets/interns.env`. Per-key behavior:
- Existing key in the file with a non-empty value — leave it alone.
- Missing key — append `<NAME>=<value-from-Phase-2>` if the user pasted one, or `<NAME>=` (blank) if the user skipped. A blank entry will fail at runtime with a clear `KeyError`; that's acceptable for the "I'll fill it later" path.
Permissions: on Linux / macOS run `chmod 600 .common/secrets/interns.env`. On Windows the default ACL is per-user, no extra step.
### Phase 6 — Register in `~/.claude.json`
Edit `~/.claude.json`. Add or update the `mcpServers.interns` block:
```json
{
"mcpServers": {
"interns": {
"command": "<ABSOLUTE_PATH_TO_PYTHON>",
"args": ["-m", "interns_mcp.server"],
"cwd": "<ABSOLUTE_PATH_TO_PROJECT_ROOT>"
}
}
}
```
`cwd` is set so the runtime resolves `.common/config/interns/config.yaml` and `.common/secrets/interns.env` relative to project root regardless of where Claude Code was launched.
Absolute interpreter path comes from Phase 0 (`sys.executable`). Forward slashes work in JSON on Windows without escaping.
After each edit, validate JSON:
```bash
# Windows (git-bash)
powershell.exe -NoProfile -c "Get-Content '<file>' -Raw | ConvertFrom-Json | Out-Null"
# Linux / macOS
jq empty <file>
# fallback
python -c "import json; json.load(open('<file>'))"
```
If validation fails → restore from `.bak-*` and abort.
### Phase 7 — Smoke test (best-effort)
Best-effort: ask the user to call `mcp__interns__bulk_text_read` against a tiny benign input (e.g. read this file's `Phase 7` section, ask "what is this section about?"). A response that is structurally valid (text + usage) means the server is reachable.
**Important caveat to relay to the user:** in the *same* session that just ran setup, the active MCP connection was bound at session start. So a passing in-session smoke test only proves "an interns server is alive" — not "the registration we just wrote is what's serving it". The real test is after Claude Code restart.
If `mcp__interns__*` tools aren't registered in this session at all, skip the smoke test and rely on Phase 8.
### Phase 8 — Restart guidance + final report
Tell the user:
```
✅ Setup complete. Restart Claude Code so the new mcpServers.interns
registration binds to a fresh stdio session.
After restart:
• mcp__interns__* tools serve from <python> -m interns_mcp.server
• Endpoint keys live in .common/secrets/interns.env (gitignored)
• Config catalog at .common/config/interns/config.yaml
• Backups saved at ~/.claude.json.bak-<ts> (and interns.env.bak-<ts>
if it existed before)
Runtime policy lives in `using-interns`. CLAUDE.md trigger:
delegate to interns when allowed
project-bootstrap 1.6.0+ adds it to new projects automatically.
If something breaks after restart:
• Restore from .bak-* and tell me — we'll roll back together.
```
## Rollback procedure
If a problem surfaces (now or after restart):
1. Stop. Don't try to fix forward.
2. Find the most recent `.bak-YYYYMMDD-HHMMSS` next to `~/.claude.json` and `.common/secrets/interns.env`.
3. `cp <file>.bak-<ts> <file>` for each.
4. Optional: `pip uninstall interns-mcp` if you want to remove the editable install.
5. Restart Claude Code.
6. Confirm `mcp__interns__*` is gone (or back to its pre-existing version).
7. Report what went wrong so we can fix the procedure.
## Cross-platform notes
The procedure is platform-agnostic. Only auxiliary tooling differs:
| | JSON validate | Backup | Permissions on `.env` |
|---|---|---|---|
| Windows (git-bash) | `powershell.exe -NoProfile -c "Get-Content '<f>' -Raw \| ConvertFrom-Json \| Out-Null"` | `cp` | per-user ACL by default |
| Linux | `jq empty <f>` (or `python -c "import json; json.load(open('<f>'))"`) | `cp` | `chmod 600` |
| macOS | same as Linux | `cp` | `chmod 600` |
POSIX-style paths (`.common/...`, `~/.claude.json`) work on all three.
## Common mistakes
- **Skipping Phase 1.** "User just said 'install interns' — let's go." No — find existing source / module / keys first; clobbering an existing `.env` over a working one loses keys you can't recover.
- **Echoing endpoint keys.** They're secrets. Edit / Write tool calls inevitably contain them (that's how they land in `.env`), but no chat output should.
