Merge origin/main into codex/local-memory-v1

2026-04-04 03:54:40 +09:00
parent 54abb5bb4e 371a3ef0ff
commit 0b4228b391
8 changed files with 304 additions and 22 deletions
--- a/container/Dockerfile
+++ b/container/Dockerfile
@@ -31,6 +31,10 @@ RUN apt-get update --fix-missing && apt-get install -y --no-install-recommends \
    zip \
    unzip \
    strace \
+    ripgrep \
+    fd-find \
+    bat \
+    tree \
    && curl -fsSL https://cli.github.com/packages/githubcli-archive-keyring.gpg \
       | dd of=/usr/share/keyrings/githubcli-archive-keyring.gpg \
    && echo "deb [arch=$(dpkg --print-architecture) signed-by=/usr/share/keyrings/githubcli-archive-keyring.gpg] https://cli.github.com/packages stable main" \
@@ -52,18 +56,24 @@ RUN curl -fsSL https://bun.sh/install | bash
 # Create app directory
 WORKDIR /app

-# Copy host's agent-runner (same code, same SDK version)
-COPY runners/agent-runner/package*.json ./
-RUN bun install
-COPY runners/agent-runner/ ./
-RUN bun run build
+# Copy agent-runner (Claude Code)
+COPY runners/agent-runner/package*.json ./agent/
+RUN bun install --cwd ./agent
+COPY runners/agent-runner/ ./agent/
+RUN bun run --cwd ./agent build
+
+# Copy codex-runner
+COPY runners/codex-runner/package*.json ./codex/
+RUN bun install --cwd ./codex
+COPY runners/codex-runner/ ./codex/
+RUN bun run --cwd ./codex build

 # Create workspace directories
 RUN mkdir -p /workspace/project /workspace/group /workspace/global \
    /workspace/ipc/messages /workspace/ipc/input

-# Create entrypoint script
-RUN printf '#!/bin/bash\nset -e\ncat > /tmp/input.json\nbun /app/dist/index.js < /tmp/input.json\n' \
+# Entrypoint (fallback — persistent containers use docker exec with explicit runner path)
+RUN printf '#!/bin/bash\nset -e\ncat > /tmp/input.json\nbun /app/agent/dist/index.js < /tmp/input.json\n' \
    > /app/entrypoint.sh && chmod +x /app/entrypoint.sh

