devils-advocate 0.9.5__tar.gz

devils_advocate-0.9.5/.github/workflows/publish.yml

@@ -0,0 +1,50 @@
+name: Publish to PyPI
+
+on:
+  push:
+    branches: [main]
+
+jobs:
+  publish:
+    runs-on: ubuntu-latest
+    permissions:
+      contents: write
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          token: ${{ secrets.GITHUB_TOKEN }}
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install build tools
+        run: pip install build twine
+
+      - name: Bump build version
+        run: |
+          current=$(grep '^version' pyproject.toml | sed 's/version = "\(.*\)"/\1/')
+          major=$(echo "$current" | cut -d. -f1)
+          minor=$(echo "$current" | cut -d. -f2)
+          build=$(echo "$current" | cut -d. -f3)
+          new="${major}.${minor}.$((build + 1))"
+          sed -i "s/^version = \"${current}\"/version = \"${new}\"/" pyproject.toml
+          echo "VERSION=${new}" >> "$GITHUB_ENV"
+          echo "Bumped ${current} -> ${new}"
+
+      - name: Build package
+        run: python -m build
+
+      - name: Upload to PyPI
+        env:
+          TWINE_USERNAME: __token__
+          TWINE_PASSWORD: ${{ secrets.PYPI_TOKEN }}
+        run: python -m twine upload dist/*
+
+      - name: Commit version bump
+        run: |
+          git config user.name "github-actions"
+          git config user.email "github-actions@github.com"
+          git add pyproject.toml
+          git commit -m "build: bump version to ${VERSION} [skip ci]"
+          git push

devils_advocate-0.9.5/.gitignore

@@ -0,0 +1,11 @@
+models.yaml
+.dvad/
+__pycache__/
+.pytest_cache/
+.venv/
+*.egg-info/
+dist/
+*.pyc
+.mypy_cache/
+tests/e2e/failures/
+tests/e2e/videos/

devils_advocate-0.9.5/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Brian Kelley
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

