Claude/supervisorai minimax m3 dcgeo2 - #3
Open
hungryshmorez wants to merge 13 commits into
Open
hungryshmorez wants to merge 13 commits into
hungryshmorez wants to merge 13 commits into
Conversation
A bare `pytest` was wandering into the gitignored deliverables/ directory and the bundled vortex-os* bash skills, causing a collection error. Restrict testpaths to tests/ so CI and local `pytest` only see the real suite. Suite: 46 passed. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MQBY7mMBZS3RVwotbB75Zm
docs/DEMO.md: exact commands, scene-by-scene beats (live M3 trace, real deliverables + passing generated tests, the VORTEX-OS CLI), and a suggested post tagging the sponsors. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MQBY7mMBZS3RVwotbB75Zm
Drop the supervisor-robot logo at the top of the README and store it under docs/assets/. The two submission artifacts (one-pager + Codex) were re-skinned to the logo's palette — warm cream, charcoal, steel-blue. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MQBY7mMBZS3RVwotbB75Zm
docs/index.html — the interactive one-pager (replay the M3 run, approve the HITL gate). docs/codex.html — the full reference encyclopedia. Cross-linked, brand- skinned, self-contained (logo embedded). .nojekyll serves the raw HTML. Enable at: Settings → Pages → Deploy from a branch → /docs. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MQBY7mMBZS3RVwotbB75Zm
Add a links table to the README covering the live demo, the Codex encyclopedia, the repo, PR, demo runbook, sample trace, and both VORTEX-OS skills. Add resource links on the pages themselves — the demo footer links to the Codex + GitHub, and the Codex sidebar/footer link to the demo, repo, and PR. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MQBY7mMBZS3RVwotbB75Zm
Rewrite the README credits: created & directed by 12matt3r (http://doesntmatter.us); originally built for MiniMax's Agent Hackathon (Honorable Mention), rebuilt since; software design & engineering of this build by Claude (Claude Code); reasoning by MiniMax-M3 on GMI Cloud. Kept the original hackathon build credit (Jules). Add the same credit line to both websites (demo + Codex), and update the VORTEX-OS skill author metadata from "MiniMax Agent" to credit 12matt3r. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MQBY7mMBZS3RVwotbB75Zm
Theme 1 — architectural hierarchy (Python core vs. bash clients): - Moved vortex-os/ and vortex-os-v3/ into implementations/ so they're no longer visual peers of src/. Added implementations/README.md framing them as downstream command-line clients. - README: reframed the bash skills as a bottom "Command-line implementations (experimental)" section with an explicit "which should I use?" recommendation (Python for production, bash for local experiments); Explore table now leads with the core engine; added a "what runs the system" orientation. - Updated pyproject norecursedirs, .gitignore, docs/DEMO.md paths, and both websites (core-vs-CLI framing, implementations/ paths). Theme 2 — configuration & the silent-mock footgun: - scripts/preflight.py: validates env + JSON configs and sends one real request to confirm live MiniMax-M3 (not the silent mock); --require-live exits non-zero for CI. Verified live (exit 0, "READY — live MiniMax-M3"). - docs/CONFIGURATION.md: the full configuration matrix (every var/file, where set, what it affects, precedence — one env source of truth, no llm_config.json), a production-readiness checklist, and how to tell mock vs live. - README quickstart now runs preflight; links both new docs. Theme 3 — feedback loop transparency (glass box): - SupervisorCore.update_weights now appends to supervisor_data/weight_history.jsonl (weights already persist to config/weights.json and survive restart). - scripts/plot_weight_drift.py charts the learned-weight drift over time (the "visual indicator" for operator bias); degrades to text without matplotlib. - docs/FEEDBACK_LOOP.md: lifecycle, persistence, version control, and the exact math (lr 0.01; ~5 overrides ≈ 0.05 shift), plus the clarification that the fixed 0.4/0.9 gates are not what overrides move. - Codex: added Configuration and Feedback-loop entries + Operations nav. Suite: 46 passing. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MQBY7mMBZS3RVwotbB75Zm
.github/workflows/ci.yml runs the 46-test suite and the config preflight on every push/PR (no key needed — passes in offline-mock mode). A final step runs `preflight --require-live` only when a GMI_API_KEY repo secret is configured, proving a deploy is on live MiniMax-M3 rather than the silent mock. README: add a grounded "Path to production" section (what's done vs. next). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MQBY7mMBZS3RVwotbB75Zm
fastmcp==2.11.3 requires rich>=13.9.4, but requirements.txt pinned rich==13.7.1, causing pip's resolver to fail (ResolutionImpossible) on a clean install. Loosen rich to >=13.9.4 so the pinned fastmcp resolves. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MQBY7mMBZS3RVwotbB75Zm
tests/test_feedback_trainer.py imported `from src.supervisor_agent...` while the rest of the suite (and the CI PYTHONPATH=src) uses the bare `supervisor_agent` package root. On a clean checkout this raised ModuleNotFoundError: No module named 'src' and aborted collection. Align the import with the rest of the suite. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MQBY7mMBZS3RVwotbB75Zm
When a GMI_API_KEY repo secret is present, GitHub injects it into every CI
step. That pushed the unit suite out of offline-mock mode and into real LLM
calls, so mock-mode tests (placeholder-response, HITL orchestration) failed
non-deterministically depending on whether the secret existed.
Two fixes:
- CI: blank GMI_API_KEY for the Test suite and offline-preflight steps so the
suite is hermetic and deterministic regardless of the repo secret. Live
coverage stays in the dedicated `preflight --require-live` step.
- Client: strip whitespace from the API key. A trailing newline from pasting
a secret into GitHub's secret box produced an illegal HTTP header value
("Illegal header value"), breaking every live call. Keys never carry edge
whitespace, so stripping is always safe.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MQBY7mMBZS3RVwotbB75Zm
12Matt3r
approved these changes
Sep 6, 2026
The GMI_API_KEY Actions secret is now configured and the require-live preflight step runs green on every push, so promote the live-M3 gate from a planned "Next" item to a done one, and narrow the remaining secrets work to graduating from a CI secret to a managed secrets store. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MQBY7mMBZS3RVwotbB75Zm
Give GMI Cloud its own row (was folded into the MiniMax-M3 line) and strengthen Jules's credit for the original Honorable-Mention build, so 12matt3r, Jules, Claude, MiniMax-M3, and GMI Cloud are each represented. Update the footer line to match. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01MQBY7mMBZS3RVwotbB75Zm
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
2 participants
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.
Fin fin