pi-v4-jspace

DeepSeek V4 Minimal Anchor + J-Space runtime integration for Pi Coding Agent — one package, two capabilities:

1. Anchored Standard bootstrap (ported from pi-dsh-minimal): the first provider request is rewritten to the exact official DeepSeek Harness minimal surface (You are a helpful software engineer assistant. + persistent bash + str_replace_editor), then the first real tool call promotes back to the full Pi surface.
2. J-Space Cognition Suite V3.6 (vendored from upstream, unmodified): auto-activated once per compaction epoch via a steer message; J-Space itself runs its own fast / full / loop gate.

Non-DeepSeek-V4 models are completely untouched.

Install

pi install npm:pi-v4-jspace      # or
pi install git:github.com/<owner>/pi-v4-jspace

then /reload (or restart pi). No other installation steps: the pi-dsh-minimal adapter and the J-Space skill ship inside this package.

How it works

user prompt
   │
   ▼
Request #1 ── payload rewrite ──► system: "You are a helpful software engineer assistant."
   │                              tools: bash, str_replace_editor
   ▼
first real tool call
   │
   ▼
Promote ──► restore full Pi tools + project context + AGENTS.md + skills
   │
   ▼
steer: /skill:j-space Continue the current task from the existing conversation state.
   │    Do not restart work that is already complete.
   ▼
Request #2 ── full Pi surface + J-Space SKILL.md expansion
   │
   ▼
J-Space runs its own Gate (fast / full / loop) and continues the task

Key behaviors:

• Bootstrap is payload-level — before_provider_request rewrites the assembled provider payload; it is not setActiveTools(["bash"]). If the payload shape is unrecognized, the original payload is passed through (fail open).
• Promotion is tool-call based (v1.0 fixed): text-only first rounds stay anchored.
• One J-Space activation per compaction epoch — parallel tool calls activate once; compaction re-anchors and the next activation uses long-gap recovery semantics.
• Session state persists via custom entries: reload / resume / fork do not re-activate J-Space.
• Thinking level is set to max once on model select; if you lower it, the status bar warns and /v4j doctor reports it — no fight with the user.

Commands

Command	Purpose
/v4j	Summary status
/v4j status	Detailed status
/v4j on / /v4j off	Enable / fully disable (restores original tools; history stays)
/v4j reanchor	Reset the current runtime epoch (session content untouched)
/v4j doctor	Diagnostic checklist (incl. optional J-Space verify_suite.py)
/v4j dump on / /v4j dump off	Toggle provider-request debug dump

Status bar shows v4j anchored → v4j promoted • jspace pending → v4j promoted • jspace, plus v4j re-anchoring • resume after compaction and v4j degraded on activation failure.

Configuration

~/.pi/agent/pi-v4-jspace.json (all fields optional):

{
  "enabled": true,
  "modelPatterns": ["deepseek-v4-pro", "deepseek-v4-flash"],
  "thinking": "max",
  "setThinkingOnModelSelect": true,
  "promotion": "tool-call",
  "jspace": {
    "enabled": true,
    "activateAfterPromotion": true,
    "resumeAfterCompaction": true
  },
  "statusLine": true,
  "debugDump": false
}

promotion is "tool-call" only in v1.0; the field is kept for future compatibility.

Debug dump

Off by default. Enable with /v4j dump on (writes ~/.pi/agent/pi-v4-jspace-dump.jsonl) or set PI_V4_JSPACE_DUMP=/path/file.jsonl.

Debug dumps may contain sensitive project or conversation content.

Development

npm install
npm run typecheck   # tsc -p tsconfig.json
npm test            # vitest run

License

MIT. Vendored components: J-Space Cognition Suite V3.6 (Apache-2.0, unmodified, see skills/j-space/UPSTREAM-LICENSE) and pi-dsh-minimal (MIT). See THIRD_PARTY_NOTICES.md and vendor-meta/ for exact source commits.