harness（harness 文档）

Harness — Long-Running Agent Framework

Executable protocol enabling any agent task to run continuously across multiple sessions with automatic progress recovery, task dependency resolution, failure rollback, and standardized error handling.

Design Principles

1. Design for the agent, not the human — Test output, docs, and task structure are the agent's primary interface
2. Progress files ARE the context — When context window resets, progress files + git history = full recovery
3. Premature completion is the #1 failure mode — Structured task lists with explicit completion criteria prevent declaring victory early
4. Standardize everything grep-able — ERROR on same line, structured timestamps, consistent prefixes
5. Fast feedback loops — Pre-compute stats, run smoke tests before full validation
6. Idempotent everything — Init scripts, task execution, environment setup must all be safe to re-run
7. Fail safe, not fail silent — Every failure must have an explicit recovery strategy

Commands

/harness init <project-path>     # Initialize harness files in project
/harness run                     # Start/resume the infinite loop
/harness status                  # Show current progress and stats
/harness add "task description"  # Add a task to the list

Activation Marker

Hooks only take effect when .harness-active marker file exists in the harness root (same directory as harness-tasks.json).

• /harness init and /harness run MUST create this marker: touch <project-path>/.harness-active
• When all tasks complete (no pending/in_progress/retryable left), remove it: rm <project-path>/.harness-active
• Without this marker, all hooks are no-ops — they exit 0 immediately

Progress Persistence (Dual-File System)

Maintain two files in the project working directory:

harness-progress.txt (Append-Only Log)

Free-text log of all agent actions across sessions. Never truncate.

[2025-07-01T10:00:00Z] [SESSION-1] INIT Harness initialized for project /path/to/project
[2025-07-01T10:00:05Z] [SESSION-1] INIT Environment health check: PASS
[2025-07-01T10:00:10Z] [SESSION-1] LOCK acquired (pid=12345)
[2025-07-01T10:00:11Z] [SESSION-1] Starting [task-001] Implement user authentication (base=def5678)
[2025-07-01T10:05:00Z] [SESSION-1] CHECKPOINT [task-001] step=2/4 "auth routes created, tests pending"
[2025-07-01T10:15:30Z] [SESSION-1] Completed [task-001] (commit abc1234)
[2025-07-01T10:15:31Z] [SESSION-1] Starting [task-002] Add rate limiting (base=abc1234)
[2025-07-01T10:20:00Z] [SESSION-1] ERROR [task-002] [TASK_EXEC] Redis connection refused
[2025-07-01T10:20:01Z] [SESSION-1] ROLLBACK [task-002] git reset --hard abc1234
[2025-07-01T10:20:02Z] [SESSION-1] STATS tasks_total=5 completed=1 failed=1 pending=3 blocked=0 attempts_total=2 checkpoints=1

harness-tasks.json (Structured State)

{
  "version": 2,
  "created": "2025-07-01T10:00:00Z",
  "session_config": {
    "concurrency_mode": "exclusive",
    "max_tasks_per_session": 20,
    "max_sessions": 50
  },
  "tasks": [
    {
      "id": "task-001",
      "title": "Implement user authentication",
      "status": "completed",
      "priority": "P0",
      "depends_on": [],
      "attempts": 1,
      "max_attempts": 3,
      "started_at_commit": "def5678",
      "validation": {
        "command": "npm test -- --testPathPattern=auth",
        "timeout_seconds": 300
      },
      "on_failure": {
        "cleanup": null
      },
      "error_log": [],
      "checkpoints": [],
      "completed_at": "2025-07-01T10:15:30Z"
    },
    {
      "id": "task-002",
      "title": "Add rate limiting",
      "status": "failed",
      "priority": "P1",
      "depends_on": [],
      "attempts": 1,
      "max_attempts": 3,
      "started_at_commit": "abc1234",
      "validation": {
        "command": "npm test -- --testPathPattern=rate-limit",
        "timeout_seconds": 120
      },
      "on_failure": {
        "cleanup": "docker compose down redis"
      },
      "error_log": ["[TASK_EXEC] Redis connection refused"],
      "checkpoints": [],
      "completed_at": null
    },
    {
      "id": "task-003",
      "title": "Add OAuth providers",
      "status": "pending",
      "priority": "P1",
      "depends_on": ["task-001"],
      "attempts": 0,
      "max_attempts": 3,
      "started_at_commit": null,
      "validation": {
        "command": "npm test -- --testPathPattern=oauth",
        "timeout_seconds": 180
      },
      "on_failure": {
        "cleanup": null
      },
      "error_log": [],
      "checkpoints": [],
      "completed_at": null
    }
  ],
  "session_count": 1,
  "last_session": "2025-07-01T10:20:02Z"
}