devils_advocate-0.9.5/PKG-INFO

@@ -0,0 +1,162 @@
+Metadata-Version: 2.4
+Name: devils-advocate
+Version: 0.9.5
+Summary: Cost-aware multi-LLM adversarial review engine with deterministic governance
+Author: Brian Kelley
+License: MIT License
+        
+        Copyright (c) 2026 Brian Kelley
+        
+        Permission is hereby granted, free of charge, to any person obtaining a copy
+        of this software and associated documentation files (the "Software"), to deal
+        in the Software without restriction, including without limitation the rights
+        to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+        copies of the Software, and to permit persons to whom the Software is
+        furnished to do so, subject to the following conditions:
+        
+        The above copyright notice and this permission notice shall be included in all
+        copies or substantial portions of the Software.
+        
+        THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+        IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+        FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+        AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+        LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+        OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+        SOFTWARE.
+License-File: LICENSE
+Keywords: adversarial,code-review,consensus,governance,llm
+Requires-Python: >=3.12
+Requires-Dist: click>=8.1
+Requires-Dist: fastapi>=0.115
+Requires-Dist: httpx>=0.27
+Requires-Dist: jinja2>=3.1
+Requires-Dist: python-multipart>=0.0.9
+Requires-Dist: pyyaml>=6.0
+Requires-Dist: rich[markdown]>=13.0
+Requires-Dist: ruamel-yaml>=0.18
+Requires-Dist: uvicorn[standard]>=0.32
+Provides-Extra: dev
+Requires-Dist: httpx>=0.27; extra == 'dev'
+Requires-Dist: pytest-asyncio>=0.23; extra == 'dev'
+Requires-Dist: pytest>=8.0; extra == 'dev'
+Requires-Dist: respx>=0.21; extra == 'dev'
+Provides-Extra: e2e
+Requires-Dist: pillow>=10.0; extra == 'e2e'
+Requires-Dist: pixelmatch>=0.3; extra == 'e2e'
+Requires-Dist: playwright>=1.48; extra == 'e2e'
+Requires-Dist: pytest-playwright>=0.5; extra == 'e2e'
+Description-Content-Type: text/markdown
+
+
+<div align="center"><img width="200" alt="dvad-color-white" src="https://github.com/user-attachments/assets/35b89380-cddd-4b70-88ac-e71cb1d867a6" /></div>
+
+# Devil's Advocate
+
+Cost-aware multi-LLM adversarial review engine with deterministic governance.
+
+**Do you have an implementation plan, codebase, or spec created by Claude, GPT, Gemini, Grok, etc and you want the flagship model from competing frontier providers to rip it apart, exposing the holes in logic and potential coding landmines, before a single line of code gets written?**
+
+Devil's Advocate pits multiple LLM reviewers against an LLM author in a structured 2-round adversarial protocol. A deterministic governance engine (straight python, no LLM calls, no probability) resolves every finding into a machine-readable outcome. The result is a vetted artifact where every finding has been accepted, defended, challenged, or escalated (to you for final decision) with full traceability from the first objection through final resolution.
+
+DVAD is borne from my frustrations with code generated inside an AI echo chamber (single source provider). Often I'll post a script to an LLM that didn't write it to see if the other providers could spot issues or problems. They usually do. I decided to turn it into an easy to use UI.
+
+<img width="1358" height="989" alt="dvad main" src="https://github.com/user-attachments/assets/686fa07d-df88-436e-81a9-b4782d722107" />
+
+<img width="1319" height="1000" alt="dvad config" src="https://github.com/user-attachments/assets/e6c0fad0-4181-4885-b064-dadb35a81a81" />
+
+## Requirements
+
+- Python 3.12+
+- Bare Minimum: 1 API key from 1 provider, 3 models - an author and two reviewers. (*The other roles: dedup, integration, normalization, revision, can use the same models the author or reviewers use*).
+- Comfortably Confrontational: 3 API keys from 3 different providers - one for the author, one per reviewer. (*Different providers mean different blind spots and that's where the friction comes from. Friction is good.*)
+- Supported providers: Every frontier provider that uses OpenAI-compatible, Anthropic, or Minimax prompt formatting. (Google Gemini, ChatGPT, Claude, DeepSeek, xAI/Grok, Minimax, Kimi, etc.) 
+
+## Quick Install & Update
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/briankelley/devils-advocate/main/install.sh | bash
+```
