sidesight AGENTS.md

AGENTS.md

Project mission

Build and maintain SideSight, a CLI-first vision sidecar that gives text-only coding agents image and video understanding through a separately configured multimodal model.

The complete product includes:

1. A production-ready sidesight CLI
2. A shared vision-analysis core
3. OpenAI-compatible multimodal provider support
4. An optional stdio MCP server
5. A Codex-compatible Agent Skill
6. Security controls for local files and remote media
7. Offline, mocked integration, and opt-in live tests
8. Complete user and developer documentation

The CLI is the primary product.

The MCP server and Agent Skill must reuse the same core behavior rather than implementing parallel logic.

---

Source of truth

The following files define the work:

• TASKS.md — ordered implementation tasks
• ACCEPTANCE.md — user-visible and system-level acceptance scenarios
• WORKLOG.md — current progress, validation results, blockers, and next task
• README.md — public product behavior and usage
• AGENTS.md — execution rules for coding agents

Treat TASKS.md and ACCEPTANCE.md as authoritative.

When documentation and implementation disagree:

1. Check TASKS.md.
2. Check ACCEPTANCE.md.
3. Check the original product requirements.
4. Update stale documentation when implementation intentionally changes.
5. Never silently weaken an acceptance requirement to make tests pass.

---

Primary execution objective

When asked to implement the project or a defined scope, complete the entire requested goal in the current run whenever technically possible.

Do not stop after:

• Creating the repository structure
• Writing a plan
• Installing dependencies
• Adding CLI command stubs
• Implementing one provider
• Implementing only the happy path
• Writing tests without running them
• Passing unit tests while acceptance tests still fail
• Producing documentation for behavior that does not work
• Reaching a convenient milestone
• Encountering an ordinary implementation bug