Keystone OS Skills & Capabilities

Convergence Loop (Σ₀)

The entire system is one loop: Observe → Remember → Reason → Act → Verify → Converge

Every skill must strengthen one stage. Nothing outside this loop is in scope.

AGI Benchmark (as of 2026-06-16)

Scores updated per-run in data/agi-benchmark.jsonl.

Live Skills (real implementations)

Audit (2026-06-29):design-only relic skills were deleted and the design-only ones banner-flagged in a Σ₀ scope-discipline pass — see docs/SKILLS-AUDIT-2026-06-29.md. The convergence skill (below) was added to give the Converge loop stage a real contract.

convergence

The Converge stage as a skill — grounded synthesis + Convergence Records. Backs the !convergance chat command.

• Synthesizes recent entries into ONE insight; grounds forward-looking claims via the research skill's task loop (1-2 bounded rounds of fan-out + gap-driven refinement — falls back to a single web search on error)
• Appends evidence-bearing records to data/convergence/records.jsonl (grounded, sources, grounding_task_id, honest confidence)
• !convergance <topic> grounds on an explicit topic; !convergance log an issue <title> files a GitHub issue (shell-free)

research (Σ₀ grounded)

Persisted, resumable long-running research tasks — the Remember + Verify stages made durable across chat turns.

• !research <topic> (or plain language: "research X" / "look into X" / "investigate X") starts a TASK, not a single search
• Each round runs the wide-search Observe→Reason→Verify→Converge loop (fan-out sub-queries, low-fidelity prune, high-fidelity cited synthesis), then a gap-check decides whether another round is warranted
• State persists to data/research-tasks/<id>.json after every round — survives server restarts; !research continue <id> resumes an unfinished task from where it left off
• Bounded by RESEARCH_TASK_MAX_ROUNDS (defaulttotal) and RESEARCH_ROUNDS_PER_TURN (defaultper HTTP turn)
• On completion, emits a Convergence Record (reasoner: "research-task") and a CSF memory entry
• Backs both !convergance's grounding (1-2 bounded rounds) and autowork's issue research (AUTOWORK_RESEARCH_ROUNDS, default 2) — one engine, three entry points
• Implementation: apps/lantern-garage/lib/research-task.js (task state) + lib/wide-search.js (per-round search loop)

dream_journal

Dream Journal entry creation, management, and RAG-backed search.

• Create dream entries with metadata (emotions, tags, symbols, lucidity)
• Search/filter across dream history
• Export to CSV/JSONL format
• CSF compression for efficient storage

lucid_dreaming

Lucid dreaming coaching and reflection tools.

• Technique suggestions (WILD, DILD, WBTB)
• Dream sign tracking
• Reality checks and sleep window planning
• Integration with dream journal for pattern analysis

archive_curator

Documentation, archival, and knowledge management.

• Markdown rendering and repo file serving
• RAG house building (flat document index)
• Knowledge base search and retrieval
• CSF/CADD memory exports

voice_curator

Text-to-speech and audio generation via ElevenLabs/OpenAI.

• Voice selection and model control
• Streaming audio output
• Provider fallback (ElevenLabs → OpenAI)
• Caching and rate-limit handling

autonomous_work (Σ₀ grounded)

Fully autonomous issue resolution via /api/convergence/autonomous-work/stream.

• Observe: fetch GitHub issue + extract keywords
• Research: grep codebase + the research skill's task loop (up to AUTOWORK_RESEARCH_ROUNDS rounds, default 2, targeting each round's gaps)
• Reason: Claude generates JSON plan via extractJson (4-strategy fallback)
• Act: apply unified diff patch to correct auto/issue-N branch
• Verify: run allowlisted test commands
• Converge: log evidence + confidence to data/convergence-autonomous-work.jsonl
• Commit + push + open draft PR (graceful "already exists" handling)
• Confidence: codebase 0.85 · web 0.80 · tests 0.90 · observable 1.0

provider_management

Live API key management without server restart.

• POST /api/providers/set-key — write key to .env.local, hot-patch process.env
• DELETE /api/providers/set-key — remove key from .env.local
• POST /api/providers/test/:provider — ping provider API to verify key
• GET /api/providers/status — which providers have keys configured
• Supports: anthropic · gemini · openai · xai

Agent Personas

Selection is automatic; override with ?agent=NAME.

Provider Chain

Cascade order for LLM calls (auto mode):

Configure via /api-keys-settings.html — keys persist to .env.local with hot-patch.

Fleet Integration

Autonomous Test Fleet (Keystone)

Trigger in dream-chat: "test the app" / "scan for issues" / "audit the system"

Scenarios: home-load · dream-chat-init · dream-chat-first-message · theme-toggle · dream-chat-agent-select · dream-chat-error-handling · home-nav-links · trader-dashboard-load · responsive-mobile · responsive-tablet · console-monitoring · network-monitoring · slow-network

Confidence gate: ≥0.8 → file immediately · 0.5–0.79 → needs-review · <0.5 → log only

MCP Server

src/mcp_server/server.py — FastAPI + SSE on port 8771: queue_status · task_intake · dispatch_work · boot_check · list_skills · get_status

Discord Bot

Optional — set DISCORD_BOT_TOKEN + LANTERN_DISCORD_GUILD_ID in .env.local.

Open Issues — Fleet Growth Queue

Run !convergence in dream-chat or Auto-work #N in the convergence panel.