+
+Installs dvad, initializes config, sets up the systemd service, and launches the web GUI on port 8411.
+
+## Manual Install
+
+```bash
+python3 -m venv ~/.local/share/devils-advocate/venv
+~/.local/share/devils-advocate/venv/bin/pip install devils-advocate
+ln -sf ~/.local/share/devils-advocate/venv/bin/dvad ~/.local/bin/dvad
+dvad install
+```
+
+## Getting Started
+
+### 1. Configure your models
+
+Run `dvad config --init` to generate `~/.config/devils-advocate/models.yaml`, then edit it. Each model needs a provider, model ID, and an environment variable name for its API key. See `examples/models.yaml.example` for a fully annotated template.
+
+### 2. Set your API keys
+
+API keys are resolved from environment variables — never stored in the config file. Set them in your shell or in `~/.config/devils-advocate/.env` (auto-loaded, won't override existing env vars).
+
+### 3. Validate
+
+```bash
+dvad config --show
+```
+
+## Web GUI
+
+The primary way to use Devil's Advocate. Covers the full workflow - submitting reviews, monitoring progress in real time, resolving escalated findings, and generating revised artifacts.
+
+```bash
+dvad gui
+```
+
+Opens at `http://127.0.0.1:8411`. The GUI includes a dashboard for submitting and browsing reviews, real-time review progress via SSE with per-model cost tracking, governance override controls, revision generation, and visual model/role configuration with a raw YAML editor.
+
+By default the GUI refuses to bind to non-localhost interfaces. `--allow-nonlocal` overrides this and requires a CSRF token header on all mutating requests.
+
+## How It Works
+
+1. **Independent review** - Multiple reviewer models analyze the input in parallel, producing findings with severity, category, location, and recommendation.
+2. **Deduplication** - A dedup model groups overlapping findings into consolidated review groups, preserving source attribution.
+3. **Author response** - The author model responds to each group: **ACCEPTED**, **REJECTED**, or **PARTIAL** with a rationale.
+4. **Rebuttal** - Reviewers issue rebuttals on contested groups only. Each votes **CONCUR** or **CHALLENGE**.
+5. **Final position** - For challenged groups, the author provides a final position.
+6. **Governance** - A deterministic engine (no LLM calls, pure rule-based logic) maps every group to an outcome: **AUTO_ACCEPTED**, **AUTO_DISMISSED**, or **ESCALATED**. No finding passes through without the author demonstrating engagement - implicit and rote acceptance both escalate to human review.
+
+Escalated findings are resolved through the GUI's override controls or `dvad override`. After governance, `dvad revise` generates the final revised artifact.
+
+## Review Modes
+
+| Mode          | Protocol      | Input                                               | Output                        |
+| ------------- | ------------- | --------------------------------------------------- | ----------------------------- |
+| `plan`        | Adversarial   | Plan file + optional reference files                | `revised-plan.md`             |
+| `code`        | Adversarial   | Exactly one code file, optional spec                | `revised-diff.patch`          |
+| `spec`        | Collaborative | Spec file(s)                                        | `revised-spec-suggestions.md` |
+| `integration` | Adversarial   | Input files or `.dvad/manifest.json`, optional spec | `remediation-plan.md`         |
+
+**spec** is non-adversarial - no author, no rebuttals, no governance. Findings are grouped by theme and compiled into a suggestion report. All other modes use the full 2-round adversarial protocol.
+
+## CLI Quick Start
+
+```bash
+dvad review --mode plan --input plan.md --input ref.py --project myproject
+dvad review --mode code --input src/app.py --spec spec.md --project myproject
+dvad review --mode plan --input plan.md --project myproject --max-cost 0.50
+dvad review --mode plan --input plan.md --project myproject --dry-run
+```
+
+## Design Notes
+
+- **No vendor SDKs.** All provider calls use `httpx` directly - full control over request shape and retry behavior.
+- **Deterministic governance.** Zero LLM calls. Every outcome is reproducible from the same inputs.
+- **Atomic operations.** File writes use `mkstemp` + `os.replace`. Locking uses `O_CREAT | O_EXCL`.
+- **XDG-compliant.** Config and data paths follow the XDG Base Directory specification.
+
+Full CLI reference, configuration schema, governance rules, and cost tracking details are available in the [documentation](docs/).
+
+## License
+
+MIT

