zeroclaw/AGENTS.md

AGENTS.md — ZeroClaw Agent Engineering Protocol

This file defines the default working protocol for coding agents in this repository. Scope: entire repository.

1) Project Snapshot (Read First)

ZeroClaw is a Rust-first autonomous agent runtime optimized for:

• high performance
• high efficiency
• high stability
• high extensibility
• high sustainability
• high security

Core architecture is trait-driven and modular. Most extension work should be done by implementing traits and registering in factory modules.

Key extension points:

• src/providers/traits.rs (Provider)
• src/channels/traits.rs (Channel)
• src/tools/traits.rs (Tool)
• src/memory/traits.rs (Memory)
• src/observability/traits.rs (Observer)
• src/runtime/traits.rs (RuntimeAdapter)

2) Deep Architecture Observations (Why This Protocol Exists)

These codebase realities should drive every design decision:

1. Trait + factory architecture is the stability backbone
   • Extension points are intentionally explicit and swappable.
   • Most features should be added via trait implementation + factory registration, not cross-cutting rewrites.
2. Security-critical surfaces are first-class and internet-adjacent
   • src/gateway/, src/security/, src/tools/, src/runtime/ carry high blast radius.
   • Defaults already lean secure-by-default (pairing, bind safety, limits, secret handling); keep it that way.
3. Performance and binary size are product goals, not nice-to-have
   • Cargo.toml release profile and dependency choices optimize for size and determinism.
   • Convenience dependencies and broad abstractions can silently regress these goals.
4. Config and runtime contracts are user-facing API
   • src/config/schema.rs and CLI commands are effectively public interfaces.
   • Backward compatibility and explicit migration matter.
5. The project now runs in high-concurrency collaboration mode
   • CI + docs governance + label routing are part of the product delivery system.
   • PR throughput is a design constraint; not just a maintainer inconvenience.

3) Engineering Principles (Normative)

These principles are mandatory by default. They are not slogans; they are implementation constraints.

3.1 KISS (Keep It Simple, Stupid)

Why here: Runtime + security behavior must stay auditable under pressure.

Required:

• Prefer straightforward control flow over clever meta-programming.
• Prefer explicit match branches and typed structs over hidden dynamic behavior.
• Keep error paths obvious and localized.

3.2 YAGNI (You Aren't Gonna Need It)

Why here: Premature features increase attack surface and maintenance burden.

Required:

• Do not add new config keys, trait methods, feature flags, or workflow branches without a concrete accepted use case.
• Do not introduce speculative “future-proof” abstractions without at least one current caller.
• Keep unsupported paths explicit (error out) rather than adding partial fake support.

3.3 DRY + Rule of Three

Why here: Naive DRY can create brittle shared abstractions across providers/channels/tools.

Required:

• Duplicate small, local logic when it preserves clarity.
• Extract shared utilities only after repeated, stable patterns (rule-of-three).
• When extracting, preserve module boundaries and avoid hidden coupling.

3.4 SRP + ISP (Single Responsibility + Interface Segregation)

Why here: Trait-driven architecture already encodes subsystem boundaries.

Required:

• Keep each module focused on one concern.
• Extend behavior by implementing existing narrow traits whenever possible.
• Avoid fat interfaces and “god modules” that mix policy + transport + storage.

3.5 Fail Fast + Explicit Errors

Why here: Silent fallback in agent runtimes can create unsafe or costly behavior.

Required:

• Prefer explicit bail!/errors for unsupported or unsafe states.
• Never silently broaden permissions/capabilities.
• Document fallback behavior when fallback is intentional and safe.

3.6 Secure by Default + Least Privilege

Why here: Gateway/tools/runtime can execute actions with real-world side effects.

Required:

• Deny-by-default for access and exposure boundaries.
• Never log secrets, raw tokens, or sensitive payloads.
• Keep network/filesystem/shell scope as narrow as possible unless explicitly justified.

3.7 Determinism + Reproducibility

Why here: Reliable CI and low-latency triage depend on deterministic behavior.

Required:

• Prefer reproducible commands and locked dependency behavior in CI-sensitive paths.
• Keep tests deterministic (no flaky timing/network dependence without guardrails).
• Ensure local validation commands map to CI expectations.

