Wizard Interactive Assistant

Overview

The Wizard is an interactive and one‑shot assistant that turns natural‑language requests into runnable zyra CLI commands. It can preview suggested commands, ask for confirmation, and execute them in sequence.

• Command: zyra wizard
• Modes: interactive REPL or one‑shot via --prompt
• Providers: OpenAI, Ollama, Gemini (Vertex/Generative Language API), or a network‑free mock

This implements the MVP for the interactive assistant. For current scope and status, see Roadmap-and-Tracking.md.

Quick Start

• Interactive:
  • zyra wizard
• One‑shot:
  • zyra wizard --prompt "Convert hrrr.grib2 to NetCDF." --dry-run

Options

• --prompt TEXT: One‑shot query; without it, starts interactive mode.
• --provider {openai,ollama,gemini,vertex,mock}: LLM backend (default: openai; Gemini defaults to gemini-2.5-flash).
• --model NAME: Model name override for the selected provider.
• --dry-run: Show suggested commands but do not execute them.
• -y, --yes: Auto‑confirm execution without prompting.
• --max-commands N: Limit the number of commands to run.
• --log: Log prompts, replies, and command executions to ~/.zyra/wizard_logs/*.jsonl.
• --log-raw-llm: Include the full raw LLM response in logs.
• --show-raw: Print the full raw LLM output to stdout before parsing.
• --explain: Show suggested commands with inline # comments preserved (preview only).
• --test-llm: Probe connectivity to the configured LLM provider and exit.
• --edit: Always open an editor to modify commands before run (edit‑before‑run).
• --no-edit: Never open the editor; run/cancel only.
• --interactive: Ask for missing required arguments in one‑shot mode.

Providers & Environment

• OpenAI
  • Requires OPENAI_API_KEY.
  • Optional: OPENAI_BASE_URL (default: https://api.openai.com/v1).
  • If credentials/network are unavailable, Wizard falls back to a safe mock suggestion and does not crash.
• Ollama
  • Requires a local Ollama server (default base URL: http://localhost:11434).
  • Override with OLLAMA_BASE_URL.
• Gemini (Vertex AI / Generative Language API)
  • Option 1 (API key / hosted tools): set GOOGLE_API_KEY to call the Generative Language REST endpoint (e.g., gemini-2.5-flash, gemini-1.5-pro). Optional overrides: GENLANG_BASE_URL, VERTEX_MODEL.
  • Option 2 (Vertex managed): provide VERTEX_PROJECT (or GOOGLE_PROJECT_ID), VERTEX_LOCATION (default us-central1), plus Application Default Credentials (e.g., GOOGLE_APPLICATION_CREDENTIALS=/path/to/service-account.json). Optional overrides: VERTEX_MODEL, VERTEX_ENDPOINT, VERTEX_PUBLISHER.
  • Install google-auth (included in the llm extra) when using Vertex-side authentication. The client falls back to the mock provider when credentials are missing.
• Mock (offline)
  • No network, returns simple deterministic suggestions for testing.

Additional env overrides:

• ZYRA_LLM_PROVIDER: default provider (openai, ollama, gemini, vertex, or mock).
• ZYRA_LLM_MODEL: default model name for the chosen provider.
• Gemini-specific:
  • GOOGLE_API_KEY: API key for the Generative Language REST API.
  • VERTEX_PROJECT: Vertex AI project ID (alias: GOOGLE_PROJECT_ID, GOOGLE_CLOUD_PROJECT).
  • VERTEX_LOCATION: Vertex AI region (alias: GOOGLE_CLOUD_REGION; default us-central1).
  • VERTEX_MODEL: Gemini model name (default gemini-2.5-flash).
  • VERTEX_ENDPOINT: Custom Vertex API root (optional).
  • GENLANG_BASE_URL: Alternate Generative Language API root (optional).
  • GOOGLE_APPLICATION_CREDENTIALS: path to a service-account JSON when using ADC.

Configuration & Precedence

Wizard supports three levels of configuration. Highest priority wins:

1. CLI flags (per-invocation)

• --provider openai|ollama|gemini|vertex|mock
• --model NAME

2. Environment variables (good for CI/CD)

• ZYRA_LLM_PROVIDER (e.g., mock for reproducible tests)
• ZYRA_LLM_MODEL
• Provider-specific: OPENAI_API_KEY, OPENAI_BASE_URL, OLLAMA_BASE_URL

3. Config file (persistent defaults)

• Path: ~/.zyra_wizard.yaml
• Keys:
  • provider: openai, ollama, or mock
  • model: model name string

Example ~/.zyra_wizard.yaml:

provider: mock
model: mistral

Examples of precedence:

• Config sets provider: openai, but CI exports ZYRA_LLM_PROVIDER=mock → mock is used.
• Env sets ZYRA_LLM_PROVIDER=openai, user passes --provider ollama → ollama is used.