- **Pinning `python` instead of `<sys.executable>`.** A bare `python` in the MCP `command:` resolves to whatever interpreter is first on `PATH` at session start — often a different env without the `interns_mcp` module. Always use the absolute interpreter path captured in Phase 0.
- **Forgetting `cwd:`.** Without it the runtime can't find `.common/config/interns/config.yaml` and bombs at startup with a config-not-found error that looks like a Claude Code bug.
- **Writing `.env` with `0644` perms on Linux/macOS.** Token leak. Always `chmod 600` after write.
- **Missing the gitignore rule.** Tokens commit to the repo on the next `git add .`. Always check `.gitignore` covers `.common/secrets/*.env` before writing — Phase 5 does it but it's worth double-checking.
- **Treating in-session smoke test as proof.** Same as the context7 / projects-meta caveat — the active MCP connection was bound at session start. Real verification happens after restart.
- **Auto-running on every "use interns".** This skill is intrusive. Trigger only on explicit "install / set up / configure interns", or when MCP tools are missing and the user is blocked.

View File

@@ -0,0 +1,68 @@
# 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`](../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 by `project-bootstrap` 1.6.0+).
- 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.md` or write a session summary from a long transcript.
- If `mcp__interns__*` tools are missing, this skill delegates to [`setup-interns`](../setup-interns/) before doing anything else.
## What it does
- **Picks the right intern** for the task from a routing table (`bulk_text_read` for reads, `transcript_distill` for summaries; future interns added in `.common/config/interns/config.yaml`).
- **Guards every call** with a per-session permission grant that mirrors `project-discipline` Rule 4 — ask-mode default, conversational grant / revoke, session-end reset.
- **Enforces always-ask paths** — `.env`, `**/secrets/**`, `**/*.key`, `**/*.pem`, `**/.ssh/**`, `**/.aws/credentials`, `**/.netrc` etc. — even when delegation is granted. Server-side enforcement plus a transitive rule: Claude can't bypass by reading the file with the local `Read` tool 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`](SKILL.md).
## Install
From the repo root:
```bash
bash scripts/install.sh using-interns
```
Works on Windows under git-bash, Linux, macOS.
## See also
- [`setup-interns`](../setup-interns/) — companion, owns the server install + `mcpServers.interns` registration.
- [`project-discipline`](../project-discipline/) — Rule 4 (per-session push grant) is the prior art for the grant mechanism here.
- [`using-projects-meta`](../using-projects-meta/), [`using-context7`](../using-context7/) — sister `using-X` skills for other MCP servers (same shape).
- `.wiki/concepts/interns-design.md` — full architecture spec.

View File

@@ -0,0 +1,191 @@
---
name: using-interns
version: 0.1.0
description: Use when delegating predictable bulk reads or summarization to cheap intern LLMs via the local `interns` MCP server (`mcp__interns__bulk_text_read`, `mcp__interns__transcript_distill`, etc.) so Claude saves Anthropic quota. Activated by `delegate to interns when allowed` in CLAUDE.md or explicit phrases like "use interns", "delegate to an intern", "разреши интернов", "allow interns". Per-session permission grant mirrors `project-discipline` Rule 4: ask-mode default, conversational grant / revoke, always-ask paths for `.env` / secrets / keys / SSH / credentials even with active grant, transitive rule (Claude can't bypass by reading the file itself and forwarding content), session-end reset. Skip for architecture, debugging, auth / payments, final commit messages, or final user-facing text.
---
# Using the `interns` MCP server
## Overview
`interns` is a local stdio MCP server that exposes a catalog of cheap-LLM "interns" — DeepSeek, Kimi, Ollama-hosted models, etc. — so Claude can delegate predictable bulk I/O and summarization tasks instead of paying Anthropic quota for them. The pattern (~23× cheaper end-to-end on summarization, ~125× per-call on bulk reads) is sourced from a Reddit thread + Medium follow-up; see `.wiki/concepts/interns-design.md` for the full design and the cost numbers.
Two MVP interns:
| Tool | What it does | When to call |
|---|---|---|
| `mcp__interns__bulk_text_read` | Reads N files end-to-end and answers a focused question with file:line citations. | Claude was about to read 3+ files or one file > 400 lines just to extract context. |
| `mcp__interns__transcript_distill` | Compresses a session transcript / log into a structured action list (decisions / open questions / next steps). | Before updating `.wiki/log.md` or producing a session summary. |
Both currently run on Ollama Cloud (`deepseek-v4-flash`, ~$0.002 / call). Adding a new intern is a config-only change — see `.wiki/concepts/interns-design.md` § "Как добавить нового интерна".