3.8 Reversibility + Rollback-First Thinking

Why here: Fast recovery is mandatory under high PR volume.

Required:

• Keep changes easy to revert (small scope, clear blast radius).
• For risky changes, define rollback path before merge.
• Avoid mixed mega-patches that block safe rollback.

4) Repository Map (High-Level)

• src/main.rs — CLI entrypoint and command routing
• src/lib.rs — module exports and shared command enums
• src/config/ — schema + config loading/merging
• src/agent/ — orchestration loop
• src/gateway/ — webhook/gateway server
• src/security/ — policy, pairing, secret store
• src/memory/ — markdown/sqlite memory backends + embeddings/vector merge
• src/providers/ — model providers and resilient wrapper
• src/channels/ — Telegram/Discord/Slack/etc channels
• src/tools/ — tool execution surface (shell, file, memory, browser)
• src/runtime/ — runtime adapters (currently native/docker)
• docs/ — architecture + process docs
• .github/ — CI, templates, automation workflows

5) Risk Tiers by Path (Review Depth Contract)

Use these tiers when deciding validation depth and review rigor.

• Low risk: docs/chore/tests-only changes
• Medium risk: most src/** behavior changes without boundary/security impact
• High risk: src/security/**, src/runtime/**, src/gateway/**, src/tools/**, .github/workflows/**, access-control boundaries

When uncertain, classify as higher risk.

6) Agent Workflow (Required)

1. Read before write
   • Inspect existing module, factory wiring, and adjacent tests before editing.
2. Define scope boundary
   • One concern per PR; avoid mixed feature+refactor+infra patches.
3. Implement minimal patch
   • Apply KISS/YAGNI/DRY rule-of-three explicitly.
4. Validate by risk tier
   • Docs-only: lightweight checks.
   • Code/risky changes: full relevant checks and focused scenarios.
5. Document impact
   • Update docs/PR notes for behavior, risk, side effects, and rollback.
6. Respect queue hygiene
   • If stacked PR: declare Depends on #....
   • If replacing old PR: declare Supersedes #....

6.1 Code Naming Contract (Required)

Apply these naming rules for all code changes unless a subsystem has a stronger existing pattern.

• Use Rust standard casing consistently: modules/files snake_case, types/traits/enums PascalCase, functions/variables snake_case, constants/statics SCREAMING_SNAKE_CASE.
• Name types and modules by domain role, not implementation detail (for example DiscordChannel, SecurityPolicy, MemoryStore over vague names like Manager/Helper).
• Keep trait implementer naming explicit and predictable: <ProviderName>Provider, <ChannelName>Channel, <ToolName>Tool, <BackendName>Memory.
• Keep factory registration keys stable, lowercase, and user-facing (for example "openai", "discord", "shell"), and avoid alias sprawl without migration need.
• Name tests by behavior/outcome (<subject>_<expected_behavior>) and keep fixture identifiers neutral/project-scoped.
• If identity-like naming is required in tests/examples, use ZeroClaw-native labels only (ZeroClawAgent, zeroclaw_user, zeroclaw_node).

6.2 Architecture Boundary Contract (Required)

Use these rules to keep the trait/factory architecture stable under growth.

• Extend capabilities by adding trait implementations + factory wiring first; avoid cross-module rewrites for isolated features.
• Keep dependency direction inward to contracts: concrete integrations depend on trait/config/util layers, not on other concrete integrations.
• Avoid creating cross-subsystem coupling (for example provider code importing channel internals, tool code mutating gateway policy directly).
• Keep module responsibilities single-purpose: orchestration in agent/, transport in channels/, model I/O in providers/, policy in security/, execution in tools/.
• Introduce new shared abstractions only after repeated use (rule-of-three), with at least one real caller in current scope.
• For config/schema changes, treat keys as public contract: document defaults, compatibility impact, and migration/rollback path.

7) Change Playbooks

7.1 Adding a Provider

• Implement Provider in src/providers/.
• Register in src/providers/mod.rs factory.
• Add focused tests for factory wiring and error paths.
• Avoid provider-specific behavior leaks into shared orchestration code.

