micro-models-agent 0.57.2 → 0.58.0

package/CHANGELOG.md

@@ -4,6 +4,34 @@ All notable changes to Micro Models Agent (MMA) will be documented in this file.
 
 The format is based on [Keep a Changelog](https://keepachangelog.com/), and this project adheres to [Semantic Versioning](https://semver.org/).
 
+## [0.58.0] - 2026-08-31
+
+### Added
+- **MoE experts**: dynamic expert registry — `config.experts` (with optional `ExpertConfig.description`) is rendered into the MoE planner prompt instead of a hardcoded code/research/browser/vision list (fallback: `tool_tags`)
+- **MoE orchestrator provider**: `OrchestratorClient` builds its provider through `ProviderManager` — per-entry `contextWindow`/`retry`/`rateLimits` and failover entries are honored; `orchestrator.contextWindow` is the global fallback (legacy `{type,baseUrl,apiKey}` configs keep working)
+- **MoE success criteria**: machine-readable `success_criteria` (`file-exists:<path>`, `substring-in-file:<path>:<text>`, `command-exit-0:<cmd>`) verified by `StepVerifier.verifySubtaskCriteria` before merge; LLM-only criteria surface as warnings in the Router context
+- **MoE selective re-plan**: `executePlan(plan, {skipIds, carriedResults})` — already-succeeded subtasks are carried over across re-plan cycles, only failed/new subtasks re-execute
+- **MoE observability**: `moe_plan` / `moe_subtask` / `moe_replan` / `moe_verify` events written to `session.jsonl` (visible in `mma session show`); token usage aggregated per `expert_tag` (`MoEExecutor.getUsageByTag`); sub-agents receive the real session id instead of the hardcoded `"moe"`
+- **MoE cert scenarios**: `5.2-moe-parallel-waves` and `5.3-moe-read-after-write` (tag `moe`, timeoutMs 600_000)
+- **ToolResult.usage**: sub-agent token usage propagated to the MoE executor for cost attribution
+
+### Changed
+- **Reasoning policy baseline**: configurable via `reasoning.baseline` (default `medium`) — the auto policy's quiet-iteration level is no longer hardcoded, so trivial Q&A turns can start at `low`
+
+### Fixed
+- **MoE hang**: sub-agents spawned by the `subagent` tool re-entered `runWithMoE` when the parent config had `moe.enabled=true` — each MoE sub-agent planned its own subtasks and spawned nested sub-agents recursively (bounded only by `maxRecursionDepth`), hanging execution for many minutes with zero progress events. Sub-agents now always run the plain single-agent loop (`moe` is a top-level orchestration mode only).
+- **MoE fallback visibility**: missing `orchestrator.model` with `moe.enabled=true` now logs a WARN with a fix hint instead of a debug-only message that made silent fallback to single-agent invisible.
+- **CLI one-shot commands hang**: `config`, `session`, `model`, `provider`, `context`, `security`, `plugins`, `changelog` printed their result but never exited — bootstrap leaves open handles (reasoning-probe fetch, indexer) and only the prompt/REPL paths called `process.exit`. Subcommand roots now exit via a `postAction` hook (verified: `session list` 2.7s / exit 0, chained `config set` persists correctly).
+- **Repeated tool calls**: identical `(tool, args)` repeat now injects a nudge into context in ALL modes — previously the nudge was gated behind `--exit-on-complete`, so interactive runs silently burned iterations re-running the same command (observed: duplicate `node sum.js` verification back-to-back).
+- **Perfectionism cycle**: repeated mutations of the SAME target (same file via `edit_file`/`write_file`, same bash command, same `plan`/`todo` action) with varying arguments now inject a finalize-nudge after 5 occurrences — the consecutive-duplicate guard cannot see this pattern (args differ each time; observed: 20+ polish iterations after the task was already complete, then 10+ bookkeeping churn iterations after an audit rejection). The nudge lands AFTER tool results, so `assistant(tool_calls) → tool(results) → user(nudge)` pairing stays valid.
+- **Audit gate friction (evidence-based auto-close)**: pending plan steps whose every file-like token resolves to an existing file are auto-closed by the final audit with a note — small models routinely finish the work but forget `plan update` bookkeeping, and the gate then rejected correct answers burning all retry budget. Steps without file tokens or with missing files stay pending (honest).
+- **Audit rejection guidance**: the audit gate's `<system-summary>` now names the resolution paths (mark done / skipped / re-plan) in en+ru instead of a bare "continue working" — model plan-drift after a rejection (wrong file names, changed approach) is common for 9B and the old message did not point at the fix.
+- **LSP false positives without type env**: `lsp_check` suppresses module-resolution diagnostics when the project root has no `tsconfig.json`/`jsconfig.json`/`node_modules` and reports an inconclusive note instead — phantom `Cannot find module 'fs'` / `Cannot find name '__dirname'` errors made the model abandon a working `.ts` approach for `.js` and spiral into plan drift.
+- **Windows Unix-command hints**: `bash` hint list and the module-side forbidden set now cover `sed`/`awk`/`uniq`/`xargs`/`cut`/`tr`/`basename`/`dirname`/`export`/`source`/`sleep`/`env` (both lists kept consistent) — the 9B model keeps reaching for Unix tools in cmd.exe.
+- **Tool visibility narrowing**: `chunk_query`, `download_file`, `mcp_call`, `pipeline_run` no longer visible by default — they moved behind `enable_tools` (`research`/`shell` tags); the default visible set drops from 23 to 19 schemas, cutting tool tokens on every turn.
+- **Tests**: replaced process-global `vi.mock` with DI seams (`AgentDeps.runWithMoEOverride`, `ToolContext.agentFactory`) — Bun module mocks leak across test files in the same worker; full suite is now green by default (1803 pass / 0 fail).
+- **Docs**: AGENTS.md / testing.md — warn that `mma` resolves the globally installed package (stale `dist/`) when run outside the repo root, silently ignoring `src/` changes; always run from the repo root with `-d <sandbox>` and verify the version in the `Environment:` line.
+
 ## [0.57.2] - 2026-08-30
 
 ### Changed