retrospect-external-babysitter-run
For a repository in the babysitter-users catalog, locate its babysitter processes and any committed runs (.a5c/runs/<runId>/) and perform a retrospective on a chosen run -- what went well, what failed, process suggestions, quality of effect design, breakpoint patterns -- mirroring the /babysitter:retrospect workflow but applied to an external repo. Invoke when asked to "retrospect on repo X's run", "analyze how someone else used babysitter", or "review an external babysitter run".
a5c-ai · AI review on file
Installation
Read the source and dependencies before running a command. Copying does not install or execute anything.
npx skills add a5c-ai/babysitter --skill retrospect-external-babysitter-runAbout this skill
name: retrospect-external-babysitter-run description: For a repository in the babysitter-users catalog, locate its babysitter processes and any committed runs (.a5c/runs//) and perform a retrospective on a chosen run -- what went well, what failed, process suggestions, quality of effect design, breakpoint patterns -- mirroring the /babysitter:retrospect workflow but applied to an external repo. Invoke when asked to "retrospect on repo X's run", "analyze how someone else used babysitter", or "review an external babysitter run".
Retrospect External Babysitter Run
Analyse a babysitter run that lives in an external public repository, using the same lens as the in-repo /babysitter:retrospect command. Produce a written retrospective with concrete suggestions for the process author (or, if the insight generalizes, for the babysitter project itself).
When to use
- User names an external repo and asks for a retrospective.
- User asks "find a babysitter run to retrospect on" (combine with the
catalog-babysitter-usersskill to pick one). - User asks "how are other people using babysitter processes? What do they get wrong?".
Prerequisites
ghCLI authenticated.docs/repo-with-babysitter-processes.mdexists (if not, run thecatalog-babysitter-usersskill first).- A workspace directory where external repos can be shallow-cloned (default:
/tmp/babysitter-retrospect/or.a5c/tmp/external-runs/).
Phase 1 -- Target selection
- Read
docs/repo-with-babysitter-processes.mdand list Active repos with stars + description. If the user already named a repo, skip to step 3. - Ask the user which repo to retrospect on (use AskUserQuestion in interactive mode; if non-interactive, pick the highest-starred Active repo that wasn't retrospected in the last 30 days -- track via
docs/retrospectives/<owner>-<name>/log.md). - Confirm the target with the user before cloning.
Phase 2 -- Discover processes and runs
Shallow clone the target:
mkdir -p .a5c/tmp/external-runs
cd .a5c/tmp/external-runs
gh repo clone <owner>/<name> -- --depth 50 --single-branch
cd <name>
Locate:
- Process files: files importing
defineTaskfrom@a5c-ai/babysitter-sdk. Use Grep:grep -rl "from '@a5c-ai/babysitter-sdk'" -- . --include='*.js' --include='*.ts'. - Committed runs:
.a5c/runs/<runId>/directories. Many repos gitignore.a5c/runs/entirely -- that's fine; note it and proceed with process-only retrospective. When runs ARE committed, look forrun.json,journal/,tasks/,state/output.json. - Historical runs via git log:
git log --all --diff-filter=A --name-only -- '.a5c/runs/'surfaces runs that existed at some commit even if later cleaned up. Check out the commit that introduced the run if you want the journal content.
Summarize to the user:
- N process files found, by top-level directory
- M run directories present in HEAD; P additional historical runs reachable via git history
- Which runs completed vs failed (grep
RUN_COMPLETED/RUN_FAILEDin the journal)
Phase 3 -- Pick a run to retrospect
If multiple runs exist, ask the user (interactive) or default (non-interactive) to:
- The most recent failed run (highest signal for process improvement), OR
- If no failures, the most recent completed run.
If no runs are committed at all, switch to a process-only retrospective: analyse the process file(s) for quality issues without run evidence. Mark the output clearly as process-only.
Phase 4 -- Load the run
Inspect, in order:
.a5c/runs/<runId>/run.json-- processId, entrypoint, prompt, createdAt.a5c/runs/<runId>/inputs.json-- user intent.a5c/runs/<runId>/journal/*.json-- event stream (RUN_CREATED, EFFECT_REQUESTED, EFFECT_RESOLVED, RUN_COMPLETED / RUN_FAILED). Read every journal entry; it is the authoritative record..a5c/runs/<runId>/tasks/<effectId>/task.json+result.json-- per-effect definition and result.a5c/runs/<runId>/state/output.json(if present) -- final output- The process file referenced by
run.json.entrypoint-- cross-reference against the journal to see what the author intended vs what happened.
Phase 5 -- Retrospective analysis
Mirror the in-repo /babysitter:retrospect workflow. Produce notes under each heading:
5.1 Outcome
- Success / partial success / failure.
- Total iterations, duration, distinct effect count, retry count.
- Final output quality (from
state/output.jsonshape + content).
5.2 What went well
- Effects that resolved on first try.
- Process sections with clear inputs/outputs and no re-runs.
- Useful breakpoints that caught real issues before they propagated.
5.3 What went poorly
- Effects that were re-dispatched (same invocationKey or similar taskId appearing repeatedly).
- Long gaps between EFFECT_REQUESTED and EFFECT_RESOLVED (external bottlenecks).
- Breakpoints that looped (approval -> reject -> retry -> reject).
- Tasks that crashed and what the error category was (Configuration / Validation / Runtime / External / Internal).
- Any RUN_FAILED: trace the last few events and the thrown error.
5.4 Process-quality review
Evaluate the process file itself against these criteria:
- Determinism: does every effect have stable invocation keys (processId:stepId:taskId)? Any non-deterministic branching based on wall-clock time, random, or unpinned env vars?
- Effect granularity: are tasks too coarse (one huge agent task vs several narrower ones) or too fine (dozens of tiny tasks)?
- Idempotency: can the process be re-run safely? Does it use
ctx.task()for all side effects, or does it write files outside a task? - Breakpoint discipline: are breakpoints used to gate irreversible actions? Do they follow the robust rejection pattern (loop with feedback)?
- Error surfacing: does the process throw with useful context, or swallow errors?
- Labels: are task labels meaningful and consistent (enables filtering / observability)?
- Re-use: cou