devils_advocate-0.9.5/README.md

@@ -0,0 +1,112 @@
+
+<div align="center"><img width="200" alt="dvad-color-white" src="https://github.com/user-attachments/assets/35b89380-cddd-4b70-88ac-e71cb1d867a6" /></div>
+
+# Devil's Advocate
+
+Cost-aware multi-LLM adversarial review engine with deterministic governance.
+
+**Do you have an implementation plan, codebase, or spec created by Claude, GPT, Gemini, Grok, etc and you want the flagship model from competing frontier providers to rip it apart, exposing the holes in logic and potential coding landmines, before a single line of code gets written?**
+
+Devil's Advocate pits multiple LLM reviewers against an LLM author in a structured 2-round adversarial protocol. A deterministic governance engine (straight python, no LLM calls, no probability) resolves every finding into a machine-readable outcome. The result is a vetted artifact where every finding has been accepted, defended, challenged, or escalated (to you for final decision) with full traceability from the first objection through final resolution.
+
+DVAD is borne from my frustrations with code generated inside an AI echo chamber (single source provider). Often I'll post a script to an LLM that didn't write it to see if the other providers could spot issues or problems. They usually do. I decided to turn it into an easy to use UI.
+
+<img width="1358" height="989" alt="dvad main" src="https://github.com/user-attachments/assets/686fa07d-df88-436e-81a9-b4782d722107" />
+
+<img width="1319" height="1000" alt="dvad config" src="https://github.com/user-attachments/assets/e6c0fad0-4181-4885-b064-dadb35a81a81" />
+
+## Requirements
+
+- Python 3.12+
+- Bare Minimum: 1 API key from 1 provider, 3 models - an author and two reviewers. (*The other roles: dedup, integration, normalization, revision, can use the same models the author or reviewers use*).
+- Comfortably Confrontational: 3 API keys from 3 different providers - one for the author, one per reviewer. (*Different providers mean different blind spots and that's where the friction comes from. Friction is good.*)
+- Supported providers: Every frontier provider that uses OpenAI-compatible, Anthropic, or Minimax prompt formatting. (Google Gemini, ChatGPT, Claude, DeepSeek, xAI/Grok, Minimax, Kimi, etc.) 
+
+## Quick Install & Update
+
+```bash
+curl -fsSL https://raw.githubusercontent.com/briankelley/devils-advocate/main/install.sh | bash
+```
+
+Installs dvad, initializes config, sets up the systemd service, and launches the web GUI on port 8411.
+
+## Manual Install
+
+```bash
+python3 -m venv ~/.local/share/devils-advocate/venv
+~/.local/share/devils-advocate/venv/bin/pip install devils-advocate
+ln -sf ~/.local/share/devils-advocate/venv/bin/dvad ~/.local/bin/dvad
+dvad install
+```
+
+## Getting Started
+
+### 1. Configure your models
+
+Run `dvad config --init` to generate `~/.config/devils-advocate/models.yaml`, then edit it. Each model needs a provider, model ID, and an environment variable name for its API key. See `examples/models.yaml.example` for a fully annotated template.
+
+### 2. Set your API keys
+
+API keys are resolved from environment variables — never stored in the config file. Set them in your shell or in `~/.config/devils-advocate/.env` (auto-loaded, won't override existing env vars).
+
+### 3. Validate
+
+```bash
+dvad config --show
+```
+
+## Web GUI
+
+The primary way to use Devil's Advocate. Covers the full workflow - submitting reviews, monitoring progress in real time, resolving escalated findings, and generating revised artifacts.
+
+```bash
+dvad gui
+```
+
+Opens at `http://127.0.0.1:8411`. The GUI includes a dashboard for submitting and browsing reviews, real-time review progress via SSE with per-model cost tracking, governance override controls, revision generation, and visual model/role configuration with a raw YAML editor.
+
+By default the GUI refuses to bind to non-localhost interfaces. `--allow-nonlocal` overrides this and requires a CSRF token header on all mutating requests.
+
+## How It Works
+
+1. **Independent review** - Multiple reviewer models analyze the input in parallel, producing findings with severity, category, location, and recommendation.
+2. **Deduplication** - A dedup model groups overlapping findings into consolidated review groups, preserving source attribution.
+3. **Author response** - The author model responds to each group: **ACCEPTED**, **REJECTED**, or **PARTIAL** with a rationale.
+4. **Rebuttal** - Reviewers issue rebuttals on contested groups only. Each votes **CONCUR** or **CHALLENGE**.
+5. **Final position** - For challenged groups, the author provides a final position.
+6. **Governance** - A deterministic engine (no LLM calls, pure rule-based logic) maps every group to an outcome: **AUTO_ACCEPTED**, **AUTO_DISMISSED**, or **ESCALATED**. No finding passes through without the author demonstrating engagement - implicit and rote acceptance both escalate to human review.
+
+Escalated findings are resolved through the GUI's override controls or `dvad override`. After governance, `dvad revise` generates the final revised artifact.
+
+## Review Modes
+
+| Mode          | Protocol      | Input                                               | Output                        |
+| ------------- | ------------- | --------------------------------------------------- | ----------------------------- |
+| `plan`        | Adversarial   | Plan file + optional reference files                | `revised-plan.md`             |
+| `code`        | Adversarial   | Exactly one code file, optional spec                | `revised-diff.patch`          |
+| `spec`        | Collaborative | Spec file(s)                                        | `revised-spec-suggestions.md` |
+| `integration` | Adversarial   | Input files or `.dvad/manifest.json`, optional spec | `remediation-plan.md`         |
+
+**spec** is non-adversarial - no author, no rebuttals, no governance. Findings are grouped by theme and compiled into a suggestion report. All other modes use the full 2-round adversarial protocol.
+
+## CLI Quick Start
+
+```bash
+dvad review --mode plan --input plan.md --input ref.py --project myproject
+dvad review --mode code --input src/app.py --spec spec.md --project myproject
+dvad review --mode plan --input plan.md --project myproject --max-cost 0.50
+dvad review --mode plan --input plan.md --project myproject --dry-run
+```
+
+## Design Notes
+
+- **No vendor SDKs.** All provider calls use `httpx` directly - full control over request shape and retry behavior.
+- **Deterministic governance.** Zero LLM calls. Every outcome is reproducible from the same inputs.
+- **Atomic operations.** File writes use `mkstemp` + `os.replace`. Locking uses `O_CREAT | O_EXCL`.
+- **XDG-compliant.** Config and data paths follow the XDG Base Directory specification.
+
+Full CLI reference, configuration schema, governance rules, and cost tracking details are available in the [documentation](docs/).
+
+## License
+
+MIT