7.2 Adding a Channel

• Implement Channel in src/channels/.
• Keep send, listen, health_check, typing semantics consistent.
• Cover auth/allowlist/health behavior with tests.

7.3 Adding a Tool

• Implement Tool in src/tools/ with strict parameter schema.
• Validate and sanitize all inputs.
• Return structured ToolResult; avoid panics in runtime path.

7.4 Memory / Runtime / Config Changes

• Keep compatibility explicit (config defaults, migration impact, fallback behavior).
• Add targeted tests for boundary conditions and unsupported values.
• Avoid hidden side effects in startup path.

7.5 Security / Gateway / CI Changes

• Include threat/risk notes and rollback strategy.
• Add/update tests or validation evidence for failure modes and boundaries.
• Keep observability useful but non-sensitive.

8) Validation Matrix

Default local checks for code changes:

cargo fmt --all -- --check
cargo clippy --all-targets -- -D warnings
cargo test

Additional expectations by change type:

• Docs/template-only: run markdown lint and relevant doc checks.
• Workflow changes: validate YAML syntax; run workflow lint/sanity checks when available.
• Security/runtime/gateway/tools: include at least one boundary/failure-mode validation.

If full checks are impractical, run the most relevant subset and document what was skipped and why.

9) Collaboration and PR Discipline

• Follow .github/pull_request_template.md fully (including side effects / blast radius).
• Keep PR descriptions concrete: problem, change, non-goals, risk, rollback.
• Use conventional commit titles.
• Prefer small PRs (size: XS/S/M) when possible.
• Agent-assisted PRs are welcome, but contributors remain accountable for understanding what their code will do.

9.1 Privacy/Sensitive Data and Neutral Wording (Required)

Treat privacy and neutrality as merge gates, not best-effort guidelines.

• Never commit personal or sensitive data in code, docs, tests, fixtures, snapshots, logs, examples, or commit messages.
• Prohibited data includes (non-exhaustive): real names, personal emails, phone numbers, addresses, access tokens, API keys, credentials, IDs, and private URLs.
• Use neutral project-scoped placeholders (for example: user_a, test_user, project_bot, example.com) instead of real identity data.
• Test names/messages/fixtures must be impersonal and system-focused; avoid first-person or identity-specific language.
• If identity-like context is unavoidable, use ZeroClaw-scoped roles/labels only (for example: ZeroClawAgent, ZeroClawOperator, zeroclaw_user) and avoid real-world personas.
• Recommended identity-safe naming palette (use when identity-like context is required):
  • actor labels: ZeroClawAgent, ZeroClawOperator, ZeroClawMaintainer, zeroclaw_user
  • service/runtime labels: zeroclaw_bot, zeroclaw_service, zeroclaw_runtime, zeroclaw_node
  • environment labels: zeroclaw_project, zeroclaw_workspace, zeroclaw_channel
• If reproducing external incidents, redact and anonymize all payloads before committing.
• Before push, review git diff --cached specifically for accidental sensitive strings and identity leakage.

Reference docs:

• CONTRIBUTING.md
• docs/pr-workflow.md
• docs/reviewer-playbook.md
• docs/ci-map.md

10) Anti-Patterns (Do Not)

• Do not add heavy dependencies for minor convenience.
• Do not silently weaken security policy or access constraints.
• Do not add speculative config/feature flags “just in case”.
• Do not mix massive formatting-only changes with functional changes.
• Do not modify unrelated modules “while here”.
• Do not bypass failing checks without explicit explanation.
• Do not hide behavior-changing side effects in refactor commits.
• Do not include personal identity or sensitive information in test data, examples, docs, or commits.

11) Handoff Template (Agent -> Agent / Maintainer)

When handing off work, include:

1. What changed
2. What did not change
3. Validation run and results
4. Remaining risks / unknowns
5. Next recommended action

12) Vibe Coding Guardrails

When working in fast iterative mode:

• Keep each iteration reversible (small commits, clear rollback).
• Validate assumptions with code search before implementing.
• Prefer deterministic behavior over clever shortcuts.
• Do not “ship and hope” on security-sensitive paths.
• If uncertain, leave a concrete TODO with verification context, not a hidden guess.