 # Set ownership to node user (non-root)
--- a/docs/architecture.md
+++ b/docs/architecture.md
@@ -0,0 +1,106 @@
+# Architecture
+
+## Service Stack
+
+Single unified service (`ejclaw`) manages all three Discord bots in one process:
+
+- `ejclaw.service` — Single unified process
+- Discord bots: `DISCORD_BOT_TOKEN` (Claude), `DISCORD_CODEX_BOT_TOKEN` (Codex-main), `DISCORD_REVIEW_BOT_TOKEN` (Codex-review)
+- Paired review: owner ↔ reviewer (agent types configurable per role)
+- Reviewer fallback: Claude exhaustion → codex-review auto-handoff
+- Shared dirs: `store/`, `groups/`, `data/`
+- SQLite WAL mode + `busy_timeout=5000` for concurrent access
+
+## Data Flow
+
+```
+Discord ──► SQLite (WAL) ──► GroupQueue ──┬──► Owner (host process)
+                                          │       │
+                                          │       ▼ (auto-trigger)
+                                          ├──► Reviewer (Docker container, :ro mount)
+                                          │       │
+                                          │   Verdict routing
+                                          │       ├─ DONE → change detection → finalize or re-review
+                                          │       ├─ BLOCKED → Arbiter (if enabled) or User
+                                          │       └─ Feedback → Owner (loop)
+                                          │
+                                          ├──► Arbiter (on-demand, fresh session each time)
+                                          │       │
+                                          │   ┌───┴─── MoA (if enabled) ───┐
+                                          │   │ Ref model A ──► opinion    │
+                                          │   │ Ref model B ──► opinion    │
+                                          │   │ → injected into prompt     │
+                                          │   └────────────────────────────┘
+                                          │       │
+                                          │   PROCEED/REVISE/RESET/ESCALATE
+                                          │
+                                     IPC polling ◄── follow-up messages
+                                          │
+                              ┌────────── Router ──────────┐
+                              ▼                            ▼
+                   paired_turn_outputs           Discord (display only)
+                   (agent ↔ agent data)      (user observation, @mention)
+```
+
+## Room Assignment Model
+
+Per-room routing uses an explicit assignment model:
+
+- `room_settings` is the room-level source of truth (SSOT)
+- Each room stores:
+  - `room_mode`: `single` or `tribunal`
+  - `owner_agent_type`: `codex` or `claude-code`
+- Public room assignment uses `assign_room`
+- `registered_groups` remains as a materialized capability/read-model layer
+
+Operationally:
+
+- `single` → one owner bot
+- `tribunal` → per-room owner + globally configured reviewer + optional arbiter
+
+Tribunal is no longer inferred from "two bots registered on one room"; it is an explicit room setting.
+
+## Tribunal 3-Agent System
+
+```
+User message
+  → Owner responds (implementation, answer, etc.)
+    → Reviewer auto-triggered (critical review, design check)
+      → Verdict:
+          DONE              → Owner finalizes → Task completed → @user ✅
+          DONE_WITH_CONCERNS → Owner addresses feedback → loop
+          BLOCKED/NEEDS_CONTEXT
+            ├─ Arbiter enabled → Arbiter judges → PROCEED/REVISE/RESET/ESCALATE
+            └─ Arbiter disabled → Escalate to user → @user ⚠️
+      → Owner BLOCKED/NEEDS_CONTEXT → Arbiter (same path as reviewer)
+      → Deadlock (3+ round trips without progress)
+          → Arbiter summoned → binding verdict → loop resumes
+```
+
+## Container Isolation
+
+Reviewer and arbiter run inside a Docker container with:
+
+- Read-only source mount (kernel-level write protection)
+- Both Claude (agent-runner) and Codex (codex-runner) supported
+- Runner selected at `docker exec` time based on agent type
+- tmpfs overlays for test runner caches
+- IPC via filesystem for follow-up messages
+- Credentials injected per-exec (never baked into container)
+
+## Key Files
+
+| File | Purpose |
+|------|---------|
+| `src/index.ts` | Orchestrator: state, message loop, agent invocation |
+| `src/agent-runner.ts` | Spawns agent processes, manages env/sessions/skills |
+| `src/container-runner.ts` | Docker container execution for reviewer/arbiter |
+| `src/channels/discord.ts` | Discord channel (8s typing refresh, Whisper transcription) |
+| `src/ipc.ts` | IPC watcher and task processing |
+| `src/router.ts` | Message formatting and outbound routing |
+| `src/config.ts` | Trigger pattern, paths, intervals |
+| `src/task-scheduler.ts` | Runs scheduled tasks |
+| `src/db.ts` | SQLite operations |
+| `runners/agent-runner/` | Claude Code runner (Agent SDK) |
+| `runners/codex-runner/` | Codex runner (SDK, `codex exec` wrapper) |
+| `groups/{name}/CLAUDE.md` | Per-group memory (isolated) |
--- a/docs/configuration.md
+++ b/docs/configuration.md
@@ -0,0 +1,83 @@
+# Configuration
+
+All configuration in a single `.env` file.
+
+## Discord Bots
+
+```bash
+DISCORD_BOT_TOKEN=               # Claude bot
+DISCORD_CODEX_BOT_TOKEN=         # Codex-main bot (owner)
+DISCORD_REVIEW_BOT_TOKEN=        # Codex-review bot (arbiter)
+```
+
+## Agent Types & Models
+
+```bash
+# Agent types per role
+OWNER_AGENT_TYPE=codex            # codex | claude-code
+REVIEWER_AGENT_TYPE=claude-code   # claude-code | codex
+ARBITER_AGENT_TYPE=codex          # codex | claude-code (optional, enables 3rd agent)
+
+# Per-role model overrides
+OWNER_MODEL=gpt-5.4
+REVIEWER_MODEL=claude-opus-4-6
+ARBITER_MODEL=gpt-5.4
+
+# Per-role effort level
+OWNER_EFFORT=xhigh
+REVIEWER_EFFORT=high
+ARBITER_EFFORT=xhigh
+
+# Per-role fallback toggle
+REVIEWER_FALLBACK_ENABLED=true    # Auto-handoff to codex when Claude exhausted
+ARBITER_FALLBACK_ENABLED=false
+
+# Response language
+AGENT_LANGUAGE=Korean             # Injected into all agent prompts
+```
+
+## Authentication
+
+```bash
+CLAUDE_CODE_OAUTH_TOKEN=          # Claude Code OAuth token
+CLAUDE_CODE_OAUTH_TOKENS=         # Comma-separated for multi-account rotation
+OPENAI_API_KEY=                   # Codex API key (if not using OAuth)
+```
+
+## Voice Transcription
+
+```bash
+GROQ_API_KEY=                     # Groq whisper-large-v3-turbo (primary)
+OPENAI_API_KEY=                   # OpenAI whisper-1 (fallback)
+```
+
+## Mixture of Agents (MoA)
+
+```bash
+MOA_ENABLED=true
+MOA_REF_MODELS=kimi,glm           # Configurable reference model list
+
+# Per-model config (pattern: MOA_{NAME}_{SETTING})
+MOA_KIMI_MODEL=kimi-k2.5
+MOA_KIMI_BASE_URL=https://api.kimi.com/coding
+MOA_KIMI_API_KEY=sk-kimi-xxx
+MOA_KIMI_API_FORMAT=anthropic
+
+MOA_GLM_MODEL=glm-5.1
+MOA_GLM_BASE_URL=https://open.bigmodel.cn/api/anthropic
+MOA_GLM_API_KEY=xxx
+MOA_GLM_API_FORMAT=anthropic
+```
+
+## Debugging Paths
+
+| Item | Path |
+|------|------|
+| **DB** | `store/messages.db` (shared, WAL mode) |
+| Service log | `journalctl --user -u ejclaw -f` or `logs/ejclaw.log` |
+| Per-group logs | `groups/{name}/logs/` |
+| Claude sessions | `data/sessions/{name}/.claude/` |
+| Codex sessions | `data/sessions/{name}/.codex/` |
+| Claude platform rules | `prompts/claude-platform.md` |
+| Codex platform rules | `prompts/codex-platform.md` |
+| Global memory | `groups/global/CLAUDE.md` |
--- a/prompts/arbiter-paired-room.md
+++ b/prompts/arbiter-paired-room.md
@@ -22,6 +22,15 @@ You have been summoned because the owner and reviewer reached a deadlock after m
  - The situation cannot be resolved without user input, regardless of technical agreement
  - The same NEEDS_CONTEXT or BLOCKED is repeated after a prior PROCEED — this means your PROCEED did not resolve the issue

