feat(harness): PI + gptme + mini-SWE wave — 3 columns, board 32/45 (RFC-006 Phase 5)#57
Merged
Merged
Conversation
…006 Phase 5)
Three model-agnostic harnesses built + smoked IN PARALLEL by a 3-agent team
(offloading the per-harness Docker recipe-discovery from the main context). All
edit the real cwd; the patch is captured by host git-diff.
- pi (@earendil-works/pi-coding-agent): models.json via config_files +
PI_CODING_AGENT_DIR; native tool_calls → runs devstral/codestral/qwen3-235b/
glm-4-32b (qwen3-coder-30b emits text-format tool calls its proxy backend can't
parse → pi no-ops; a real compat finding).
- gptme: env-only (OPENAI_BASE_URL + MODEL=local/<m>, prefix stripped); tiktoken
cache baked (no-egress); no turn cap → wall-clock bounds it (patch written
before the verify loop, so a timeout-kill still yields a valid patch).
- mini-SWE-agent: the bare validator loop (L0+L2+L7). --environment-class local
(bash in /workspace, NO nested Docker) + litellm_textbased (fenced bash, not
native tool_calls) + MSWEA_CONFIGURED/COST_TRACKING guards. Trajectory → /tmp.
Cross-harness finding: a model's PROXY BACKEND tool-call-parser determines native-
tool support — qwen3-coder-30b returns text-format calls (explains opencode 1/4);
devstral/codestral/qwen3-235b/glm-4-32b emit native tool_calls.
ruff + mypy --strict clean; 15/15 stack specs valid. supported_harnesses =
{aider, goose, opencode, crush, cline, pi, gptme, mini-swe}.
Refs: rfc-006-stack-executor
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The 3-harness wave's scored columns (--add-stack each), merged into board.json: - PI (native tool_calls): devstral 7.46, codestral 6.67, qwen3-235b 6.17 (3/4; glm-4-32b fail). Native-tool models only — the tool-parser finding holds. - gptme (text-tolerant): codestral 7.17 (1/4 — no turn cap → verify-loop timeouts on the other models; wall-clock kills before a clean finish). - mini-SWE (textbased loop): qwen3-coder-30b 7.08, qwen-3-14b 6.54 (2/4). Board now 32/45 scored, 8 harness columns. Harness display name pi→PI. Refs: rfc-006-stack-executor Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
Three model-agnostic harnesses — PI, gptme, mini-SWE-agent — built + smoked
in parallel by a 3-agent team (offloading per-harness Docker recipe-discovery),
integrated serially. Board now 32/45 scored, 8 harness columns.
Results (be_01)
★ The tool-call-parser finding (explains the whole matrix)
A model's proxy/vLLM backend decides whether it emits native
tool_callsortext-format (XML/markdown fences). qwen3-coder-30b emits text → native-tool
harnesses (PI, opencode) no-op on it (this is why opencode is 1/4!). Native-tool
models on the proxy: devstral, codestral, qwen3-235b, glm-4-32b. Text-tolerant
harnesses (aider, goose, crush, gptme, mini-SWE-textbased) work on text-format models.
How
Docker) + litellm_textbased; MSWEA guards; trajectory → /tmp.
Verify
734 tests green; ruff + mypy --strict clean; 15/15 stack specs valid.
Refs: rfc-006-stack-executor
🤖 Generated with Claude Code