SgtPooki · April 28, 2026 21:13 · Apr 28, 2026
diff --git a/dirac-gist-final.md b/dirac-gist-final.md
@@ -0,0 +1,75 @@
+# Dirac retry-loop bug — full repro and logs
+
+Companion to [issue dirac-run/dirac#XXX](https://github.com/dirac-run/dirac/issues).
+
+## Environment
+
+- `dirac-cli` 0.3.3, Node v24.11.1, macOS (Darwin 25.3.0)
+- Provider: OpenAI-compatible (vLLM) via `--baseurl https://my-vllm-endpoint/v1`
+- Failing model: `qwen36-27b-aeon` (27B, AWQ-quantized Qwen 3, served by vLLM)
+- Passing model: `qwen3-32b` (32B, served by vLLM)
+- Mode: `--act --yolo --reasoning-effort medium`
+
+## Reproduction
+
+```bash
+mkdir -p /tmp/dirac-repro/src && cd /tmp/dirac-repro
+echo '{"name":"repro","version":"0.0.1"}' > package.json
+echo "# repro" > README.md
+echo "export const x = 1" > src/index.ts
+echo "export const y = 2" > src/util.ts
+git init -q && git add -A && git commit -qm init
+
+dirac auth -p openai -k EMPTY -b <vllm-base-url> -m <model>
+dirac -y -a -c /tmp/dirac-repro --max-consecutive-mistakes 6 \
+  "Investigate this repo. First list all files. Then read each TypeScript file. Then summarize the structure and any quality issues."
+```
+
+## Result with `qwen36-27b-aeon`
+
+```
+Task started: 1777410056605
+Error: Dirac tried to use attempt_completion without providing a value for 'result'. Retrying...
+Error: Dirac tried to use attempt_completion without providing a value for 'result'. Retrying...
+Dirac tried to use attempt_completion without providing a value for 'result'. Retrying...
+
+Task Summary:
+Tokens: 37,419 in, 1,089 out
+```
+
+Aborted after 6 consecutive mistakes. Zero file reads, zero useful output.
+
+## Result with `qwen3-32b` (same prompt, same repo, same flags)
+
+```
+Repository investigated. It's a minimal 4-file TypeScript project:
+Files: package.json, src/index.ts, src/util.ts, README.md
+TypeScript files analyzed:
+- src/index.ts: exports constant `x = 1`
+- src/util.ts: exports constant `y = 2`
+Quality issues identified:
+1. No tsconfig.json for TypeScript configuration
+...
+
+Task Summary:
+Tokens: 29,575 in, 984 out
+```
+
+Completes cleanly.
+
+## Second observation (real repo, separate task)
+
+Driving Dirac on a multi-package monorepo (CI-fix task on a dependabot PR) with `qwen36-27b-aeon`:
+- v1 invocation: hit the same loop on `list_files` missing `paths`. Aborted after the retry budget. 53,296 input / 715 output tokens consumed, no work done.
+- v2 invocation with the prompt explicitly steering away from `list_files` ("use `execute_command` + `ls` instead") and `--max-consecutive-mistakes 10`: completed end-to-end. 167,664 input / 1,632 output. Single working commit, all CI commands green locally.
+
+So the retry loop is reproducible on real workloads, and the workaround (prompt the small model away from the problematic tool) actually works.
+
+## Possible explanation (hypothesis)
+
+Dirac's tool docs in the system prompt are large (line-hash protocol + `edit_file` examples). The `Missing value for required parameter '<x>'. Please retry with complete response.` feedback IS injected as a tool error on each retry — verified by reading `dirac-cli/dist/cli.mjs` (`missingToolParameterError`). At 27B, the model appears to not act on that feedback, possibly because:
+- The original tool docs still dominate attention on resampling.
+- The repair message is generic; it doesn't include the malformed call or an example of the correct shape.
+- The error is delivered as a tool result, not as a strong steering directive.
+
+This is hypothesis, not diagnosis — would need to instrument the chat history or test multiple repair-message phrasings to confirm.
No results found