## Prerequisites
This skill assumes `mcp__interns__*` tools are available. If they aren't (tools missing from the session, or calls fail with a connection error), the server isn't running for this session. Trigger the **`setup-interns`** skill to `pip install -e` the runtime, write `.common/secrets/interns.env`, and register `mcpServers.interns` in `~/.claude.json`. It's a one-time procedure with confirmation gates.
## When to use
- A read that's **bulky and not central to editing**: scanning long config files, log dumps, transcript files, generated docs — anything where Claude would burn input tokens summarizing for itself.
- A read across **3+ files** to assemble context.
- A read of **one file > 400 lines** that won't be edited.
- **Transcript / session distillation** before writing a wiki log entry, summary, or post-mortem.
- Any future intern documented in `.common/config/interns/config.yaml` (PDF extraction via Marker, repo packaging via Repomix, JS-rendered web fetches via Firecrawl, etc.) — same trigger / grant rules apply.
## When NOT to use
- **Architecture / design decisions.** Cheap models miss subtle trade-offs.
- **Debugging.** Cheap models hallucinate in ways that turn a 5-minute fix into a 1-hour wild goose chase.
- **Auth / payments / PII / production data / deletion.** Recommendation, not enforced — but obvious.
- **Final commit messages, PR descriptions, user-facing answers.** The task explicitly delegated *to* Claude. Don't sub-delegate the deliverable.
- **Anything that fits into Claude's own working context cheaply.** A 50-line file to skim is not a delegation candidate; it's just a Read.
## Per-session permission grant (mirrors `project-discipline` Rule 4)
The grant mechanism is identical in shape to the push-grant in `project-discipline`. Reuse the same vocabulary so the user doesn't have to learn two policies.
### Default — ask-mode
Every session starts in ask-mode. Before the first `mcp__interns__*` call, ask:
> Я бы делегировал чтение `<files>` интерну `bulk_text_read` (DeepSeek Flash via Ollama Cloud, ~$0.002 за вызов). Ок?
(or its English equivalent if the user is communicating in English) and wait for explicit `yes` / `да` / `go` / `поехали`. Without confirmation — don't call.
### Conversational grant
When the user says any of:
- "разреши интернов" / "разреши интерна" / "разреши делегировать"
- "allow interns" / "use interns freely" / "go ahead with interns"
— grant the right to call `mcp__interns__*` without further confirmation **for the rest of the session**, subject to always-ask paths below. Acknowledge the grant in one short line ("ок, делегирую интернам без подтверждения, кроме secrets-путей" / "got it, free delegation except secrets").
### Conversational revoke
When the user says any of:
- "отзови интернов" / "снова спрашивай" / "делай сам"
- "revoke interns" / "stop using interns" / "ask me again every time"
— return to ask-mode for the rest of the session. Acknowledge in one short line.
### Session-end reset
The next session starts in ask-mode regardless of prior grants. Same reasoning as `project-discipline` Rule 4: a grant is given for the current context (user nearby, attention available), and should not survive a context switch.
### Always-ask paths (even with active grant)
These never go to an intern silently, even when delegation is granted:
- `**/.env`, `**/.env.*`
- `**/secrets/**` (including `.common/secrets/`, `~/.config/projects-mcp/`)
- `**/credentials*`, `**/credentials.json`
- `**/*.key`, `**/*.pem`
- `**/.ssh/**`
- `**/.aws/credentials`, `**/.aws/config`
- `**/.netrc`, `**/.npmrc`, `**/.pypirc`
The `interns` MCP server enforces this server-side: matched paths return `BlockedByPolicy{path, pattern, reason}` instead of going out to the endpoint. Claude's job: **don't try to bypass** by reading the file with the local `Read` tool and forwarding the content as a string argument. The transitive rule applies — once a file matched, its content stays out of every intern call for the rest of the session unless the user explicitly grants per-path.
When a path is blocked, ask the user explicitly:
> Файл `<path>` matched always-ask pattern `<pattern>` (reason: `<reason>`). Передавать интерну на endpoint `<endpoint>` несмотря на это?
Wait for an explicit per-path "yes". Don't generalize the answer to other paths in the same call.
