Let an agent improve its own workflow

A debugging workflow repairs its own navigation helper, then uses the repaired helper to diagnose and fix an application.

The helper reads a failed run's captured evidence before any model call. You end with a checked candidate for each fix and the source files you started with.

Run the example

Initialize a fresh copy and run the script with your OpenRouter environment file:

ptc init debug-a-failed-run --example debug-a-failed-run
sh debug-a-failed-run/run-self-improvement.sh /absolute/path/to/.env

The script captures two failures, lets an agent repair the helper, checks that helper on two captures without a model, and then diagnoses and repairs the application. A successful run ends with:

Completed: helper checks, trace navigation, and three application validation cases. Artifacts: self-improvement-results

Model runs vary. The script stops at the first failed stage and leaves its artifacts in self-improvement-results.

Read the proposed edit

Print the component the agent chose and its explanation:

ptc repl --project debug-a-failed-run/self-improver.ptc-project.json \
  --profile private-run-analysis-v2 --private-unattended \
  -e '(let [r (first (get (analysis/runs {"status" "ok"}) "items"))] (select-keys (get-in (analysis/open (get r "run_id")) ["result" "value"]) ["component_id" "cause"]))'

{"cause" "debug.start/context blindly selected the first relationship from generated_sources without checking rel, state, or filters, attempting to follow an unavailable relationship" "component_id" "debug.start"}

Add "candidate_source" to the key list to print the replacement source. The example README explains what navigation means, shows what the agent saw before it wrote that edit, and inspects the application fix the same way.

See debug navigation for the evidence contract and components for how a candidate is checked.