@zhixuan92/multi-model-agent 4.0.3 → 4.0.4

package/README.md

@@ -90,7 +90,7 @@ Two ways — pick one:
 
 ```bash
 mmagent serve                          # 127.0.0.1:7337 by default
-curl -s http://localhost:7337/health   # → {"ok":true,"version":"4.0.3",...}
+curl -s http://localhost:7337/health   # → {"ok":true,"version":"4.0.4",...}
 ```
 
 For an always-on background install (survives reboots): [launchd / systemd templates](./scripts/README.md).
@@ -290,20 +290,17 @@ Full design rationale: [DIRECTION.md](https://github.com/zhixuan312/multi-model-
 | TLS `handshake_failure` to a known-good telemetry endpoint | Local DNS cache is stale. `sudo dscacheutil -flushcache && sudo killall -HUP mDNSResponder` (macOS); restart the daemon so its Node process re-resolves |
 | Local telemetry queue stops draining | Daemon's flusher is in exponential backoff after a transport failure (capped at 1 hr). Restart the daemon to force an immediate boot-flush |
 
-## What's new in 4.0.3
+## What's new in 4.0.4
 
-- **Required headers `X-MMA-Main-Model` + `X-MMA-Client` on every tool route.** Server returns `400 main_model_required` / `400 client_required` if missing. Replaces `defaults.parentModel` config (removed) and `PARENT_MODEL_NAME` env. Shipped skills set both headers automatically from `MMAGENT_MAIN_MODEL` + `MMAGENT_CLIENT` env vars.
-- **Wire field rename:** `parentModel*` → `mainModel*` (matches DB column `main_model`). `costDeltaVsParentUSD` → `costDeltaVsMainUSD`. Internal record now matches wire 1:1.
-- **Canonical model-name preservation.** `claude-opus-4-7` no longer collapses to `claude-opus`. Best-effort extraction handles arbitrary wrappers (`bedrock.claude-opus-4-7`, `vertex_ai/anthropic.claude-sonnet-4-6@2024-10-22`).
-- **`contextBlockIds` actually reach the worker prompt.** Round-over-round audit recipes were broken pre-4.0.3 — the dispatcher dispatched the unexpanded task. Now the same expanded reference flows through both `state.task` and `executionContext.task`.
-- **File-backed context blocks survive daemon restarts.** Stored at `<projectCwd>/.mma/context-blocks/<id>.txt` with atomic writes, mode `0700`/`0600`, 7-day TTL, 1 MiB / 100 MiB caps. `.mma/` belongs in `.gitignore` (daemon prints a stderr breadcrumb on first creation).
-- **`totalDurationMs` reflects real wall-clock** (was implementer-only). Per-stage durations stay truthful — drops the proportional scale-down that was masking the under-counting.
-- **Audit/review/delegate headlines** fall back to `runResult.annotatedFindings` and `runResult.filesWritten` when the structured report lacks them. Pre-fix headlines reported `0 findings (0 high)` and `(0 files)` for narrative-emitting tools.
-- **`batch_failed` fires when the executor packages an error envelope.** Operator visibility — verbose stream no longer says `batch_completed` while the run actually failed.
-- **`run_shell` write tracking.** Workers writing via `cat >`, `sed -i`, `tee`, etc. correctly increment the polling headline's write count.
-- **Stage-progression denominator derives from the StagePlan.** Audit `(1/3)`, delegate `(1/9)`, register-context-block `(1/1)`. Single source of truth.
+- **Reviewers see the actual diff.** New `DiffTracker` (snapshot-based, works in non-git directories) gives spec / quality / diff reviewers the cumulative unified diff against the pre-task baseline. Pre-fix the reviewer judged the worker's text claim alone, defaulted to `changes_required`, and triggered rework spirals on already-correct work. Verdicts must now point to specific diff lines.
+- **Coherent prompts via shared rubric.** `finding-criteria.ts` is the single source of truth for severity ladder, evidence-grounding, scope discipline, and stage awareness. Read-only tools share `ANNOTATOR_CHECK_AWARENESS_RO`; artifact-producing tools share `REVIEWER_AWARENESS_AP`. Workers self-align with what the reviewer will judge → cleaner first-round outputs.
+- **Lenient JSON parsers** in both reviewer and annotator output paths. Accepts ` ```json ` fenced, ` ``` ` (no language tag), bare JSON, and embedded objects/arrays. Caused `verdict: 'error'` and `findings_low: 0` regressions despite valid model output.
+- **Cumulative `filesWritten` across rework rounds.** Pre-fix the implementer's writes were wiped when a no-op rework round overwrote `lastRunResult`. Now unioned across rounds.
+- **Headlines unified across all tools** — `[<status>] <route>: <summary>`. `execute-plan` (was `execute_plan` snake-case), `retry: N/N tasks complete` now reflects per-task status, debug carries file path + finding count.
+- **Implementer system prompt + per-tool prompts hardened.** "Trust `edit_file`/`write_file` — do NOT re-read just to verify your own successful edit" saves 4-6 min per artifact task. Read-only prompts include severity calibration, evidence-grounding, scope discipline, stage awareness.
+- **Per-tool fixes:** investigate prompt aligned with parser, verify `findings_low` correctly populated, debug rewritten as proper read-only (PROPOSE — do NOT apply), spec / quality concerns accumulated across rounds.
 
