the-agency-ai/src/claude/skills/dispatch-read

<!-- Flag #62/#63: allowed-tools removed. Inherits Bash(*) from .claude/settings.json. Restricting to specific subcommand patterns at the skill level silently blocks agents on permission prompts the agent cannot see — see dispatch #171 for the devex incident that surfaced this trap. -->

Read Dispatch

Read a dispatch by ID, display its contents, and mark as read. Payloads resolve transparently across branches and worktrees — no merge needed.

Arguments

• $ARGUMENTS: An integer dispatch ID. If empty, list available dispatches first.

Instructions

Step 1: List or read

If $ARGUMENTS is empty:

1. Run ./agency/tools/dispatch list to show dispatches for the current agent.
2. If no dispatches, try ./agency/tools/dispatch list --all for all agents.
3. Ask which dispatch to read.

If $ARGUMENTS is an integer ID:

1. Run ./agency/tools/dispatch read $ARGUMENTS to display and mark as read.

To peek without marking as read, use ./agency/tools/dispatch fetch $ARGUMENTS instead.

Step 2: Evaluate findings

After reading the dispatch:

1. For each finding or directive, assess: is it actionable? Has context changed?
2. For review dispatches: classify each finding as valid (fix it), stale (code changed), or disputed (explain why).
3. Report your assessment before starting work.

Step 3: Act on valid items

For review findings, follow the red-green cycle:

1. Write a bug-exposing test that fails (red).
2. Fix the code.
3. Verify the test passes (green).

For directives, execute the work as specified.

Step 4: Respond and resolve

After all items are addressed:

1. Run ./agency/tools/dispatch reply <id> "summary of what was done" to send a response.
2. Run ./agency/tools/dispatch resolve <id> --response <reply-id> to mark as resolved.
3. Run /iteration-complete to commit your code changes.