@burnmydays/signaf

signaf — token-cascade coach

Interactive token-cascade agent. Reads your AI coding session logs locally, computes the yield cascade (Υ, SNR, Leverage, Velocity), builds a behavioral taste profile, measures your Appropriate Steering Index (ASI), and coaches you on token efficiency.

Everything stays local. Nothing leaves your machine.

→ signalaf.com

The SigRank ecosystem

signaf is one of three pieces:

Repo	What it is	Install
sigrank-mcp	The instrument — extracts 4 token pillars, computes the cascade, submits to the leaderboard. MCP server + TUI dashboard.	npx sigrank
sigrank-app	The leaderboard — signalaf.com. Privacy-preserving operator profiles, class tiers, board rankings.	signalaf.com
signaf (this repo)	The coach — reads all 3 signal layers from your logs, builds a taste profile, measures ASI, coaches you on what your tokens were worth.	npx signaf

sigrank-mcp is the calorie counter. signaf is the metabolic panel.

Table of Contents

• The SigRank ecosystem
• What this is
• Install
• Quick start
• The REPL
• MCP server mode
• Skills
• Appropriate Steering Index (ASI)
• The taste profile
• The taste → cascade bridge
• Privacy
• File layout
• Architecture
• Contributing
• License

What this is

signaf reads the same session logs that sigrank-mcp's tokenpull reads — but instead of just extracting four token pillars, it reads all three signal layers:

• Layer 1 (metadata): tool distribution, file edit counts, edit sizes, reject/error rates
• Layer 2 (structural): correction loops, convergence patterns, session shape
• Layer 3 (content): user feedback directives — distilled into preferences, not retained (opt-in)

From these layers it computes:

• The cascade (Υ, SNR, Leverage, Velocity, class) — the efficiency metrics
• Appropriate Steering Index (ASI) — 8 dimensions measuring whether your interventions were the RIGHT ones, not just how often you accepted
• A behavioral taste profile — 5 dimensions: steering signature, iteration fingerprint, workflow rhythm, cascade personality, correction taxonomy

Then it coaches you: diagnose weak pillars, simulate changes, suggest improvements, track trends, set goals, analyze costs, detect anomalies, and run the full self-improvement cycle.

Install

npx signaf                 # no install needed
# or install globally:
npm install -g signaf
signaf --help

Requires Node.js ≥ 18.

Quick start

# 1. Start the interactive REPL (auto-scans on first run)
signaf

# 2. Or run one-shot commands
signaf scan                 # Read logs, compute everything, save
signaf diagnose             # "How am I doing?"
signaf simulate input -50%  # "What if I cut my input in half?"
signaf suggest              # "What should I do differently?"
signaf taste                # "What's my taste profile?"
signaf asi                  # "How well do I steer?"
signaf bridge               # "Connect my behavior to my cascade"
signaf goal transmitter     # "How do I hit TRANSMITTER?"
signaf cost                 # "How much did I spend?"
signaf compare transmitter  # "How do I compare to TRANSMITTER avg?"
signaf self-improve         # Full cycle: diagnose → suggest → actions
signaf watch                # Background daemon (auto-scan on log changes)

# 3. Or expose signaf as MCP tools for your AI agent
signaf --mcp                # starts stdio MCP server (12 tools)

The REPL

$ signa

signaf — interactive token-cascade agent. Type "help" for commands.

signa> how am I doing today?
═══ DIAGNOSE ═══
Class: POWER  ·  Υ 283.17  ·  SNR 0.522  ·  Leverage 259.3×  ·  Velocity 1.092
...

signa> what should I do differently?
═══ SUGGEST ═══
1. Increase output by 19.43M (→ 42.79M)
   impact: Υ 283.17 → 518.64 (83.2% gain)
...

signa> simulate input -50%
═══ SIMULATE ═══
Base:    Υ 283.17  ·  POWER
Projected: Υ 1,132.66  ·  ARCHITECT+
Delta:   ↑ 849.49 (300.0%)
Class change: POWER → ARCHITECT+
...

