Autonomous Coding Workflows

This guide shows real-world EdgeCrab workflows that go beyond simple Q&A. Each pattern is a prompt strategy plus a set of skills that produces reliable results.

---

Pattern 1 — Bug Hunt

Goal: Find and fix a regression without knowing which file it’s in.

There's a regression where authenticated users are being logged out after 5 minutes
even though the session should last 24 hours. Find the root cause and fix it.

What EdgeCrab does:

1. Searches codebase for session-related code (web_search-free)
2. Reads session configuration files
3. Reads the authentication middleware
4. Reads recent commit history (git log --oneline -20)
5. Narrows to the likely commit with git show
6. Reads the changed lines
7. Proposes and implements a fix
8. Offers to run the test suite

Tips:

• Include a reproduction step in your prompt if you have one
• Name the feature area when you know it (“in the session expiry logic”)

---

Pattern 2 — Guided Refactor

Goal: Rename a type across the entire codebase safely.

Rename the Rust type `UserRecord` to `User` throughout the entire codebase.
Run cargo check after each file to confirm no new errors. If cargo check fails,
fix any type mismatches before moving on.

What EdgeCrab does:

1. Finds all occurrences with file_search
2. Renames each file with file_write
3. Runs cargo check after each batch
4. Fixes any resulting type errors
5. Reports a summary of all changed files

With a skill: Create a rust-rename-type skill with these steps pre-written so you can invoke it with a single line: Rename UserRecord to User using the rust-rename-type skill.

---

Pattern 3 — Test Generation

Goal: Write a comprehensive test suite for an under-tested module.

The `edgecrab-security` crate has no tests for the SSRF guard. Write a comprehensive
test suite covering: private IP ranges (IPv4 and IPv6), DNS-based bypasses, valid
public URLs, and edge cases. Use Rust's built-in test framework.

What EdgeCrab does:

1. Reads the SSRF guard implementation (edgecrab-security/src/ssrf.rs)
2. Reads any existing tests for context
3. Identifies the main code paths and edge cases
4. Writes test functions covering all identified cases
5. Runs cargo test -p edgecrab-security to verify they pass

---

Pattern 4 — Documentation Run

Goal: Add inline documentation to all public API items.

Add Rust doc comments (///) to every public function, struct, and enum in
edgecrab-core/src/agent.rs. Comments should explain what the item does,
not just restate the name. Add examples for the most important functions.

What EdgeCrab does:

1. Reads the file to understand the public API
2. Writes documentation for each item
3. Runs cargo doc --no-deps to verify it renders correctly
4. Checks for any cargo clippy -- -D warnings doc-comment issues

---

Pattern 5 — Security Audit

Goal: Find security issues in a specific module.

Perform a security audit of the file tool implementation in
edgecrab-tools/src/tools/file.rs. Focus on path traversal, symlink attacks,
and TOCTOU races. Cite specific line numbers for any issues found.

Even better with the security-review skill:

Use the security-review skill to audit edgecrab-tools/src/tools/file.rs.

---

Pattern 6 — Codebase Onboarding

Goal: Understand an unfamiliar codebase quickly.

I'm new to this codebase. Start with the top-level Cargo.toml and work
down: explain what each crate does, how they depend on each other, and
what the overall data flow is from a user prompt to a tool call. Draw an
ASCII architecture diagram at the end.

What EdgeCrab does:

1. Reads Cargo.toml (workspace manifest)
2. Reads each crate’s Cargo.toml and src/lib.rs or src/main.rs
3. Understands the dependency graph
4. Summarizes the architecture
5. Produces an ASCII diagram

---

Pattern 7 — CI Failure Triage

Goal: Diagnose and fix a failing CI build without running CI yourself.

The CI build is failing. Here's the error output:
[paste CI output]

Find the root cause, fix it, and make sure `cargo test` passes locally before
you're done.

What EdgeCrab does:

1. Parses the error output
2. Reads the relevant source files
3. Identifies the root cause
4. Implements a fix
5. Runs cargo test to confirm it passes

---

Chaining Workflows

For complex multi-step projects, give EdgeCrab an explicit plan:

We're going to refactor the auth module to use JWT. Here's the plan:

1. Read all files in src/auth/ to understand the current implementation
2. Design the new JWT-based approach and explain it to me before touching code
3. After my approval, implement the changes one file at a time
4. Write tests for the new implementation
5. Run cargo test and fix any failures
6. Summarize all changes in a PR description format

Start with step 1.

The explicit numbered plan gives EdgeCrab a clear loop termination condition and lets you review the design before any code is written.

---

Prompt Tips for Best Results

Do	Don’t
Specify the file or module name when you know it	Give open-ended “improve this codebase”
Ask for tests to be run after changes	Trust the agent to self-validate without running tests
Request a summary of changes at the end	Let the loop end without a completion report
Give constraints (“don’t modify the test files”)	Let constraints be implied
Provide context (“this broke after commit abc123”)	Make the agent guess the context

---

Workflow Pattern Summary

Pattern	Best For	Key Prompt Element
Bug Hunt	Regressions, unknown root cause	Symptom + expected behaviour
Guided Refactor	Rename, extract, restructure	Old name → New name + “run check after each file”
Test Generation	Missing tests, edge-case coverage	Module path + coverage goals
Documentation Run	Public API doc comments	File path + quality bar
Security Audit	Module-level hardening	File path + threat categories
Codebase Onboarding	Unfamiliar repo	”Start with Cargo.toml, explain each crate”
CI Failure Triage	Broken CI, pasted error log	Paste full error + “fix and run tests”

---

Pro Tips

• Give a termination condition: “You’re done when cargo test --workspace passes” prevents the agent from over-iterating.
• Ask for a diff summary at the end: “Summarize all changed files and why” produces useful commit message material.
• Use a skill for repeat workflows: Patterns 1–7 above can each be packaged as a skill so you invoke them with one line instead of a multi-sentence prompt.
• Set max_iterations higher for large refactors: A 50-file rename needs more loop iterations. Use EDGECRAB_MAX_ITERATIONS=150 or --max-iterations 150 for big tasks.
• Gate on human review for destructive ops: Prepend “Don’t delete or overwrite any file without showing me the change and asking for confirmation first” to any large-scale workflow.

---

FAQ

What if the agent is stuck in a loop? Use /stop to cancel the current run. Then either refine your prompt or set a lower max_iterations to force an early summary.

Can I chain skills with inline instructions? Yes: Use the security-review skill on src/auth/, then use the test-generation skill to write tests for any issues you find.

How do I resume a workflow that was cancelled? Use /retry to resend the last message, or start a new message with “Continue where you left off” if the session history is intact.

What is a session? A session is the conversation history for a single edgecrab run context. Sessions are stored in ~/.edgecrab/state.db.

---

See Also

• Skills System — package any pattern as a reusable skill
• Slash Commands — /stop, /retry, and live session controls
• Sessions — session persistence and the --session-id flag