Super Brain

Implementation workspace for the Fold Platform described in Fold_Platform_Super_Brain_Unified_Reference_v8.md.

The repository now contains the complete local vertical slice: authenticated harness ingestion, a transactional PostgreSQL Fold, resumable event subscriptions, historical transcript capture, automatic project-aware memory formation, reviewed promotion, full-corpus recall, and an operator UI. The original v0.7 risk-retirement gate remains covered by executable conformance tests.

Packages

• @_89/fold implements the v0.7 Change Record schema, deterministic replay, inclusive fork cursors, provenance/capture metadata, and sensor lifecycle freshness.
• @_89/fold-trace implements the projection feasibility contract, coverage, mapped-edge aggregation, first-divergent-edge analysis, tool-call divergence mining, support-based structural merge, and explicit verified/unknown outcome accounting. Its two-model test experiment remains a structural fixture, not empirical model evidence.
• @_89/fold-trajectory records shared decision trees and projected runs as scoped canonical Fold events, then rebuilds task coverage, observed routes, first divergence, and review-oracle results without hiding projection gaps.
• @_89/fold-eval implements the review verdict protocol, command oracle, fail-closed oracle configuration, explicit absent results, configurable confidence aggregation, and confidence-kernel history integration.
• @_89/fold-activity normalizes terminal output, classifies bounded recent state, preserves classifier provenance, and maps terminal manager signals to canonical lifecycle and observation records with complete source identity. Its envelope validator prevents generic or human-authored events from masquerading as terminal sensor evidence.
• @_89/fold-fleet reconstructs agent/session state from those records at boot, separates last-known status from lifecycle freshness, and plans timeout-gated orphan reconciliation without interpreting silence as offline.
• @_89/fold-transcript owns immutable project, artifact, source-qualified run, context-segment, turn, and observable-action records for historical Claude Code and Codex imports. Canonical records never contain transcript text or private reasoning.
• @_89/fold-drives incrementally advances drive and wear state, records causal discontinuities as canonical Fold events, and rebuilds surfaced, committed, declined, acted, and ended intentions without performing host cognition. The SDK and API expose that lifecycle as replayed human steering.
• @_89/fold-epistemic records, revises, forgets, and recalls personal memory plus immutable memory proposals, decisions, evidence revisions, and usefulness feedback. Memory is explicitly scoped by workspace, optional space, audience, and projects; access is reapplied after external ranking.
• @_89/fold-sdk provides journal-compatible producer and consumer APIs with capture-scope enforcement, canonical ordering, access-filtered projection, personal-memory lifecycle and authorized ranker orchestration, trajectory task/report facades, replay-built fleet reads over validated activity signals, and idempotent transcript import/catalog queries.
• @_89/fold-postgres is the transactional production store for canonical events, durable consumer offsets, and rebuildable projection checkpoints. Its optional pgvector ranker is derived from authorized memory and never becomes canonical state.
• @_89/super-brain-client is the harness-neutral authenticated client for append, recall, candidate review, resumable SSE, and principal-scoped durable consumer offsets. Codex, Claude, Hermes, or another harness can use the same boundary without a repository-specific adapter.
• @_89/super-brain-api serves those SDK operations over authenticated HTTP, derives authorship and personal-memory capture identity from credentials, resolves membership on every request, supplies a pluggable memory-ranker port with a deterministic local lexical provider, supplies a pull-reasoner port with an explicitly extractive local provider, and gates canonical steering. It uses transactional PostgreSQL for multi-process canonical persistence, durable consumer offsets, and projection checkpoints when FOLD_DATABASE_URL is configured; the fsynced JSONL store remains the local fallback and migration source. Hosted mode verifies Clerk tokens, exposes tenant-safe session discovery, applies signed idempotent organization webhooks, and lets organization admins enroll or revoke scoped machine identities without restarting the API. Owner-authorized transcript imports and workspace-readable history queries use dedicated routes that cannot be bypassed through generic event append. Its runtime adds exact-origin CORS, bounded per-address request limiting, connection timeouts, graceful lease-releasing shutdown, and an in-process validated-entry cache held behind the exclusive writer lease.