-**Migration from 4.0.2:** custom HTTP callers must add `X-MMA-Main-Model` and `X-MMA-Client`. Skills users get this for free after `mmagent sync-skills`.
+**Migration from 4.0.3:** none. Wire envelope, schema fields, and route names are unchanged. `npm update` to take the bug fixes.
 
 Full history: [CHANGELOG](https://github.com/zhixuan312/multi-model-agent/blob/master/CHANGELOG.md).
 

package/dist/skills/mma-audit/SKILL.md

@@ -8,7 +8,7 @@ when_to_use: >-
   User asks for a doc/spec/config audit OR a methodology skill
   (superpowers:dispatching-parallel-agents, /security-review) points at one AND
   mmagent is running. Audit on PROSE/SPEC docs — use mma-review for source code.
-version: 4.0.3
+version: 4.0.4
 ---
 
 # mma-audit

package/dist/skills/mma-context-blocks/SKILL.md

@@ -12,7 +12,7 @@ when_to_use: >-
   Register once here, then pass the ID via `contextBlockIds` on mma-delegate /
   mma-execute-plan / mma-audit / mma-review / mma-verify / mma-debug /
   mma-investigate. Cheaper and faster than inlining the same content N times.
-version: 4.0.3
+version: 4.0.4
 ---
 
 # mma-context-blocks

package/dist/skills/mma-debug/SKILL.md

@@ -10,7 +10,7 @@ when_to_use: >-
   read files, reproduce, trace — OR a methodology skill
   (superpowers:systematic-debugging) points at the investigation step. Delegate
   the read/reproduce/trace; the main agent stays on the hypothesis and the fix.
-version: 4.0.3
+version: 4.0.4
 ---
 
 # mma-debug

package/dist/skills/mma-delegate/SKILL.md

@@ -11,7 +11,7 @@ when_to_use: >-
   and keep main context free. If a plan file exists → use mma-execute-plan. If
   the task is audit / review / verify / debug / investigate → use the matching
   specialized skill.
-version: 4.0.3
+version: 4.0.4
 ---
 
 # mma-delegate

package/dist/skills/mma-execute-plan/SKILL.md

@@ -10,7 +10,7 @@ when_to_use: >-
   superpowers:subagent-driven-development / superpowers:executing-plans —
   workers are cheaper and don't pollute main context. Task descriptors must
   match plan headings verbatim.
-version: 4.0.3
+version: 4.0.4
 ---
 
 # mma-execute-plan

package/dist/skills/mma-explore/SKILL.md

@@ -11,7 +11,7 @@ when_to_use: >-
   Delegating the read+grep+web-search to a worker keeps your main context on
   judgment. DO NOT use for convergent single-answer questions (where is X
   called, how does Y work) — those are mma-investigate.
-version: 4.0.3
+version: 4.0.4
 ---
 
 # mma-explore

package/dist/skills/mma-investigate/SKILL.md

@@ -12,7 +12,7 @@ when_to_use: >-
   git-history queries. OR you are about to read 3+ files / run any grep in main
   context — that's the inline-labor-leakage anti-pattern (AP2); delegate to this
   skill instead.
-version: 4.0.3
+version: 4.0.4
 ---
 
 # mma-investigate

package/dist/skills/mma-retry/SKILL.md

@@ -10,7 +10,7 @@ when_to_use: >-
   you want to re-try the failed indices only. Prefer this over re-dispatching
   the whole batch or inline-retrying — it's idempotent and preserves the
   original batch's diagnostics.
-version: 4.0.3
+version: 4.0.4
 ---
 
 # mma-retry

package/dist/skills/mma-review/SKILL.md

@@ -10,7 +10,7 @@ when_to_use: >-
   AND mmagent is running. Delegate so each file reviews on its own worker; the
   main agent only decides what to merge. Review on SOURCE CODE — use mma-audit
   for prose specs / configs.
-version: 4.0.3
+version: 4.0.4
 ---
 
 # mma-review

package/dist/skills/mma-verify/SKILL.md

@@ -10,7 +10,7 @@ when_to_use: >-
   against implemented work BEFORE claiming success. Delegate so each checklist
   item gets independent evidence-gathering on a worker. Use this BEFORE saying
   "done" — never after.
-version: 4.0.3
+version: 4.0.4
 ---
 
 # mma-verify

package/dist/skills/multi-model-agent/SKILL.md

@@ -11,7 +11,7 @@ when_to_use: >-
   tasks — AND mmagent is running. Read this once, pick the matching mma-* skill,
   and delegate there. Applies equally whether the user invoked a superpowers
   methodology skill or asked directly.
-version: 4.0.3
+version: 4.0.4
 ---
 
 # multi-model-agent (router)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@zhixuan92/multi-model-agent",
-  "version": "4.0.3",
+  "version": "4.0.4",
   "type": "module",
   "license": "MIT",
   "description": "Standalone HTTP server for multi-model-agent. Routes tool-invocation work to Claude, Codex, or OpenAI-compatible sub-agents with async-polling REST dispatch and installable skills for Claude Code, Gemini CLI, Codex CLI, and Cursor.",
@@ -53,7 +53,7 @@
   },
   "dependencies": {
     "@asteasolutions/zod-to-openapi": "^8.5.0",
-    "@zhixuan92/multi-model-agent-core": "^4.0.3",
+    "@zhixuan92/multi-model-agent-core": "^4.0.4",
     "gray-matter": "^4.0.3",
     "minimist": "^1.2.8",
     "proper-lockfile": "^4.1.2",