signa> quit

LLM mode (optional)

The REPL works without an LLM — it pattern-matches input to skills directly. For conversational responses, enable the Claude API adapter:

# Set your API key
export ANTHROPIC_API_KEY=sk-ant-...

# Or in ~/.signa/settings.json:
{ "llm": "claude" }

When enabled, signaf sends only computed metrics (class, yield, pillars, ASI dimensions, taste dimensions) to Claude for conversational formatting. No session logs, no code, no message content. Default mode (stub) sends nothing — zero API calls, zero data leaves.

MCP server mode

signaf --mcp starts a stdio MCP server that exposes 12 tools. Your AI agent (Claude Code, Cursor, Windsurf) can call them through MCP. You bring your own LLM; signaf provides the skills.

// In .mcp.json:
{
  "mcpServers": {
    "signaf": {
      "command": "signaf",
      "args": ["--mcp"]
    }
  }
}

MCP tool	What it does
signa_scan	Read logs, compute cascade + ASI + taste
signa_diagnose	Pillar-level audit + issue detection
signa_simulate	Project Υ/class delta from a pillar change
signa_suggest	Ranked recommendations with impact
signa_taste	5-dimension behavioral taste profile
signa_asi	8-dimension Appropriate Steering Index
signa_bridge	Taste → cascade coaching insights
signa_cost	Token-to-cost analysis + cache savings
signa_goal	Path to a target class
signa_compare	Head-to-head vs class benchmark
signa_track	Metrics over time from history
signa_anomaly	Detect metric drops

Context is cached for 60 seconds to avoid re-reading logs on every tool call. All data stays local.

Skills

Skill	Trigger	What it does
scan	"scan", "refresh"	Read logs, compute cascade + ASI + taste profile, save to history
diagnose	"how am I doing", "audit"	Pillar-level audit: which pillar is weak, why
simulate	"simulate", "what if"	Project Υ/class delta from a hypothetical pillar change
suggest	"suggest", "what should I do"	Ranked recommendations with simulated impact
track	"track", "am I improving"	Metrics over time from local history
taste	"taste", "profile"	Show your behavioral taste profile (5 dimensions)
asi	"asi", "steering"	Show your Appropriate Steering Index (8 dimensions)
bridge	"bridge", "connect"	Taste → cascade coaching insights
goal	"goal", "how do I hit"	Path to a target class (TRANSMITTER, ARCHITECT, etc.)
cost	"cost", "how much"	Token-to-cost analysis (Claude pricing)
anomaly	"anomaly", "did anything drop"	Detect metric drops, pinpoint when
self-improve	"self-improve", "coach"	Full cycle: diagnose → suggest → simulate → next actions
compare	"compare", "vs"	Head-to-head vs class average
watch	"watch", "daemon"	Background daemon: auto-scan on .jsonl changes

Appropriate Steering Index (ASI)

ASI measures whether your interventions were the RIGHT ones, not just how often you accepted. Based on Anthropic's autonomy research.

8 dimensions:

• Acceptance rate — fraction of turns used as-is
• Correction rate — fraction of turns you re-edited
• Rejection rate — fraction of turns explicitly rejected
• Correction precision — how targeted your corrections were (single-file vs scattered)
• Intervention timing — how long you let the agent work before intervening
• Reliance slope — trend of your acceptance rate over the session (stable/improving/declining)
• Over-correction index — fraction of corrections that were potentially unnecessary
• Under-steering index — fraction of turns where you should have intervened but didn't

Each dimension reports a confidence level (high/medium/low) based on sample size.

Why ASI, not just SE? SE v1 (the legacy metric) measured acceptance rate — a high SE just means you said "yes" a lot. ASI measures whether your "yes" was the right call. An SE of 0.99 can correspond to an ASI of 0.686 — revealing that 49% of corrections were potentially unnecessary.