Task statuses: pending → in_progress (transient, set only during active execution) → completed or failed. A task found as in_progress at session start means the previous session was interrupted — handle via Context Window Recovery Protocol.

In concurrent mode (see Concurrency Control), tasks may also carry claim metadata: claimed_by and lease_expires_at (ISO timestamp).

Session boundary: A session starts when the agent begins executing the Session Start protocol and ends when a Stopping Condition is met or the context window resets. Each session gets a unique SESSION-N identifier (N = session_count after increment).

Concurrency Control

Before modifying harness-tasks.json, acquire an exclusive lock using portable mkdir (atomic on all POSIX systems, works on both macOS and Linux):

# Acquire lock (fail fast if another agent is running)
# Lock key must be stable even if invoked from a subdirectory.
ROOT="$PWD"
SEARCH="$PWD"
while [ "$SEARCH" != "/" ] && [ ! -f "$SEARCH/harness-tasks.json" ]; do
  SEARCH="$(dirname "$SEARCH")"
done
if [ -f "$SEARCH/harness-tasks.json" ]; then
  ROOT="$SEARCH"
fi

PWD_HASH="$(
  printf '%s' "$ROOT" |
    (shasum -a 256 2>/dev/null || sha256sum 2>/dev/null) |
    awk '{print $1}' |
    cut -c1-16
)"
LOCKDIR="/tmp/harness-${PWD_HASH:-unknown}.lock"
if ! mkdir "$LOCKDIR" 2>/dev/null; then
  # Check if lock holder is still alive
  LOCK_PID=$(cat "$LOCKDIR/pid" 2>/dev/null)
  if [ -n "$LOCK_PID" ] && kill -0 "$LOCK_PID" 2>/dev/null; then
    echo "ERROR: Another harness session is active (pid=$LOCK_PID)"; exit 1
  fi
  # Stale lock — atomically reclaim via mv to avoid TOCTOU race
  STALE="$LOCKDIR.stale.$$"
  if mv "$LOCKDIR" "$STALE" 2>/dev/null; then
    rm -rf "$STALE"
    mkdir "$LOCKDIR" || { echo "ERROR: Lock contention"; exit 1; }
    echo "WARN: Removed stale lock${LOCK_PID:+ from pid=$LOCK_PID}"
  else
    echo "ERROR: Another agent reclaimed the lock"; exit 1
  fi
fi
echo "$$" > "$LOCKDIR/pid"
trap 'rm -rf "$LOCKDIR"' EXIT

Log lock acquisition: [timestamp] [SESSION-N] LOCK acquired (pid=<PID>) Log lock release: [timestamp] [SESSION-N] LOCK released

Modes:

• Exclusive (default): hold the lock for the entire session (the trap EXIT handler releases it automatically). Any second session in the same state root fails fast.
• Concurrent (opt-in via session_config.concurrency_mode: "concurrent"): treat this as a state transaction lock. Hold it only while reading/modifying/writing harness-tasks.json (including .bak/.tmp) and appending to harness-progress.txt. Release it immediately before doing real work.

Concurrent mode invariants:

• All workers MUST point at the same state root (the directory that contains harness-tasks.json). If you are using separate worktrees/clones, pin it explicitly (e.g., HARNESS_STATE_ROOT=/abs/path/to/state-root).
• Task selection is advisory; the real gate is atomic claim under the lock: set status="in_progress", set claimed_by (stable worker id, e.g., HARNESS_WORKER_ID), set lease_expires_at. If claim fails (already in_progress with a valid lease), pick another eligible task and retry.
• Never run two workers in the same git working directory. Use separate worktrees/clones. Otherwise rollback (git reset --hard / git clean -fd) will destroy other workers.