• @_89/super-brain is the responsive operator client for workspace activity, private personal-memory lifecycle and ranked recall, trajectory import and analysis, historical project/run inspection, live agent fleet inspection, pull reasoning, human steering, canonical/draft event inspection, and materialized Fold state. It talks only to the authenticated HTTP API. When Clerk is configured it provides sign-in, sign-out, organization switching, server-derived workspace selection, and in-memory token refresh; the manual local connection remains available when Clerk is absent.
• @_89/super-brain-importer streams local Claude Code and Codex JSONL, performs metadata-only dry runs, writes explicitly requested redacted artifacts to a restrictive local vault, and delivers confirmed canonical bundles through the authenticated API without modifying source histories.
• @_89/super-brain-capture-daemon is the loopback-only live sensor for Claude Code, Codex, and compatible harnesses. It acknowledges lifecycle hooks from a durable encrypted/redacted local inbox, then spools prompt, tool, file, verification, structured-reasoning, decision, trajectory, and final transcript-import work. Raw prompt and tool bodies stay in its secret-redacted private vault rather than canonical Fold. Append-only step journals remove the old long-session cap, retained encrypted hook evidence repairs historical omissions, and daemon-owned transcript snapshots eliminate mutable source-path delivery races.
• @_89/super-brain-memory-worker reads only the redacted transcript vault, proposes deterministic project-aware memories, consolidates repeated evidence, and consumes transcript, live-checkpoint, decision, and trajectory events from a durable cursor. A narrow opt-in policy automatically promotes high-confidence structured observations, explicit human decisions, and reasoning checkpoints cited by a verified successful trajectory; ambiguous proposals remain reviewable. With a real model provider it also periodically proposes cited cross-project synthesis, contradictions, procedures, and investigations without silently promoting model output.
• @_89/super-brain-mcp-server gives any MCP-compatible harness authenticated memory search, cited context assembly, structured checkpoint and proposal capture, and immutable helpful/unhelpful/superseded feedback tools.
• @_89/confidence-kernel is the pinned 0.2.0 history-scoring, drift, pooling, oracle, and journal implementation imported under its MIT license.
• @_89/fold-narrative projects canonical Fold events into arc state, apply-time knowledge, journeys, curves, resolution peaks, and convergence. Its first slice is pinned to the Mythopia Fellowship fixture without a production Mythopia adapter.
• @_89/fold-storage provides versioned JSONL append/replay, strict reopen, torn-tail recovery, atomic rewrites, and verified materialized-state checkpoints. Its format is specified in docs/storage/JSONL_FORMAT_v1.md.

The target package graph, source ownership rules, and per-repository extraction map are recorded in docs/EXTRACTION_BLUEPRINT.md. Referenced repositories are prior art and behavioral evidence; production packages here do not depend on their worktrees.

This repository does not modify the referenced repositories in /Users/jakobgrant/Workspaces. Their observed baselines are recorded in docs/EVIDENCE_MANIFEST.md.

Commands

pnpm install
pnpm verify

Local client setup and proxy configuration are documented in apps/brain/README.md. Historical transcript identity, privacy, import, and query behavior is recorded in docs/TRANSCRIPT_INGESTION.md. The live architecture, memory lifecycle, harness contract, and PostgreSQL operations are documented in docs/ARCHITECTURE.md. The implementation ledger and remaining deployment choices are in docs/ROADMAP.md, with acceptance evidence tracked in docs/EXECUTION_BACKLOG.md.

The repository is public at project-89/super-brain. New Fold packages remain UNLICENSED until package ownership and release terms are settled; public visibility does not grant a reuse license. The imported confidence kernel retains its MIT license.

Security reports and contribution requirements are documented in SECURITY.md and CONTRIBUTING.md.