How It Works

1. Wizard sends your prompt to the selected provider (or mock) with system guidance.
2. It extracts zyra ... commands (supports multiple fenced blocks, and bash prompts like $ zyra...). Inline # comments outside quotes are stripped.
3. Safety: only zyra ... lines are accepted for execution; any other lines are ignored.
4. In --dry-run mode, it prints the commands only (no session state changes). With execution, it asks for confirmation (unless -y) and runs commands sequentially using the internal CLI entrypoint.
5. Execution stops at the first non‑zero return code; that code is returned by wizard.

Session Memory (Interactive)

Wizard maintains lightweight context across turns during interactive sessions:

• Tracks the last output file from suggested commands (via --output/-o).
• Shows suggested commands to the user before execution (unless -y).
• Feeds a short context header to the model on the next turn, e.g.:
  • Last file: /path/to/previous_output.nc
  • recent commands (up to 5).

This lets prompts like “make a heatmap” automatically refer to the prior result.

Notes:

• Session context is updated only when executing (non‑dry‑run) or during the interactive REPL as commands are run.
• Context is in‑memory for now; logs can be enabled with --log.

Command History

Wizard keeps a command history you can browse and reuse.

• Persistent file: ~/.zyra/wizard_history (JSONL)

  • One line per successful execution:

    { "ts": "2025-08-20T12:34:56Z", "cmd": "zyra visualize heatmap --input x.nc --var temp" }

  • Corrupted lines are skipped on load with a warning.
  • Consecutive duplicates are deduplicated when loading.

• REPL helpers:

  • :history [N]: show history (optionally last N).
  • !N or :retry N: re-run the Nth command as-is.
  • :edit N: open the Nth command in the edit-before-run flow.
  • :clear-history: clear in-memory and persisted history.
  • :save-history <file>: export commands (one per line) for sharing.

Examples:

> :history 5
[1] zyra process subset --input a.nc --var t --output co.nc
[2] zyra visualize heatmap --input co.nc --var temp
> !2   # retry
> :edit 1   # edit before running
> :save-history history.txt
Saved 2 command(s) to /path/to/history.txt

Usage Examples

• One‑shot (preview only):

  • zyra wizard --provider mock --prompt "Convert hrrr.grib2 to NetCDF." --dry-run

• One‑shot (execute with auto‑confirm, OpenAI):

  • OPENAI_API_KEY=... zyra wizard --prompt "Subset HRRR to Colorado and make a heatmap." -y

• Interactive (basic):

  • zyra wizard
  • Flow example:
    • > subset HRRR to Colorado → suggests ... --output co.nc
    • > make a heatmap → suggests using --input co.nc automatically

• Config file (persistent defaults):

  • Create ~/.zyra_wizard.yaml:

    provider: mock
    model: mistral

  • Run one‑shot with defaults:
    • zyra wizard --prompt "convert input.nc to GeoTIFF" --dry-run

• Logging (JSONL events):

  • Minimal logging:
    • zyra wizard --provider mock --prompt "convert" --dry-run --log
  • Include raw LLM text:
    • zyra wizard --provider mock --prompt "convert" --dry-run --log --log-raw-llm

• Explain and Raw Output (preview aids):

  • Show raw model output first:
    • zyra wizard --provider mock --prompt "convert" --dry-run --show-raw
  • Preserve inline comments in preview (execution still strips them safely):
    • zyra wizard --provider mock --prompt "subset then plot" --dry-run --explain

Sample JSONL Log

{"schema_version":1,"session_id":"3f1c2a6b2bdf4e2f98d9a3a3d1e2c4ab","event_id":"2e7b1c3a4d5e6f708192a3b4c5d6e7f0","ts":"2025-08-19T18:25:43Z","type":"user_prompt","prompt":"convert","provider":"mock","model":null}
{"schema_version":1,"session_id":"3f1c2a6b2bdf4e2f98d9a3a3d1e2c4ab","event_id":"9a8b7c6d5e4f3a2b1c0d9e8f7a6b5c4d","ts":"2025-08-19T18:25:43Z","type":"dry_run","commands":["zyra process convert-format input.nc --format geotiff --output output.tif"],"provider":"mock","model":null}
{"schema_version":1,"session_id":"3f1c2a6b2bdf4e2f98d9a3a3d1e2c4ab","event_id":"0f1e2d3c4b5a69788776655443322110","ts":"2025-08-19T18:25:43Z","type":"result","returncode":0,"ok":true,"commands":["zyra process convert-format input.nc --format geotiff --output output.tif"],"provider":"mock","model":null}

Logging

• Enable with --log. Logs are written to ~/.zyra/wizard_logs/YYYYMMDDTHHMMSSZ.jsonl.
• Each line is JSON with structured events to aid debugging and replay:
  • Common fields on every event: schema_version: 1, session_id, and unique event_id (UUIDs) for correlation.
  • user_prompt: includes provider and model used.
  • assistant_reply (optional): includes raw LLM text when --log-raw-llm is set.
  • dry_run: records suggested commands without execution.
  • exec: per-command execution with cmd, returncode, ok, provider, model.
  • result: summary of the whole turn with returncode, ok, commands, provider, model.
• Toggle raw LLM capture with --log-raw-llm (off by default to reduce noise and avoid storing sensitive model output inadvertently).

