amp-code — Delegate Coding Tasks to Sourcegraph Amp

Use this skill to hand off coding work to Amp, an autonomous coding agent. Amp can read, write, refactor, and test code across an entire codebase without supervision.

---

When to delegate to Amp

Delegate when the task involves:

• Multi-file changes — refactoring, moving modules, updating APIs
• New features — implementing something that requires creating/editing several files
• Bug fixes — especially when the root cause requires investigation and code changes
• Test writing — generating test suites, adding coverage
• Anything > ~5 min of coding — if you'd need multiple edit calls, let Amp do it

Do NOT delegate:

• Simple single-line edits (just use edit)
• Purely diagnostic/read-only questions (just use exec/read)
• Tasks where you need fine-grained control over each file change

---

Modes

Mode	Description	Use when
rush	Fast, lightweight model	Small/clear tasks, quick fixes
smart	Balanced model (default)	Most tasks
deep	Powerful model, slower, higher cost	Complex architecture, hard bugs

---

How to run a task

Simple one-shot (recommended)

cd /path/to/project && \
  amp \
    --dangerously-allow-all \
    --no-notifications \
    --no-ide \
    -m smart \
    -x "Your task description here"

The -x flag (execute mode) makes amp non-interactive: it runs the task and exits, printing only the agent's final message to stdout.

Using the wrapper script

bash {baseDir}/scripts/amp-task.sh \
  --task "Add pagination to the /invoices endpoint" \
  --dir /path/to/your/project \
  --mode smart

The script handles: cd to project dir, thread creation (for auditability), execute mode, and clean output. It prints the thread ID and final agent response.

---

Key flags

Flag	Effect
--dangerously-allow-all	No confirmation prompts — agent acts autonomously
--no-notifications	Suppress sound/system notifications
--no-ide	Don't connect to IDE (safe for headless runs)
-x "message"	Execute mode: non-interactive, prints final output
-m rush/smart/deep	Select agent mode (model + system prompt)
-l "label"	Tag the thread with a label (repeatable)

---

Working directory matters

Amp reads the codebase from the current directory. Always cd to the project root first, or pass --dir to the wrapper script. Without the right cwd, Amp won't see the right files.

---

Checking results

After a run, the wrapper script outputs:

1. The thread ID — use this to inspect or continue the conversation
2. The final agent message — summary of what was done

To see the full thread as markdown (all messages, tool calls, etc.):

amp threads markdown <thread-id>

To continue a thread (e.g., to follow up or fix something):

cd /path/to/project && \
  echo "Now also add tests for the pagination logic" | \
  amp threads continue <thread-id> \
    --dangerously-allow-all --no-notifications --no-ide -x

To list recent threads:

amp threads list

---

Example invocations

One-shot fix

cd /path/to/your/project && \
  amp --dangerously-allow-all --no-notifications --no-ide -x \
  "Fix the TypeError in src/api/invoices.js line 42 — amount is sometimes null"

Feature implementation (via wrapper)

bash {baseDir}/scripts/amp-task.sh \
  --task "Implement CSV export for the invoices list page — add a button in the UI and a /api/invoices/export endpoint" \
  --dir /path/to/your/project \
  --mode smart

Deep refactor

bash {baseDir}/scripts/amp-task.sh \
  --task "Migrate all database calls from raw SQL to Knex query builder. Maintain existing behaviour, add comments." \
  --dir /path/to/your/project \
  --mode deep

Quick test generation

bash {baseDir}/scripts/amp-task.sh \
  --task "Write Jest unit tests for all functions in src/utils/formatting.js" \
  --dir /path/to/your/project \
  --mode rush

---

Notes

• Amp commits are tagged with Amp-Thread: trailer in git. Use this to find amp-authored commits.
• The --dangerously-allow-all flag bypasses all tool permission checks. Only use on trusted projects.
• Amp may make mistakes. Review the diff after large changes: git diff HEAD~1
• For very long tasks, amp may time out. Break into smaller sub-tasks if needed.
• Thread IDs are UUIDs. Save them if you want to audit or continue work later.