Infinite Loop Protocol

Session Start (Execute Every Time)

1. Read state: Read last 200 lines of harness-progress.txt + full harness-tasks.json. If JSON is unparseable, see JSON corruption recovery in Error Handling.
2. Read git: Run git log --oneline -20 and git diff --stat to detect uncommitted work
3. Acquire lock (mode-dependent): Exclusive mode fails if another session is active. Concurrent mode uses the lock only for state transactions.
4. Recover interrupted tasks (see Context Window Recovery below)
5. Health check: Run harness-init.sh if it exists
6. Track session: Increment session_count in JSON. Check session_count against max_sessions — if reached, log STATS and STOP. Initialize per-session task counter to 0.
7. Pick next task using Task Selection Algorithm below

Task Selection Algorithm

Before selecting, run dependency validation:

1. Cycle detection: For each non-completed task, walk depends_on transitively. If any task appears in its own chain, mark it failed with [DEPENDENCY] Circular dependency detected: task-A -> task-B -> task-A. Self-references (depends_on includes own id) are also cycles.
2. Blocked propagation: If a task's depends_on includes a task that is failed and will never be retried (either attempts >= max_attempts OR its error_log contains a [DEPENDENCY] entry), mark the blocked task as failed with [DEPENDENCY] Blocked by failed task-XXX. Repeat until no more tasks can be propagated.

Then pick the next task in this priority order:

1. Tasks with status: "pending" where ALL depends_on tasks are completed — sorted by priority (P0 > P1 > P2), then by id (lowest first)
2. Tasks with status: "failed" where attempts < max_attempts and ALL depends_on are completed — sorted by priority, then oldest failure first
3. If no eligible tasks remain → log final STATS → STOP

Task Execution Cycle

For each task, execute this exact sequence:

1. Claim (atomic, under lock): Record started_at_commit = current HEAD hash. Set status to in_progress, set claimed_by, set lease_expires_at, log Starting [<task-id>] <title> (base=<hash>). If the task is already claimed (in_progress with a valid lease), pick another eligible task and retry.
2. Execute with checkpoints: Perform the work. After each significant step, log: [timestamp] [SESSION-N] CHECKPOINT [task-id] step=M/N "description of what was done" Also append to the task's checkpoints array: {"step": M, "total": N, "description": "...", "timestamp": "ISO"}. In concurrent mode, renew the lease at each checkpoint (push lease_expires_at forward).
3. Validate: Run the task's validation.command with a timeout wrapper (prefer timeout; on macOS use gtimeout from coreutils). If validation.command is empty/null, log ERROR [<task-id>] [CONFIG] Missing validation.command and STOP — do not declare completion without an objective check. Before running, verify the command exists (e.g., command -v <binary>) — if missing, treat as ENV_SETUP error.

- Command exits 0 → PASS - Command exits non-zero → FAIL - Command exceeds timeout → TIMEOUT

1. Record outcome:

- Success: status=completed, set completed_at, log Completed [<task-id>] (commit <hash>), git commit - Failure: increment attempts, append error to error_log. Verify started_at_commit exists via git cat-file -t <hash> — if missing, mark failed at max_attempts. Otherwise execute git reset --hard <started_at_commit> and git clean -fd to rollback ALL commits and remove untracked files. Execute on_failure.cleanup if defined. Log ERROR [<task-id>] [<category>] <message>. Set status=failed (Task Selection Algorithm pass 2 handles retries when attempts < max_attempts)

1. Track: Increment per-session task counter. If max_tasks_per_session reached, log STATS and STOP.
2. Continue: Immediately pick next task (zero idle time)

Stopping Conditions

• All tasks completed
• All remaining tasks failed at max_attempts or blocked by failed dependencies
• session_config.max_tasks_per_session reached for this session
• session_config.max_sessions reached across all sessions
• User interrupts

Context Window Recovery Protocol

When a new session starts and finds a task with status: "in_progress":

• Exclusive mode: treat this as an interrupted previous session and run the Recovery Protocol below.
• Concurrent mode: only recover a task if either (a) claimed_by matches this worker, or (b) lease_expires_at is in the past (stale lease). Otherwise, treat it as owned by another worker and do not modify it.