Capabilities Manifest

The Wizard can ground suggestions using a structured capabilities manifest that lists available commands and their options.

• Generation:
  • poetry run zyra generate-manifest (writes to src/zyra/wizard/zyra_capabilities.json by default).
  • Use -o /path/to/file.json to write elsewhere.
  • Note: When new CLI commands or options are added (e.g., search flags), regenerate the manifest so assistants and tools stay in sync.
• Packaging: The manifest under src/zyra/wizard/zyra_capabilities.json is bundled with the package.
• Usage in Wizard: On each turn, Wizard loads the manifest (if present), selects relevant entries by keyword match against the user prompt, and prepends a short context like:
  • - Relevant commands:
  • - visualize heatmap: Generate heatmap images. Options: --input, --var, --output

This keeps prompts short, accurate, and focused on commands likely to be needed.

If the model suggests no runnable commands, the Wizard emits a minimal, safe default that encourages interactive argument resolution (e.g., a heatmap or time‑series visualization with placeholder arguments).

Manifest Schema (Draft)

Consumers should support both legacy and enriched forms for backward compatibility.

• Top-level: object mapping full command strings to metadata.
• Command object:
  • description: short text
  • options: mapping of flag → value

Option entry formats

• Simple string (legacy):

  "--flag": "help text"

• Rich object (enriched):

  "--output": {
    "help": "Path to write output file",
    "path_arg": true,
    "type": "path",
    "choices": ["geotiff", "netcdf"],
    "required": false
  }

Field semantics

• help: human-readable help string.
• path_arg (bool): the next token is a filesystem path; Wizard uses this to enable file path completion.
• type (string): basic type of the argument; values may include int, float, str, path, or a callable/type name.
• choices (array): enum of acceptable values as exposed by argparse.
• required (bool): option must be provided (rare for short flags but included for completeness).

Example: Before → After

Before (legacy):

{
  "visualize heatmap": {
    "description": "zyra visualize heatmap",
    "options": {
      "--input": "Path to .nc or .npy input",
      "--output": "Output PNG path"
    }
  }
}

After (enriched):

{
  "visualize heatmap": {
    "description": "zyra visualize heatmap",
    "options": {
      "--input": {"help": "Path to .nc or .npy input", "path_arg": true, "type": "path"},
      "--output": {"help": "Output PNG path", "path_arg": true, "type": "path"},
      "--cmap": {"help": "", "choices": ["viridis", "plasma", "inferno", "magma"], "type": "str", "default": "viridis"}
    }
  }
}

How Wizard Uses path_arg

• The REPL completer detects options marked with path_arg: true and invokes a path completer for the next token.
• When path_arg is absent, Wizard falls back to a small heuristic set (e.g., --input, --output) to maintain compatibility with older manifests.

Incremental Adoption

• Producers may emit a mix of string and object entries; consumers must handle both.
• Additional fields (choices, type, required) are optional and can be adopted as needed by clients and UIs.

Groups and Defaults

• Each command may include groups: a list of { title, options[] } reflecting argparse groups such as “Input Options” or “Output Options”.
• Enriched option entries may include default. Completers and UIs can display these in tooltips (e.g., --cmap (default: viridis)).

Limitations (MVP)

• The LLM must output explicit zyra commands; Wizard does not synthesize flags/arguments on its own beyond parsing.
• Safety is basic (confirm or --dry-run). No sandbox or permission checks beyond what the underlying CLI performs.
• None on persistence: config file is supported (~/.zyra_wizard.yaml).

Roadmap

• Interactive session memory (retain context between prompts).
• Config UX enhancements (validation, helpful errors).
• Enhanced parsing and validation of suggested commands.
• Expanded provider options and improved offline behavior.
• Command explanations and optimization suggestions.

Troubleshooting

• “No zyra commands were suggested.”
  • Ask the model to output runnable commands in a fenced code block.
• Provider errors (network/credentials):
  • Wizard falls back to mock suggestions and continues; set --provider mock to avoid network entirely.
• Nothing happens on execute:
  • Try --dry-run to inspect commands; then run with -y. Note on semantic dataset search

---

Semantic dataset discovery now lives under the search command for a more natural fit:

• Use: zyra search --semantic "your dataset request" --limit 10 --show-plan.
• Wizard --semantic remains available for compatibility, but new workflows should prefer zyra search --semantic.