OpenClaw-Setup/openclaw-setup-max/PRD.md

# PRD: OpenClaw Local Model Cost Control (Setup Max)

## 1) Document Purpose

Define requirements for a reusable OpenClaw setup that supports:
- Instant manual switching between paid and free model profiles
- Optional scheduled switching (day/night)
- Optional budget protection for expensive models

Audience:
- Primary: local operator (non-expert friendly)
- Secondary: future maintainer cloning this setup for another machine

## 2) Problem Statement

Running a paid model 24/7 can create unnecessary spend.  
Operators need an easy way to:
- Switch to free local models on demand
- Automate free mode during off-hours
- Prevent accidental long-running use of expensive models

## 3) Goals

- G1: One-command live switch between `paid` and `free` profiles.
- G2: Support day/night automatic profile switching.
- G3: Provide guardrail warning + optional auto-revert from high-cost model usage.
- G4: Make setup operable by a newbie using copy/paste commands.

## 4) Non-Goals

- Building a new OpenClaw provider plugin.
- Replacing OpenClaw internal routing.
- Managing billing dashboards directly.

## 5) User Stories

- As an operator, I can run one command to switch to free local models immediately.
- As an operator, I can switch back to paid model for deep tasks.
- As an operator, I can set free mode from 9pm-7am automatically.
- As an operator, I get warned if expensive model remains active too long.
- As a maintainer, I can copy this folder to another Mac and follow docs to bootstrap.

## 6) Functional Requirements

### FR-1 Profile Configuration
- System must define named model profiles in JSON:
  - `paid`
  - `free`
- Each profile must include:
  - `primary` model
  - ordered fallback list

Implementation:
- `config/model-profiles.config.json`

### FR-2 Manual Live Switch
- Must support:
  - `model_profile_switch.sh paid`
  - `model_profile_switch.sh free`
- Must:
  - update OpenClaw primary model
  - reset + apply fallback list
  - optionally push `/model <primary>` to active session key
- Must support flags:
  - `--no-live`
  - `--no-status`
  - `status`

Implementation:
- `scripts/model_profile_switch.sh`

### FR-3 Schedule Guard
- Must read schedule config and apply profile by hour.
- Must support day and night windows.
- Must run idempotently and avoid unnecessary switches.
- Must correct drift if current default model does not match expected scheduled profile.

Implementation:
- `config/model-schedule.config.json`
- `scripts/model_schedule_guard.sh`
- `scripts/install_model_schedule_guard_launchd.sh`

### FR-4 Budget Guard
- Must detect if active session model is in high-cost allowlist.
- Must warn after configurable threshold.
- Must auto-revert after configurable threshold.
- Must throttle warnings/reverts to avoid spam loops.

Implementation:
- `config/model-budget-guard.config.json`
- `scripts/model_budget_guard.sh`
- `scripts/install_model_budget_guard_launchd.sh`

### FR-5 Install Convenience
- Must provide one script to install both schedule and budget guards.

Implementation:
- `scripts/install_local_model_guardrails.sh`

### FR-6 Documentation
- Must include a step-by-step README for first-time users.
- Must include verification commands and disable/uninstall steps.

Implementation:
- `README.md`

## 7) Non-Functional Requirements

- NFR-1 Reliability: scripts should be safe to run repeatedly.
- NFR-2 Observability: each LaunchAgent writes stdout/stderr logs to `/tmp`.
- NFR-3 Compatibility: target macOS + zsh/bash + OpenClaw CLI.
- NFR-4 Safety: avoid exposing secrets in script output.
- NFR-5 Usability: all common actions available as copy/paste commands.

## 8) Architecture Overview

Inputs:
- OpenClaw session state (`~/.openclaw/agents/main/sessions/sessions.json`)
- Profile/schedule/budget JSON configs

Executors:
- Manual CLI script (`model_profile_switch.sh`)
- Launchd workers (`model_schedule_guard.sh`, `model_budget_guard.sh`)

Outputs:
- Updated OpenClaw model + fallbacks in config
- Optional live `/model` message into active session
- Log files in `/tmp`

## 9) Operational Flow

### Flow A: Manual Switch
1. Operator runs switch command.
2. Script loads profile from config.
3. Script applies primary + fallbacks.
4. Script pushes live model command to active session (unless `--no-live`).
5. Script prints status.

### Flow B: Scheduled Switch
1. LaunchAgent triggers every 5 minutes.
2. Guard computes desired profile from local hour.
3. Guard checks last applied profile + current default model.
4. If mismatch, applies desired profile.
5. Guard stores state timestamp.

### Flow C: Budget Protection
1. LaunchAgent triggers every 2 minutes.
2. Guard inspects active session model.
3. If model is high-cost:
   - warn after threshold
   - auto-revert after threshold
4. Guard updates state file for rate-limiting.

## 10) Acceptance Criteria

- AC-1: Running `bash ./scripts/model_profile_switch.sh free` changes default to `ollama/qwen3:14b`.
- AC-2: Running `bash ./scripts/model_profile_switch.sh paid` restores default to `moonshot/kimi-k2.5`.
- AC-3: With schedule enabled (`21`/`7`), profile changes correctly by local time window.
- AC-4: If operator manually drifts model during scheduled window, schedule guard re-aligns on next run.
- AC-5: Budget guard warns and reverts when high model remains active beyond thresholds.
- AC-6: New operator can follow README from zero knowledge and complete setup without editing scripts.

## 11) Risks and Mitigations

- Risk: model names differ across machines.
  - Mitigation: update `config/model-profiles.config.json` per machine.
- Risk: LaunchAgents not loaded after reboot/login changes.
  - Mitigation: include explicit `launchctl print` checks and reinstall command.
- Risk: missing CLI dependencies (`jq`, `openclaw`).
  - Mitigation: prerequisite checks in README.
- Risk: user confusion between ChatGPT subscription and API billing.
  - Mitigation: clear FAQ note in README.

## 12) Rollout Plan

1. Validate scripts (`bash -n` and status commands).
2. Perform manual switch tests (`free` then `paid`).
3. Enable schedule (`enabled=true`) and install schedule LaunchAgent.
4. Install budget LaunchAgent.
5. Monitor logs for one day.
6. Clone/copy to future machine and adjust profile model IDs as needed.

## 13) Future Enhancements

- Optional quiet-hours policy to disable non-essential notifications.
- Per-channel session key mapping (Telegram/Discord-specific switching).
- Daily spend estimator from model usage telemetry.
- Optional provider locking profile policies.