+## MoA (Mixture of Agents) Reference Opinions
+
+You may receive reference opinions from external models appended to your prompt. When present:
+
+- **Cite them explicitly** in your verdict — e.g., "Reference model A agrees that...", "Reference model B raises a concern about..."
+- **Cross-reference** their opinions against the owner/reviewer conversation and code evidence
+- **Resolve conflicts** — if reference opinions disagree with each other or with owner/reviewer, state which view you adopt and why
+- Do NOT blindly follow reference opinions — they are inputs to your judgment, not authorities
+
 ## Rules

 - Base your verdict on evidence (code, test output, logs), not on who said what first
--- a/src/agent-error-detection.ts
+++ b/src/agent-error-detection.ts
@@ -102,7 +102,8 @@ export type AgentTriggerReason =
  | 'session-failure'
  | 'overloaded'
  | 'network-error'
-  | 'success-null-result';
+  | 'success-null-result'
+  | 'session-failure';

 export type ClaudeRotationReason = Extract<
  AgentTriggerReason,
--- a/src/agent-runner.ts
+++ b/src/agent-runner.ts
@@ -97,6 +97,26 @@ export async function runAgentProcess(
        memoryBriefing: input.memoryBriefing,
        role: containerRole,
      });
+      // For codex: also write AGENTS.md to the reviewer session dir, because
+      // the container's /home/node/.claude always mounts the reviewer session.
+      // Arbiter and reviewer never run simultaneously, so this is safe.
+      if (containerRole === 'arbiter') {
+        const reviewerSessionDir = path.join(
+          path.dirname(sessionDir),
+          `${group.folder}-reviewer`,
+        );
+        const reviewerCodexDir = path.join(reviewerSessionDir, '.codex');
+        if (fs.existsSync(reviewerSessionDir)) {
+          fs.mkdirSync(reviewerCodexDir, { recursive: true });
+          const arbiterAgentsMd = path.join(sessionDir, '.codex', 'AGENTS.md');
+          if (fs.existsSync(arbiterAgentsMd)) {
+            fs.copyFileSync(
+              arbiterAgentsMd,
+              path.join(reviewerCodexDir, 'AGENTS.md'),
+            );
+          }
+        }
+      }
    }

    return runReviewerContainer({
--- a/src/container-runner.ts
+++ b/src/container-runner.ts
@@ -321,6 +321,28 @@ export function buildReviewerMounts(
    readonly: false,
  });

+  // Owner session directory: read-only so reviewer can verify runtime state
+  // files (cron state, configs, etc.) that the owner references by absolute path.
+  const ownerSessionDir = path.join(DATA_DIR, 'sessions', group.folder);
+  if (fs.existsSync(ownerSessionDir)) {
+    mounts.push({
+      hostPath: ownerSessionDir,
+      containerPath: ownerSessionDir,
+      readonly: true,
+    });
+  }
+
+  // Codex OAuth: mount host's ~/.codex (read-only) so codex-runner can authenticate
+  const hostCodexHome =
+    process.env.CODEX_HOME || path.join(os.homedir(), '.codex');
+  if (fs.existsSync(hostCodexHome)) {
+    mounts.push({
+      hostPath: hostCodexHome,
+      containerPath: '/home/node/.codex',
+      readonly: true,
+    });
+  }
+
  return mounts;
 }

