driftwatch-oc（漂移观察 OC）

Driftwatch — Workspace Health Scanner

Running a Scan

python3 {baseDir}/scripts/scan.py --workspace <workspace_path>

The --workspace argument is optional. If omitted, Driftwatch checks the OPENCLAW_WORKSPACE environment variable, then falls back to ~/.openclaw/workspace/. Most of the time just run:

python3 {baseDir}/scripts/scan.py

Output is JSON. Parse it, then present findings conversationally — never dump raw JSON at the operator.

Agent Delivery Rules

Always generate the HTML report when running a scan:

mkdir -p ~/.driftwatch/reports
python3 {baseDir}/scripts/scan.py --html ~/.driftwatch/reports/driftwatch-$(date +%Y-%m-%dT%H%M).html

Reports are stored with timestamps in ~/.driftwatch/reports/ so the operator can review past scans over time. Never overwrite — each scan gets its own file.

After generating:

1. On messaging surfaces (Telegram, Discord, Signal): send the HTML file to the user, then follow with a 1-3 sentence text summary
2. In terminal/CLI: print the file path so the operator can open it in a browser

The text summary covers: warning/danger/truncation counts (or "all clear"), notable changes since last scan, one actionable callout if needed. The HTML is the proof — always send both.

Exception: --check mode (cron) skips HTML unless --html is also passed explicitly.

What the Scanner Checks

Four analysis modules run in sequence:

truncation — Measures every bootstrap file's character count against the 20,000-char per-file limit and the 150,000-char aggregate budget. Tracks sequential budget consumption so you can see when MEMORY.md (last in injection order) is getting starved.

compaction — Checks whether AGENTS.md contains two recommended anchor sections: ## Session Startup and ## Red Lines. These are workspace conventions (not source-enforced) that ensure critical instructions are always present. Verifies each is present and within a 3,000-char budget. AGENTS.md is re-injected every turn as a bootstrap file — these sections matter as best-practice conventions, not because OpenClaw source code references them.

hygiene — Checks for duplicate memory files, empty bootstrap slots, missing subagent-required files, and stray markdown files the operator may think are being loaded but aren't.

simulation — Maps which lines and sections fall inside the truncation danger zone. For files over 18,000 chars, it shows exactly what content would be cut first. For files already over 20,000 chars, it shows what's being cut *right now*.

Severity Levels

• critical — address immediately. Something is broken or will break.
• warning — review soon. Not broken yet, but trending bad.
• info — awareness only. Nothing's wrong, just worth knowing.

Presenting Findings

Lead with critical findings. If there are none, say so first — that's the good news. Then work through warnings and info grouped by module.

Translate numbers into meaning. Don't say "char_count: 18500, percent_of_limit: 92.5". Say: "AGENTS.md is at 18,500 characters — 92% of its 20,000-char limit. If it grows much more, content near the bottom will start getting cut silently."

What not to do: Don't modify any files. Don't attempt to auto-fix anything. Present findings and let the operator decide what to change.

Flags Reference

--save — Persist scan for drift tracking

python3 {baseDir}/scripts/scan.py --save

Saves the scan result to ~/.driftwatch/history/ as a timestamped JSON file. Run with --save regularly (daily cron is ideal) to build a history baseline. Nothing happens if the directory doesn't exist — it's created automatically.

When to suggest this: When the operator asks about trends, when AGENTS.md is growing, or when you want to establish a baseline for future comparisons.

--history — Show drift trends from saved scans

python3 {baseDir}/scripts/scan.py --history

Loads stored scan history and adds a trends section to the output showing per-file growth rates and days-to-limit projections. Requires at least one prior --save scan to produce trend data.

The common combo: --save --history — saves today's scan and shows trends based on all prior saves.

When presenting trends: "AGENTS.md has been growing at about 150 chars per day over the past two weeks. At that rate it hits its 20,000-char limit in roughly 12 days."

--html <path> — HTML report

python3 {baseDir}/scripts/scan.py --html ~/.driftwatch/reports/report.html

Generates a self-contained HTML report at the specified path — no external dependencies, no CDN calls. The file includes zone-based progress bars, truncation overlays, and danger zone callouts for at-risk files. This is the primary output format — see Agent Delivery Rules above.

--check — Cron-friendly alert mode

python3 {baseDir}/scripts/scan.py --check

Designed for automation and cron monitoring. Outputs a single summary line to stdout and exits with:

• 0 — all healthy
• 1 — warnings present
• 2 — criticals present

This is the only mode that uses non-zero exit codes. Normal scan mode always exits 0 — the agent interprets the JSON findings. --check is for cron jobs and shell scripts that watch exit codes.

