Moved Gemma4-12B from OpenCode primary model to MCP tool.
This avoids GUI issues and token burn from monitoring.
MCP tool: gemma4
- Calls local llama-server at 127.0.0.1:8081
- Returns reasoning + content
- No API key required
- ~40 tokens/sec generation
Usage: call gemma4 tool with a question, get answer back.
No parent model burns tokens waiting.
Hermes already handles model routing. Remove MCP server, keep skill
as a cost-optimized model selection guide.
Key insight: DeepSeek V4 Pro is 55x cheaper than Claude Opus and
nearly as good at math/code. Default to DeepSeek, escalate to Opus
only when stuck.
Removed: break_glass_mcp.py (unnecessary — Hermes handles routing)
- Updated .mcp.json with WOLFRAM_APP_ID env var (fallback to hardcoded HYJE3R3R63)
- Toolchain manifest documents all tools used in hachimoji citation pipeline
Build: 3314 jobs, 0 errors (lake build)
- Replace repo .mcp.json with minimal config (10 essential servers instead
of 29). Full backup retained as .mcp.json.full.
- Remove inline MCP section from ~/.config/opencode/config.json so OpenCode
loads only repo .mcp.json (also removes hardcoded Linear token leak).
- Add singleton guards to opencode_prover_mcp.py, token_saver_mcp.py, and
remote_lean_proof_mcp.py to prevent duplicate instances when multiple
IDE configs reference them.
Build: N/A (Python shims)
Drop the pist-trace-classify server (and any lingering aws block) from
.mcp.json and opencode.json. It depended on RDS/AWS env vars. Already
removed from ~/.kimi/mcp.json as well.
Replace the TODO(lean-port) sorry with a complete proof of the
projectionOrdering theorem: for positive SourceValue pairs s1 < s2
with s2 ≤ maxExpected, projectToCoding preserves strict ordering
of the Q0_64 values.
The proof uses Nat-only arithmetic (no Float) and handles two cases:
- a2 < d: both values fit in Q0_64 range, ordering follows from
monotonicity of integer division
- a2 = d: a2*s/d = s clamped to q0_64MaxRaw; a1*s/d < q0_64MaxRaw
via the key inequality (d-1)*s < (s-1)*d
Build: 8598 jobs, 0 errors (lake build)
This squashes all local history (768 commits) onto the scrubbed PR #90
baseline. Individual commits were lost during filter-repo corruption;
the working tree content is preserved intact.
Build: N/A (working tree state only)
Adds automated guardrails so mathematical rigor is enforced by tooling
instead of by convention. See docs/math-first-tooling.md for the full
contract.
Schemas + registry:
- shared-data/schemas/deepseek-review-receipt.schema.json
Draft 2020-12 schema for the existing ollama_deepseek_review_receipt_v1
and ollama_deepseek_review_continuation_receipt_v1 receipt formats. Pins
sha256:<hex> hashes, non-negative token counts, repo-relative POSIX
paths, and rejects additional fields.
- shared-data/schemas/claims-registry.schema.json
Schema for claims.yaml. Requires review_receipts when status is
verified-by-ai and a lean source when status is formally-proven.
- claims.yaml
Initial registry entry: prime-gap-entropy-collapse (verified-by-ai)
linked to the two existing receipts under
shared-data/artifacts/deepseek_review/.
Validators (scripts/math-first/):
- validate_deepseek_receipts.py: validates tracked or passed receipts
against the JSON Schema; shared by pre-commit and CI.
- test_validate_deepseek_receipts.py: positive + 7 negative fixtures
asserting exit-code behaviour.
- validate_claims_registry.py: schema check + unique id check + on-disk
existence check for every referenced repo-relative path.
- require_math_evidence.py: gate that requires a DeepSeek receipt, a
Lean change, or a claims.yaml update alongside edits to math-track
surfaces (Lean Semantics kernels, ArithmeticSpec docs, stack
solidification receipts).
Pre-commit (.pre-commit-config.yaml):
- check-json, check-yaml, end-of-file-fixer, trim trailing whitespace,
detect-private-key (scoped to math-first files only per AGENTS.md
Do Not Sweep).
- Local hooks wiring all three math-first validators above.
CI (.github/workflows/math-check.yml):
- validate-schemas: compiles every schema, runs both validators, runs
the validator self-tests, then re-invokes the canonical Ollama
emitter in --verify-only mode against every tracked receipt to
re-check answer_sha256 against the answer-file bytes on disk.
- require-evidence: enforces the math-track evidence rule at PR scope.
- pre-commit: runs all pre-commit hooks against the PR diff so the
contract holds even for contributors who skip installing hooks
locally.
MCP (.mcp.json):
- filesystem, sympy, wolfram-alpha, lean, deepseek-review entries
pointing at off-the-shelf upstream servers and at the canonical
ollama_deepseek_review_emitter.py. Secrets stay in the runtime env
(WOLFRAM_ALPHA_APPID, OLLAMA_API_KEY) and are never embedded.
Docs (docs/math-first-tooling.md):
- Philosophy, surfaces, schema reference, registry workflow, hook
catalogue, CI catalogue, MCP catalogue, end-to-end verify command.
shared-data/schemas/*.schema.json and claims.yaml live under paths the
top-level .gitignore would normally exclude; they are force-added via
git add -f the same way existing promoted receipts under
shared-data/artifacts/deepseek_review/ are tracked (per AGENTS.md).
Co-Authored-By: Allaun Silverfox <bigdataiscoming+9i37y6j2@protonmail.com>