@@ -367,12 +389,20 @@ function buildCreateArgs(
    args.push('-e', `SENTRY_AUTH_TOKEN=${process.env.SENTRY_AUTH_TOKEN}`);
  }

-  // Extra env overrides from paired-execution-context
+  // Extra env overrides from paired-execution-context.
+  // Model/effort keys are excluded here — they are injected per-exec so the
+  // correct agent-type-specific values are used (claude vs codex).
+  const perExecKeys = new Set([
+    'CLAUDE_MODEL',
+    'CLAUDE_EFFORT',
+    'CODEX_MODEL',
+    'CODEX_EFFORT',
+  ]);
  if (envOverrides) {
    for (const [key, value] of Object.entries(envOverrides)) {
-      // Already set above — skip duplicates
      if (key === 'ANTHROPIC_API_KEY' || key === 'CLAUDE_CODE_OAUTH_TOKEN')
        continue;
+      if (perExecKeys.has(key)) continue;
      if (key === 'EJCLAW_WORK_DIR') {
        args.push('-e', 'EJCLAW_WORK_DIR=/workspace/project');
        continue;
@@ -469,22 +499,38 @@ export async function runReviewerContainer(args: {
      '';
    execArgs.push('-e', `CLAUDE_CODE_OAUTH_TOKEN=${oauthToken}`);
  }
-  // Inject model/effort overrides at exec-time so per-role config
-  // takes effect even with persistent containers.
-  const modelEnvKeys = [
-    'CLAUDE_MODEL',
-    'CLAUDE_EFFORT',
-    'CODEX_MODEL',
-    'CODEX_EFFORT',
-  ];
+  const isCodexAgent = (group.agentType || 'claude-code') === 'codex';
+  logger.info(
+    {
+      containerName,
+      groupAgentType: group.agentType,
+      isCodexAgent,
+      runnerPath: isCodexAgent
+        ? '/app/codex/dist/index.js'
+        : '/app/agent/dist/index.js',
+    },
+    'Container exec runner selection',
+  );
+  // Inject only agent-type-appropriate model/effort overrides per exec.
  if (envOverrides) {
+    const modelEnvKeys = isCodexAgent
+      ? ['CODEX_MODEL', 'CODEX_EFFORT']
+      : ['CLAUDE_MODEL', 'CLAUDE_EFFORT'];
    for (const key of modelEnvKeys) {
      if (envOverrides[key]) {
        execArgs.push('-e', `${key}=${envOverrides[key]}`);
      }
    }
  }
-  execArgs.push(containerName, 'bun', '/app/dist/index.js');
+  if (isCodexAgent) {
+    // Use session-local .codex dir (contains AGENTS.md with role prompts)
+    // instead of the host-mounted ~/.codex (which has owner-only config).
+    execArgs.push('-e', 'CODEX_HOME=/home/node/.claude/.codex');
+  }
+  const runnerPath = isCodexAgent
+    ? '/app/codex/dist/index.js'
+    : '/app/agent/dist/index.js';
+  execArgs.push(containerName, 'bun', runnerPath);

  return new Promise<AgentOutput>((resolve) => {
    const proc = spawn(CONTAINER_RUNTIME_BIN, execArgs, {
--- a/src/unified-dashboard.ts
+++ b/src/unified-dashboard.ts
@@ -591,6 +591,7 @@ function buildModelConfigSection(): string {
    },
  ];

+  const failover = getGlobalFailoverInfo();
  const lines = ['🤖 *모델 구성*'];
  for (const role of roleConfigs) {
    if (!role.agentType && role.label === 'Arbiter') continue;
@@ -600,7 +601,15 @@ function buildModelConfigSection(): string {
        ? process.env.CODEX_MODEL || 'codex'
        : process.env.CLAUDE_MODEL || 'claude';
    const model = role.model || defaultModel;
-    lines.push(`  **${role.label}** — ${type} \`${model}\``);
+
+    // Show fallback status for claude-code roles when global failover is active
+    const isFallback = failover.active && type === 'claude-code';
+    if (isFallback) {
+      const fallbackModel = process.env.CODEX_MODEL || 'codex';
+      lines.push(`  **${role.label}** — codex \`${fallbackModel}\` (fallback)`);
+    } else {
+      lines.push(`  **${role.label}** — ${type} \`${model}\``);
+    }
  }

  // MoA status
@@ -612,8 +621,6 @@ function buildModelConfigSection(): string {
    lines.push(`  **MoA** — ${refs}`);
  }

-  // Global failover
-  const failover = getGlobalFailoverInfo();
  if (failover.active) {
    lines.push(`  ⚠️ **Failover 활성** — ${failover.reason || '알 수 없음'}`);
  }