The taste profile

Saved at ~/.signa/taste-profile.json. Generated from your last 30 days of logs. Behavioral, not content-based — 5 dimensions:

• Steering signature — your ASI dimensions + SE legacy
• Iteration fingerprint — which files you iterate on, loop depth, convergence patterns
• Workflow rhythm — tool distribution, investigate-to-edit ratio, workflow style
• Cascade personality — pillar distribution tendencies (cache-hoarder, input-minimizer, output-light, high-leverage)
• Correction taxonomy — what you correct (design vs logic vs config), categorized by file type

Layer 3 (content-based) is opt-in. Use signaf taste --deep or pass { deepTaste: true } to the MCP tool. Raw content is not retained — only distilled preferences.

The profile is operator-owned: you can read it, edit it, share it, or delete it. It never leaves your machine by default.

The taste → cascade bridge

The bridge connects your behavioral taste profile to your cascade performance, generating coaching insights unique to SigRank. No other tool can do this — it requires both the taste profile AND the cascade formula.

Example insights:

• "Bash-heavy workflow" (high severity) — you run a lot of Bash commands, which tend to reset context. Impact: lower cache reads → lower leverage → lower Υ. Recommendation: batch your commands.
• "Diverging file loops" (high severity) — you're iterating on files without converging. Impact: high input, low output per turn. Recommendation: stop the loop and give explicit taste guidance.
• "Output-light personality" (high severity) — your output-to-input ratio is low. Impact: low velocity → low Υ. Recommendation: ask the agent for complete implementations, not pieces.

Privacy

Everything stays local. The agent reads all three signal layers from your logs, builds the taste profile, computes metrics — all on-device. Nothing is transmitted.

MCP server mode: All computation happens locally. The MCP server only exposes computed results to your AI agent via stdio. No data is sent to any server.

LLM mode (optional): When enabled, signaf sends only computed metrics (class, yield, pillars, ASI dimensions, taste dimensions) to Claude for conversational formatting. No session logs, no code, no message content. Default mode (stub) sends nothing.

File layout

~/.signa/
  taste-profile.json   — your taste profile (regenerated on each scan)
  history.json         — cascade metrics over time (append-only, capped at 1000)
  settings.json        — codename, platform, log root path, llm config

Architecture

signa/
  src/
    index.mjs           — entry: CLI dispatch + REPL boot + --mcp flag
    repl.mjs            — interactive chat loop (readline, pattern-matches to skills)
    mcp-server.mjs      — MCP server: 12 tools, stdio transport, 60s context cache
    logreader.mjs       — rich session-log reader (all 3 signal layers)
    cascade.mjs         — Υ/SNR/Leverage/Velocity/class + simulate + cost (pure math)
    store.mjs           — local JSON persistence (~/.signa/)
    watch.mjs           — daemon: auto-scan on .jsonl change
    taste/
      extractor.mjs     — extract taste signal from logs (3 layers, Layer 3 opt-in)
      profile.mjs       — build + save + load taste profile
      se.mjs            — Steering Efficiency (SE v1) + Appropriate Steering Index (ASI v2)
      bridge.mjs        — taste → cascade coaching insights
    skills/
      index.mjs         — all 13 skills (diagnose, simulate, suggest, taste, asi, etc.)
    llm/
      stub.mjs          — LLM interface (delegates to claude.mjs when configured)
      claude.mjs        — Claude API adapter (operator brings own key, metrics-only)

Brainstorm

This agent was built from the SigRank brainstorm package. The brainstorm stays untouched — this is the build that came out of it. See PLAN.md for the build plan.

Contributing

Contributions welcome. signaf is built in the open.

• Report bugs via GitHub Issues
• PRs: fork → branch → tests pass → open PR against main
• See the SigRank ecosystem section for how this repo relates to the others

License

CC-BY-NC-4.0 — see LICENSE.