agent-memory-inbox-capture

You are running the /cc-tools:agent-memory-inbox-capture skill. Move the calling agent's own memory entries that outlive this run out of its throwaway worktree and into the session's inbox, where the curator can reach them.

You are the only stage that still has the run in context — what the task was, what was tried, which observations were about this branch and which about the repo. Every later stage reads an entry file stripped of all of it, and resolves the resulting ambiguity by publishing, which is how something that was true inside one worktree for an hour reaches CLAUDE.md. So the session-scope filter in step 3 is yours and nobody else's.

Do not rewrite an entry, and do not touch any other agent's subdirectory.

Read skills/lib/agent-memory-inbox.md for the inbox path, its <plugin>-<agent>/ layout, and how the branch and the scratchpad are resolved. That contract is the only statement of the path; this skill does not restate it.

Invocation

/cc-tools:agent-memory-inbox-capture

No arguments. Everything the skill needs — the calling agent's memory directory, the current branch, and the session scratchpad — comes from the invoking agent's own context and working tree.

Execution

1. Resolve the branch, before any end-of-run detach:

   git branch --show-current
   

   An empty result means HEAD is detached and the branch is unrecoverable. Report that and stop; capture nothing rather than writing to a wrongly-named inbox.

2. Find the entries. The calling agent's own memory directory is .claude/agent-memory/<plugin>-<agent>/ relative to its cwd, where <plugin>-<agent> is its own name — sdlc-issue-developer, say. Take that name from your own identity, never by looking at what the tree happens to hold, and sweep that one directory:

   find .claude/agent-memory/<plugin>-<agent> -type f -name '*.md'
   

   A sibling directory belongs to another agent and is never read. In a repo that gitignores .claude/agent-memory/ a fresh worktree has none; in a repo that commits the tree the worktree carries every agent's committed entries, and copying those would file another run's lessons under this run's capture.

   If the directory is absent, or holds no file other than a MEMORY.md, report that nothing was captured and stop. That is a valid outcome: it means the run learned nothing worth recording.

3. Ask of each entry whether it survives the run. Would this mean anything to a reader who was not here — someone with the repo in front of them and no knowledge of this task, this branch, or this worktree? Answer it from the run itself, which is the context only you have; you are not grading whether the entry is worth publishing, which is the curator's call on the entries that reach it.

   An entry that fails is dropped, which means not copying it and nothing more. Nothing is deleted: the agent's own files stay where they are and die with the worktree.

4. Copy each surviving entry file into the same-named <plugin>-<agent>/ subdirectory of this branch's inbox, creating the directories as needed. Copy the file's bytes unchanged, under its own name.

   • MEMORY.md is skipped. It is an index for a tree nothing reads back, and the inbox keeps no index of its own.
   • A same-named file in the inbox is overwritten. Two runs of the same agent on one branch — a first and a second issue-fixer round — write the same subdirectory, and the later run saw more.

5. Commit nothing. This skill never runs git add, git commit, or git push, and it never writes inside the repository.

Output

One block:

## Agent memory captured

Inbox:   <inbox path for this branch>
Entries: <N>
  - <plugin>-<agent>/<file>

Dropped (session-scoped) (<N>):
  - <plugin>-<agent>/<file> — <what about it was local to this run>

Name every drop: it is the one record anyone gets of an entry no later stage will ever see. Report that nothing was captured when no entry survived, or when there was none to begin with.