Example one-line outputs:

✓ All clear — 8 files healthy, 18.6% aggregate budget used
⚠ Warning — AGENTS.md at 82% of limit
✗ Critical — MEMORY.md TRUNCATED

Thresholds (configurable in ~/.driftwatch/config.json):

• Per-file warning: 70% of limit (default)
• Per-file critical: 90% of limit
• Aggregate warning: 60%
• Aggregate critical: 80%
• Growth rate warning: 200 chars/day

Flag combinations:

• --check --save — the cron use case: check thresholds AND save a history point
• --check --json — one-line summary AND full JSON in stdout (both outputs)

--json — Full JSON with --check

python3 {baseDir}/scripts/scan.py --check --json

When --check is used alone, it suppresses the JSON blob and shows only the summary line. Add --json to get both the summary line and the full JSON report.

Truncation Simulation

The simulation module tells you exactly what content is at risk before truncation becomes a problem.

How it works:

OpenClaw keeps the first 14,000 chars (head) and last 4,000 chars (tail) of each bootstrap file. Everything between is cut. The "danger zone" is the gap between those boundaries.

When presenting simulation findings:

If a file is in the danger zone, name the specific sections at risk:

"AGENTS.md has 1,200 characters in the danger zone — lines 350 to 380. If this file grows past 20,000 chars, the ## Delegation Templates section would be the first content your agent loses. Consider moving that section earlier in the file or trimming it."

If a file is actively truncated:

"MEMORY.md is 25,000 characters — 5,000 over the limit. Your agent cannot see lines 290 through 450 right now. That's roughly 7,000 characters of context that's being silently cut every session."

If all files are safe:

"No truncation risk — all files are well under 18,000 characters. The danger zone simulation shows nothing at risk right now."

Suggested Workflows

Baseline + monitoring:

# Establish baseline
python3 {baseDir}/scripts/scan.py --save

# Daily cron check (in crontab: 0 8 * * *)
python3 {baseDir}/scripts/scan.py --check --save

# Weekly review with trends + HTML report
mkdir -p ~/.driftwatch/reports
python3 {baseDir}/scripts/scan.py --history --html ~/.driftwatch/reports/driftwatch-$(date +%Y-%m-%dT%H%M).html

OpenClaw cron integration — add to openclaw.json:

{
  "cron": [
    {
      "schedule": "0 8 * * *",
      "command": "python3 ~/.openclaw/skills/driftwatch/scripts/scan.py --check --save",
      "label": "driftwatch-daily"
    }
  ]
}

When the cron fires and exits non-zero, report the summary line to the operator. Don't be noisy about exit 0.

Sharing a report with teammates:

python3 {baseDir}/scripts/scan.py --history --html ~/.driftwatch/reports/workspace-health.html

Sample Output Interpretation

Here's what a healthy workspace looks like and how to present it:

{
  "summary": { "critical": 0, "warning": 1, "info": 2 },
  "truncation": {
    "files": [
      {
        "file": "AGENTS.md",
        "char_count": 9200,
        "limit": 20000,
        "percent_of_limit": 46.0,
        "status": "ok"
      }
    ],
    "aggregate": {
      "total_chars": 54000,
      "percent_of_aggregate": 36.0,
      "aggregate_status": "ok"
    }
  },
  "compaction": {
    "anchor_sections": [
      { "heading": "Session Startup", "found": true, "char_count": 1200, "status": "ok" },
      { "heading": "Red Lines", "found": true, "char_count": 800, "status": "ok" }
    ]
  },
  "hygiene": {
    "findings": [
      { "severity": "warning", "check": "empty_bootstrap", "message": "IDENTITY.md exists but is empty" }
    ]
  }
}

How to present this:

Workspace looks healthy — no critical issues. One thing to note: IDENTITY.md exists but is empty. It's taking up a bootstrap slot without contributing anything. Worth filling in or removing. Bootstrap files are using 54,000 of 150,000 characters (36% of aggregate budget) — plenty of room. AGENTS.md is at 46% of its individual limit, well clear of truncation territory. Both post-compaction anchor sections are present in AGENTS.md and within the 3,000-char cap.

Notes

• Character counts, not token counts. OpenClaw enforces char-based limits.
• Findings are stamped with the OpenClaw version tag they were calibrated against (currently 2026.03). If you're on a different version, limits may differ.
• The scanner makes zero network calls. Everything runs locally.
• History data is stored locally in ~/.driftwatch/ — never transmitted.
• Auto-retention pruning removes history files older than 90 days (configurable in ~/.driftwatch/config.json).