delegator-eval-loop

---

name: delegator-eval-loop description: "Fully autonomous eval loop for the delegator agent. Runs real-mode tests, analyzes output for usability gaps, spawns fix agents, rebuilds, and re-tests until clean. disable-model-invocation: true

Delegator Eval Loop

Autonomous test→diagnose→fix cycle for the delegator agent. Max 5 rounds. Exit early when all gaps are closed.

Loop

1. Build: bazel build //bramble:bramble

2. Read context: docs/design/delegator-real-mode-testing.md and bramble/cmd/delegator/delegator.go

3. Check prior runs: Read .claude/skills/delegator-eval-loop/data/eval-runs.log for patterns from past iterations

4. Run tests — choose the right harness for the eval type:

   Codewalk eval (multi-turn, questions from file):

   python3 scripts/delegator-eval.py \
     --questions-file scripts/delegator-eval-questions.txt \
     --work-dir /path/to/repo \
     --model sonnet --child-model gemini-3-flash-preview \
     --log-dir "$LOG_DIR" --timeout 900 2>"$LOG_DIR/stderr.txt"
   

   The script drives interactive mode via PTY + --status-fd pipe. It sends questions one per idle, waits for children to complete, then sends quit. Edit scripts/delegator-eval-questions.txt to change the question set.

   Single-prompt eval (one-shot task, non-interactive):

   echo "<PROMPT>" | bazel-bin/bramble/bramble_/bramble delegator \
     --mode real --work-dir "$TEST_DIR" --model sonnet \
     --log-dir "$LOG_DIR" --timeout 8m 2>"$LOG_DIR/stderr.txt"
   

   Each single-prompt test needs a fresh git init'd directory with a minimal go.mod + main.go.

   Run sequentially (real API calls). Use run_in_background + timeout=600000 so you can analyze completed results while the next runs.

5. Analyze — read .claude/skills/delegator-eval-loop/references/checklist.md and check every item against both rendered output and JSONL logs. Check .claude/skills/delegator-eval-loop/references/known-gaps.md for regressions.

6. Exit check — zero high/medium findings, tool restriction correct, no regressions, costs reasonable → exit to summary

7. Fix — spawn Plan agents for diagnosis, then worker agents (sonnet) for implementation. Fix root causes, not symptoms.

8. Quality gates (must pass before re-testing):

   scripts/lint.sh
   bazel test //... --test_timeout=60
   bazel build //bramble:bramble
   

9. Commit, then go to step 4 with the rebuilt binary and fresh test directories.

Final Summary

Always produce a summary table (rounds, findings fixed, remaining gaps, verdict). After a successful loop, add a "Test Run N" section to docs/design/delegator-real-mode-testing.md. Append results to .claude/skills/delegator-eval-loop/data/eval-runs.log.

Gotchas

1. JSONL is truth, not rendered output. The JSONL session logs are ground truth for tool counts, session lifecycle, and costs. Rendered output can mask issues (e.g., a tool might execute but not display).

2. Text fragmentation is subtle. Streaming TextEvent chunks have arbitrary boundaries. Look for concatenated words like "TheI" or "GreatHere" — these indicate the render layer isn't buffering at word boundaries. See CLAUDE.md's "Streaming Text Event Pipeline" section.

3. Fix root causes, not symptoms. If the delegator calls ToolSearch, don't filter it from display — fix why it's available. If stderr is noisy, fix log levels at the source, don't redirect.

4. The 8-minute timeout is real. Complex prompts can hit it. If a test times out, check whether the child session hung (JSONL will show no end_turn after the last message) vs. the task was genuinely too large.

5. Fresh directories for single-prompt evals. Never reuse test directories between single-prompt tests. Git state from a prior run will confuse the delegator's child sessions. Codewalk evals use the real repo as work-dir (read-only codetalk sessions).

6. Tests must run sequentially. Each test makes real Claude API calls. Run one at a time, but use run_in_background + timeout=600000 so you can analyze completed results while the next runs.

7. Capture stderr separately. The delegator's stderr contains log output that won't appear in stdout or JSONL. Always use 2>"$LOG_DIR/stderr.txt" and check it for unexpected warnings.

8. Model name format matters for non-Claude children. Child sessions inherit the model from delegator config. Verify via JSONL init messages that children got the expected model, not a default.

Key Files

Area	Files
Harness & rendering	bramble/cmd/delegator/delegator.go
Multi-turn eval	scripts/delegator-eval.py + scripts/delegator-eval-questions.txt
Session setup	bramble/session/delegator_runner.go
Tool handlers	bramble/session/delegator_tools.go
Session lifecycle	bramble/session/manager.go
SDK options	agent-cli-wrapper/claude/session_options.go
CLI arg building	agent-cli-wrapper/claude/process.go
Design doc	docs/design/delegator-real-mode-testing.md
Analysis checklist	.claude/skills/delegator-eval-loop/references/checklist.md
Known gaps	.claude/skills/delegator-eval-loop/references/known-gaps.md
Run history	.claude/skills/delegator-eval-loop/data/eval-runs.log