Skip to content

CHANGELOG.md

Latest commit

 

History

History
90 lines (67 loc) · 5.11 KB

CHANGELOG.md

File metadata and controls

90 lines (67 loc) · 5.11 KB
Raw

Changelog

All notable changes to momo Code are documented in this file.

The format is based on Keep a Changelog.

[Unreleased]

Added

  • Pi config bridge (src/pi-config-sync.ts) — the v2.0 Pi runtime now inherits momo's provider config (CC Switch → momo.jsonc → MOMO_* env). The resolved key/endpoint/model are written to Pi's auth.json / models.json / settings.json under the Pi agent directory, so momo starts with working credentials instead of "No API key found for the selected model". Built-in providers get an auth.json entry; relays and non-bundled providers (zhipu/doubao/stepfun/alibaba/moonshot/custom) are declared as a momo custom provider. Existing OAuth credentials are never overwritten. Kill switch: MOMO_PI_CONFIG_SYNC=false

Changed

  • Pi SDK 1.1.0 → 1.4.0 (@earendil-works/pi-coding-agent, @earendil-works/pi-codemode) — adapts to the Pi 1.2 before_agent_start union (messageType: "user" | "custom") in the experience and memory extensions; version pins in README / CLI hints updated

Fixed

  • src/cli/chat.ts now imports os as a default import so tests can mock the home directory (same convention as src/provider/cc-switch.ts)

[2.0.0] - 2026-10-11

Major architecture upgrade: converged Pi-based agent architecture.

Added

  • Unified CLI entry (bin/momo) — single entry point supporting three modes: interactive REPL, Pi agent, and legacy /evolve /fine-tune handlers
  • JEV decision layer (src/jev/) — 5 decision points, shadow mode, and ECE calibration
  • DecisionPort contract (src/decision/port.ts) — typed decision interface (choice / bool / score / abstain)
  • RSI 3-level self-evolution loop (src/rsi/) — R1/R2/R3 loop with budget control and AbortController
  • momo-bench (src/momo-bench/) — three-tier evaluation framework (selection / holdout / canary) plus promotion gate with 7 reverse test cases
  • EvolutionCandidate contract (src/evolve/candidate.ts) — candidate + PromotionResult with CAS guard
  • Experience closure (src/experience/closure.ts) — idempotent TaskExposure → TaskOutcome → Solidify loop
  • Pi extensions (src/pi-extensions/) — 6 extensions: brand banner, CoT panel, decision capsules, tactical experience injection, sensitive-operation confirmation, RSI signal collection
  • Interactive REPL (src/repl/) — TTY entry with keyboard bindings (⌃T / ⌃O / ⌃C), collapsed/expanded CoT rendering, decision capsules with evidence bars
  • Runtime event envelope (src/runtime/types.ts) — TaskResult + 8 event envelope types
  • momo.jsonc v2 config schema (src/config/schema.ts)
  • Tests — e2e smoke suite (23 checks) and P0–P6 integration tests

Changed

  • LICENSE / NOTICE: original-work declaration; Pi (@earendil-works/pi-coding-agent) is consumed as a third-party Apache-2.0 dependency, not a derivative work
  • Simulated Math.random() paths in evaluator and gate are now explicitly marked with [SIMULATED] warnings
  • Dependencies: added @earendil-works/pi-coding-agent@1.1.0 and tsx; added smoke / loop / bench scripts

[1.0.0] - 2026-06-16

Added

  • Experience Fast Loop (/evolve) — Knowledge Embedding Protocol (KEP) for second-level learning
    • Tactic model with Beta(α, β) distribution tracking
    • Thompson sampling and UCB1 selection strategies
    • Signal pattern matching from session trajectories
    • Prompt injection with token budget management
    • Solidify step for verdict feedback
    • Promotion gate with A/B benchmark + ratchet check
    • Two-speed bridge: promoted tactics → fine-tune curriculum
    • Safety guardrails (command whitelist, stagnation detection, PII scrubbing)
    • Append-only ledger (JSONL) for auditability
  • Weight Slow Loop (/fine-tune) — Monte Carlo Graph Search (MCGS) training pipeline
    • Signal mining from session trajectories
    • Curriculum synthesis (Gold/Hard-negative/Replay slices)
    • MCGS exploration (EXPAND/SCORE/FUSE/PRUNE)
    • LoRA fine-tuning integration
    • Ratchet gate for monotonic improvement
    • Model registry with promote/rollback
  • CLI Command System — momo /evolve, momo /fine-tune, momo models, momo help
  • 53 TypeScript modules across provider / evolve / experience / cli layers
  • Dual-speed evolution architecture — experience (seconds) + weights (hours)

Changed

  • Product name: kqq Code → momo Code
  • bin/momo: Rewritten from CJS to ESM command router
  • package.json: Production-ready configuration (exports, files, engines, keywords)
  • README: Complete rewrite with /evolve documentation and architecture diagrams

[0.1.0] - 2026-06-15

Added

  • Initial momo Code implementation
  • 19 LLM provider integrations (Claude, GPT-4, Gemini, OpenRouter, Groq, Mistral, etc.)
  • Model tiers (ultra / standard / lite)
  • Claude Code interoperability (.claude/ config inheritance, MCP servers, prompts)
  • SSE streaming with 8-minute chunk timeout watchdog
  • Auth system with provider credential management
  • Configuration system (JSONC with schema)
  • Prompt routing with multi-model support
  • Effect-powered architecture