### Cost-based always-ask
Any call with estimated cost > $0.10 (per the `tokens × price_per_M` config in `.common/config/interns/config.yaml`) goes to ask-mode regardless of grant state. Sanity bound — catches a misrouted huge-payload request before it bills.
## Routing hints (when to pick which intern)
| Situation | Intern |
|---|---|
| File > 400 lines, not editing it | `bulk_text_read` |
| Need to read 3+ files for context | `bulk_text_read` |
| Updating `.wiki/log.md` or session-summary doc | `transcript_distill` |
| Compressing a long brainstorm transcript before quoting in a design doc | `transcript_distill` |
| (Future interns in `config.yaml`) | per the description in the catalog |
## Workflow
```
1. Identify the candidate task (read-heavy, bulk, predictable).
2. Pick the intern from the routing table.
3. Check grant state:
• ask-mode → ask before calling
• granted → check always-ask paths; if any matched, ask per-path
4. Call the intern with `paths` (list of file paths — the server reads them)
and `question` (focused, specific).
5. Cite the response (file:line refs come from the intern's reply).
```
Don't pass file *contents* as arguments — only paths. The MCP server reads and applies the always-ask matcher before any path content reaches an endpoint. If you forward content yourself, the gate is bypassed and the user can't trust the policy.
## Tool quick reference
| Tool | Required args | Optional | Purpose |
|---|---|---|---|
| `mcp__interns__bulk_text_read` | `paths: list[str]`, `question: str` | `max_tokens: int` | Read N files end-to-end, answer the question with citations. |
| `mcp__interns__transcript_distill` | `paths: list[str]`, `question: str` | `max_tokens: int` | Distill transcript / log into structured action list. |
(Future interns expose tools at `mcp__interns__<intern_id>` as the catalog grows.)
## Examples
### Read-bulk delegation
User: "summarize what these four config files do for the build pipeline."
```
1. Routing → bulk_text_read (4 files, context only).
2. Grant state → ask-mode (session start).
3. Ask: "Bulk-read these 4 files via bulk_text_read (~$0.002)? Ок?"
4. User: "ok"
5. Call mcp__interns__bulk_text_read
paths: ["build/config.ts", "build/plugins.ts", "build/env.ts", "build/output.ts"]
question: "What does each file contribute to the build pipeline? Cite file:line."
6. Cite the response. Done.
```
### Always-ask path triggered
User: "scan all `.env*` files in the repo and tell me which keys are duplicated."
```
1. Routing → bulk_text_read.
2. Grant state → granted (user said "allow interns" earlier).
3. Path matcher → every input matches `**/.env`. Server returns BlockedByPolicy.
4. Ask user: "These are .env files (always-ask). Pass them to bulk_text_read despite the secrets exposure?"
5. Wait for explicit per-path yes/no.
```
### Transcript distill
User: "compress this brainstorm transcript into action items before I ingest it into the wiki."
```
1. Routing → transcript_distill.
2. Call mcp__interns__transcript_distill
paths: [".meeting-room/.brainstorm/foo.md"]
question: "Extract decisions, open questions, and next steps. Output structured markdown."
3. Quote the result in the wiki ingest commit.
```
## Common mistakes
| Mistake | Fix |
|---|---|
| Calling an intern in ask-mode without asking first | Ask before the first call every session, even for trivial reads. |
| Forwarding file *content* instead of *path* | Pass paths. The server reads. The matcher only sees paths. |
| Trying to bypass always-ask by reading the file with `Read` and pasting content | Transitive rule — that path stays out of intern calls for the session unless user says otherwise per-path. |
| Sub-delegating final deliverables (commit messages, user-facing answers) | Those are Claude's job. Interns are for predictable bulk I/O, not for the answer that ships. |
| Using an intern for debugging | Cheap models miss subtle bugs. Pay the Anthropic call for debugging — the savings aren't worth a wrong fix. |
| Persistent grant across sessions | The next session resets to ask-mode. By design — same as project-discipline Rule 4. |
## Red flags
- "I'll just paste the content as a string arg" → no. The matcher only sees paths. If you're tempted to bypass it, you're about to leak a secret.
- "This grant should survive the session" → no. The next session asks again. The grant is bound to the current context where the user is paying attention.
- "It's just a small file, why ask?" → small files don't need an intern. Read them yourself. The ask is *because* the call goes to a third-party endpoint, not because of size.
- "Estimated cost is fine, $0.08 < $0.10" → still ask if the call is novel or the file set is unfamiliar. The cost gate is a hard ceiling, not the only signal.