devils_advocate-0.9.5/RELEASING.md

@@ -0,0 +1,23 @@
+# Releasing
+
+## Version Bumps
+
+1. Update `__version__` in `src/devils_advocate/__init__.py`
+2. Update `version` in `pyproject.toml`
+3. Commit: `git commit -am "release: vX.Y.Z"`
+4. Tag: `git tag vX.Y.Z`
+
+## Build and Publish
+
+```bash
+pip install build twine
+python -m build
+twine upload dist/*
+```
+
+## Pre-release Checklist
+
+- [ ] All tests pass: `pytest`
+- [ ] Version strings match in `__init__.py` and `pyproject.toml`
+- [ ] CHANGELOG updated (if maintained)
+- [ ] No debug prints or development-only code

devils_advocate-0.9.5/docs/config-page-iconography.md

@@ -0,0 +1,20 @@
+# Functional Specification: Role Assignment & Chain-of-Thought Iconography
+
+## Role Assignment Icons (pen-tool, scan-eye, combine, scale, file-pen, puzzle)
+
+- **Behavior:** Bidirectional state machine between Models table and Role Assignments table. Click toggles membership in `{roles}` namespace. DOM syncs horizontally — assignment in either card immediately reflects in the other.
+- **Cardinality:** Radio-select for singular roles (author, dedup, normalization, revision, integration); checkbox with ceiling=2 for reviewer.
+- **Visual State:** Binary. `role-active` class (illuminated) = assigned; absence = unassigned.
+- **Mutation:** Updates `config.roles` in memory; persists via `/api/config` POST on explicit save.
+
+## Chain-of-Thought Icons (brain)
+
+- **Models Table:** Toggle with eligibility gating — click handler active only when model has role assignment (`thinking-eligible`). Mutates `model.thinking` boolean via `/api/config/model-thinking`. Visual state reflects `thinking` property from YAML.
+- **Role Assignments Table:** Read-only. Pure function of `modelThinking[assignedModelName]` — no event listeners.
+- **State Dependency:** Row existence derives from role assignment; visual state derives independently from `thinking` boolean on the assigned model.
+
+## Critical Constraints
+
+1. **Idempotent init:** `initRolePills()` and `initThinkingToggle()` guard double-attachment via `_rolePillsInitialized` / `_thinkingToggleInitialized` flags.
+2. **Hydration order:** Client-side CoT reconciliation deferred until `modelThinking` context is defined (populated by inline script injection post-`DOMContentLoaded`).
+3. **Soft navigation:** Logo click → config page must re-evaluate icon states against current `roles` and `modelThinking` without assuming empty initial state.