OptimusPrime v2.0.0 — Intelligence layer + Agentic Conductor

Claude is stateless. Your project isn't.

v2.0.0 makes it intelligent.

v1.0 gave Claude memory and enforcement.
v2.0 gives Claude self-awareness and autonomy.

---

What's new in v2.0.0

Self-Model Engine
Claude learns its own failure patterns on your project.
After enough sessions it knows what it handles well, what
it consistently gets wrong, what conditions cause loops.
Every session injects pre-emptive warnings before Claude
makes the same mistake again.

Codebase Map
Before any file edit, OptimusPrime scans what already
exists — utilities, installed deps, patterns in use,
things this project never uses. Ponytail tells Claude
to check if something exists. OptimusPrime tells Claude
what it would find.

Pre-Response Intelligence
Fires before Claude even responds to your message. Reads
what you asked, checks against the self-model and decision
history, injects warnings before Claude forms its approach.
The earliest possible intervention point.

Living Task Document
Maintained in real time during a session. Updated after
every significant tool call. Claude never loses the thread
of a long task.

Convention Extractor
Reads your actual codebase and extracts implicit rules —
naming patterns, error handling style, testing approach.
Not documentation you wrote. Conventions inferred from
real code. Checked after every file write.

op watch — Live Dashboard
Real-time terminal dashboard alongside Claude Code.
Tokens, cost, scope violations, decisions, loops, active
skills, self-model confidence. Updates every 3 seconds.

op autopilot — Pre-Session Brief
Before you open Claude Code, reads git diff, last snapshot,
open TODOs, known risks, and writes a ready-to-paste first
message. Every session starts oriented.

op replay — Session Debugger
Step through any past session like a debugger. Every
decision, every failure, every scope block, every loop
caught. See exactly where tokens were wasted.

op diff-intel — Between-Session Intelligence
Analyzes what changed in git since your last session.
Flags out-of-scope changes, rejected deps re-added,
test count drops, decision contradictions.

Conductor — Agentic Loop with Full Guardrails
You define a goal. Conductor breaks it into subtasks,
runs Claude headlessly on each, evaluates every output,
escalates when stuck, continues autonomously when not.

The entire enforcement stack acts as guardrails:

• Scope contract enforced during autonomous execution
• Loop detection escalates instead of burning tokens
• Self-model warnings injected before each subtask
• Done checker runs before marking anything complete

op conductor start --goal "refactor auth to use JWT"
op conductor resume --context "use RS256 not HS256"
op conductor status

---

Benchmarks

Metric	Result
Output compression	63.9% (non-destructive)
Scope guard	78.7ms avg (n=1000)
Loop detection	100% accuracy
Decision search	0.03ms avg
Self-model build	2.6ms avg
Codebase map scan	334ms first, 2.18ms warm
Conductor plan	0.001s avg

---

Numbers

• 374 tests passing
• 16 benchmarks all green
• 13 sessions to build
• 167 decisions logged during its own build
• Built with itself from session one

---

Install

git clone https://github.com/aarshsharma2803-stack/optimusprime
cd optimusprime
bash install.sh

MIT License.

v1.0.1 — Post-launch bug fixes

10 bugs found and fixed via end-to-end verification (53/53 tests).

Critical fixes:

• install.sh was missing predictive-context, learner-hook,
  SubagentStop and PreCompact hook registrations — new installs
  were incomplete
• scope-guard now blocks bash redirect operator targeting
  out-of-scope files (echo x > .env)
• MCP server now works on Python 3.9 with stdlib fallback
  (was crashing on FastMCP import)
• op skills list now wired to real registry (was still a stub)
• MCP tool public aliases restored (10 tests were broken)

All 145 tests passing. 53/53 verification checks passing.

OptimusPrime v1.0.0 — The session state protocol for AI coding

Claude is stateless. Your project isn't.

OptimusPrime is the first hook-level session state
protocol for AI coding. Runtime enforcement that fires
before execution. Cross-session memory that compounds
over time. Intelligence that gets smarter every session.

What's in v1.0.0

9 hooks — scope enforcement, loop detection,
dependency analysis, breaking change detection, output
compression, attempt logging, TODO accountability, done
checking, session snapshots

7 skills — scope contract extraction, decision
logging, confidence signaling, CLAUDE.md generation,
context restoration, cost awareness, production output
mode

Intelligence layer — contradiction detection across
decisions, topic clustering, predictive context injection
ranked by relevance not recency, cross-session learning
engine that adapts to your actual usage patterns

MCP server — 9 tools including reason_about(),
get_contradictions(), get_patterns() exposing full
project intelligence to any MCP-capable agent

CLI — op command with 9 command groups including
op intel ask, op intel patterns, op intel learned

Ecosystem Skills Hub — one-command install for
Superpowers, gstack, UI/UX Pro Max, caveman, ponytail.
Auto-updates. Contextual activation. Learns your
real activation thresholds from usage.

Benchmarks

Metric	Result
Output compression	63.9%
Scope guard latency	75ms (n=1000)
Loop detection	100% accuracy
Decision search	0.02ms avg
Session logger	0.12s
Input token reduction	40%+
Contradiction detection	163ms (n=50)
Context prediction	0.77ms avg
Learning cycle	1.8ms avg

Compression is non-destructive — code blocks untouched.

Install

git clone https://github.com/aarshsharma2803-stack/optimusprime
cd optimusprime
bash install.sh

Compatible agents

Claude Code · Antigravity · Cursor · Codex · Any MCP agent

What makes it different

Every other tool makes requests to Claude via CLAUDE.md.
Claude forgets them mid-session. OptimusPrime enforces
at the hook level — before execution, every time,
regardless of context window state.

And it gets smarter. Session 50 knows your patterns,
your preferred libraries, your failure history, your
real skill activation thresholds. Session 1 doesn't.

Built across 12 sessions. 123 decisions logged in
.optimusprime/decisions.md — ate our own cooking
through the entire build.

145 tests passing. MIT License.