1. Check git state: git diff --stat # Uncommitted changes? git log --oneline -5 # Recent commits since task started? git stash list # Any stashed work?
2. Check checkpoints: Read the task's checkpoints array to determine last completed step
3. Decision matrix (verify recent commits belong to this task by checking commit messages for the task-id):

1. Log recovery: [timestamp] [SESSION-N] RECOVERY [task-id] action="<action taken>" reason="<reason>"

Error Handling & Recovery Strategies

Each error category has a default recovery strategy:

JSON corruption: If harness-tasks.json cannot be parsed, check for harness-tasks.json.bak (written before each modification). If backup exists and is valid, restore from it. If no valid backup, log ERROR [ENV_SETUP] harness-tasks.json corrupted and unrecoverable and STOP — task metadata (validation commands, dependencies, cleanup) cannot be reconstructed from logs alone.

Backup protocol: Before every write to harness-tasks.json, copy the current file to harness-tasks.json.bak. Write updates atomically: write JSON to harness-tasks.json.tmp then mv it into place (readers should never see a partial file).

Environment Initialization

If harness-init.sh exists in the project root, run it at every session start. The script must be idempotent.

Example harness-init.sh:

#!/bin/bash
set -e
npm install 2>/dev/null || pip install -r requirements.txt 2>/dev/null || true
curl -sf http://localhost:5432 >/dev/null 2>&1 || echo "WARN: DB not reachable"
npm test -- --bail --silent 2>/dev/null || echo "WARN: Smoke test failed"
echo "Environment health check complete"

Standardized Log Format

All log entries use grep-friendly format on a single line:

[ISO-timestamp] [SESSION-N] <TYPE> [task-id]? [category]? message

[task-id] and [category] are included when applicable (task-scoped entries). Session-level entries (INIT, LOCK, STATS) omit them.

Types: INIT, Starting, Completed, ERROR, CHECKPOINT, ROLLBACK, RECOVERY, STATS, LOCK, WARN

Error categories: ENV_SETUP, CONFIG, TASK_EXEC, TEST_FAIL, TIMEOUT, DEPENDENCY, SESSION_TIMEOUT

Filtering:

grep "ERROR" harness-progress.txt                    # All errors
grep "ERROR" harness-progress.txt | grep "TASK_EXEC" # Execution errors only
grep "SESSION-3" harness-progress.txt                # All session 3 activity
grep "STATS" harness-progress.txt                    # All session summaries
grep "CHECKPOINT" harness-progress.txt               # All checkpoints
grep "RECOVERY" harness-progress.txt                 # All recovery actions

Session Statistics

At session end, update harness-tasks.json: set last_session to current timestamp. (Do NOT increment session_count here — it is incremented at Session Start.) Then append:

[timestamp] [SESSION-N] STATS tasks_total=10 completed=7 failed=1 pending=2 blocked=0 attempts_total=12 checkpoints=23

blocked is computed at stats time: count of pending tasks whose depends_on includes a permanently failed task. It is not a stored status value.

Init Command (/harness init)

1. Create harness-progress.txt with initialization entry
2. Create harness-tasks.json with empty task list and default session_config
3. Optionally create harness-init.sh template (chmod +x)
4. Ask user: add harness files to .gitignore?

Status Command (/harness status)

Read harness-tasks.json and harness-progress.txt, then display:

1. Task summary: count by status (completed, failed, pending, blocked). blocked = pending tasks whose depends_on includes a permanently failed task (computed, not a stored status).
2. Per-task one-liner: [status] task-id: title (attempts/max_attempts)
3. Last 5 lines from harness-progress.txt
4. Session count and last session timestamp

Does NOT acquire the lock (read-only operation).

Add Command (/harness add)

Append a new task to harness-tasks.json with auto-incremented id (task-NNN), status pending, default max_attempts: 3, empty depends_on, and no validation command (required before the task can be completed). Prompt user for optional fields: priority, depends_on, validation.command, timeout_seconds. Requires lock acquisition (modifies JSON).

Tool Dependencies

Requires: Bash, file read/write, git. All harness operations must be executed from the project root directory. Does NOT require: specific MCP servers, programming languages, or test frameworks.

Concurrent mode requires isolated working directories (git worktree or separate clones). Do not run concurrent workers in the same working tree.