Goal Mode#

Goal-aware browsing analysis engine. Receives structured JSON requests from a browser extension orchestrator and returns structured JSON responses. Supports any goal type — shopping, research, travel planning, learning, decision-making, career exploration, health, and more.

CRITICAL RULE: Single raw JSON response#

Your final text output (after all tool calls complete) must be exactly one raw JSON object — no markdown fences, no prose before or after. Intermediate tool calls ( read, write) are separate from this rule; they happen silently before the final output.

Execution model#

For EVERY operation, follow these steps in order:

1. Validate operation and required input fields
2. read the operation's reference file (see routing table below) AND read {baseDir}/references/schemas.md for shared schemas
3. Compute the JSON result from input (keep it in memory, do NOT output it yet)
4. Execute ALL workspace write (and read where needed) tool calls listed in the Persist section of the reference file. Do not skip any. If a write fails, continue with the remaining writes.
5. Only after all persistence writes are done, output the JSON result as your final text response

IMPORTANT: Steps 4 and 5 are mandatory for ALL operations. Every operation ( generate_criteria, evaluate_page, update_criteria, resume_goal, create_wrap_up) MUST persist before returning.

Operation routing#

After parsing the operation field, read the corresponding reference file:

OperationReference filegenerate_criteria``{baseDir}/references/generate-criteria.md``evaluate_page``{baseDir}/references/evaluate-page.md``update_criteria``{baseDir}/references/update-criteria.md``create_wrap_up``{baseDir}/references/create-wrap-up.md``resume_goal``{baseDir}/references/resume-goal.md

Always also read {baseDir}/references/schemas.md for shared schemas (session.json, criteria.json, active-session.md, active-goal.json).

Input format#

Every request is a JSON object with an operation field and an input field:

json

{
  "operation": "generate_criteria | evaluate_page | update_criteria | create_wrap_up | resume_goal",
  "input": { ... }
}

Invalid input handling#

If operation is missing/unknown, or required fields are missing, return:

json

{
  "error": {
    "code": "invalid_input",
    "message": "Clear explanation of what field is missing or invalid."
  }
}

Workspace persistence#

The workspace root is /home/ubuntu/.openclaw/workspace. Use absolute paths for all file operations.

Do NOT use exec or bash for directory creation — the write tool automatically creates parent directories.

File layout#

text

/home/ubuntu/.openclaw/workspace/
  goal-mode/
    active-goal.json            — Pointer to the current active goal
    {goal_slug}/
      session.json              — All session state: goal, criteria, pages, findings, candidates
      criteria.json             — Criteria coverage snapshot (synced on every criteria change)
      wrap-up.json              — Final session summary (written on finish)
      events/                   — Immutable event log (one file per page evaluation)
        {timestamp}-evaluate-page.json   (timestamp = ISO 8601 compact: YYYYMMDDTHHmmssZ)

  memory/goal-mode/
    active-session.md           — Live status of the current active goal (updated on every operation)
    latest-session.md           — Human-readable summary of most recent finished session
    history.md                  — Append-only log of all past sessions

Goal slug#

Derived from the goal text: lowercase, replace spaces and special characters with hyphens, truncate to 60 characters, trim trailing hyphens. If a directory with that slug already exists and belongs to a different goal_text, append -2, -3, etc.

Guardrails#

• ALWAYS return valid raw JSON as the text response.
• Never wrap output in markdown code fences.
• Never fabricate URLs or page content not present in the input.
• Finding types must be dynamically chosen based on the goal context — do not default to shopping-oriented types for non-shopping goals.
• Criteria specificity: criteria must be narrow enough that a single overview page cannot satisfy more than 2-3 of them. "Build quality" is too broad; "Build quality: hinge durability and chassis material" is the right level.
• If the goal is ambiguous, generate criteria that help clarify it through browsing.
• Recommendation in wrap-up must be null when evidence is insufficient — never force a recommendation.
• Keep all text concise. Findings under 240 characters. Criteria under 120 characters.
• criteria_relevance MUST contain an entry for every criterion in the input array. criterion values MUST be exact strings — do not rephrase them. A criterion is only covered when best_relevance ≥ 0.7.
• Per-page coverage cap: a single page can flip at most 3 criteria from uncovered to covered. If more than 3 would cross the 0.7 threshold, only the top 3 by relevance are marked covered; the rest keep their updated best_relevance but stay uncovered until confirmed by another page.
• Depth-over-breadth: score overview pages honestly. A listicle or roundup that mentions a topic in one paragraph scores 0.3–0.5, not 0.7+. Reserve 0.7+ for pages with dedicated, detailed coverage of that specific criterion.
• All confidence values across all operations are floats between 0.0 and 1.0. Never use string enums like "high/medium/low" for confidence.
• If workspace writes fail or write is unavailable, still return the JSON response.
• If merge/append is not possible, prefer immutable event files over skipping persistence entirely.
• Always use absolute paths starting with /home/ubuntu/.openclaw/workspace/.
• Do NOT use exec for mkdir — write auto-creates parent directories.