ganjihong
beac59fc8b
Phase 4 of the decoupling refactor (behavior-preserving):
- profile_schema.json: every operation contract now carries
runtime_capability {body_mutating, requires_active_body,
replayable, requires_selector, open_profile_ok} (single source of truth)
- operation_contracts.py: validates and forwards the flags
- capabilities.py: the five data-classification frozensets are now
derived from the schema at import time; dispatch-logic sets
(_HOLE_ATOMICS, _PATTERN_ATOMICS, extent constants) stay in code
- test_profile_schema.py: completeness + structural invariants test
Equivalence proven by flag counts (27/13/25/6/2) matching the previous
hand-written sets and by the unchanged test-baseline failure set.
Backend
The backend owns the application API and CAD generation workflow:
app/: HTTP API, jobs, orchestration, and persistence adapters.agent/: AI prompts, tools, and skills used by the generation agent.engine/: CDSL compiler, sketch solver, and STEP generation runtime.cdsl_library/: Official CDSL examples, metadata, and search index.tests/: Engine, API, and end-to-end generation tests.
Expected development entrypoint: app.main:app, served by Uvicorn.
Reasoning Effort
The backend uses the Chat Completions API. Configure a provider's reasoning
budget with CDSL_<PROVIDER>_REASONING_EFFORT; for the current OpenAI setup:
CDSL_OPENAI_REASONING_EFFORT=medium
Use low, medium, or high according to the latency/cost versus quality
tradeoff. The setting is sent as Chat Completions' reasoning_effort field to
authoring, streaming, and visual-review requests. Leave it empty to use the
provider/model default. The selected OpenAI-compatible endpoint must support
the requested value.
Autonomous CDSL Agent Configuration
The autonomous agent writes one frozen free-form requirements.md, then
observes, measures, renders and appends one CDSL feature at a time. Its author
uses normal function calls; no provider strict JSON Schema capability or
complete modelling DAG is required. Candidate fragments are rebuilt in a
staging directory through cdsl_only before a checkpoint can be committed.
Final publication requires a separately configured vision-capable review model and the Python OpenCascade/Pillow technical renderer. The agent may build and inspect intermediate checkpoints without image review; a final run fails closed if its independent review configuration is unavailable.
# Must name one configured provider and one model listed in that provider's
# CDSL_<PROVIDER>_VISION_MODELS setting. It is intentionally not inferred
# from the authoring model.
CDSL_REVIEW_PROVIDER=deepseek
CDSL_REVIEW_MODEL=deepseek-v4-flash-vision-exp
CDSL_DEEPSEEK_VISION_MODELS=deepseek-v4-flash-vision-exp
# Install Python rendering dependencies. The renderer reads the revision STEP
# file and creates canonical images without a browser or GPU driver.
pip install -r requirements.txt
# Limits apply to the current checkpoint head, never to total task complexity.
CDSL_AGENT_TOOL_CALLS_PER_CYCLE=12
CDSL_AGENT_CANDIDATE_ATTEMPTS_PER_HEAD=3
CDSL_AGENT_CONSECUTIVE_NO_PROGRESS_LIMIT=6
CDSL_AGENT_MAX_FEATURES_PER_FRAGMENT=6
CDSL_AGENT_CONTEXT_CHAR_LIMIT=24000
CDSL_AGENT_RENDER_CACHE=true
The author chooses each coherent 1-6 feature batch. Every rebuilt batch is
rendered and independently reviewed before it can become a checkpoint; only
an accepted reviewer verdict advances the working model. Every checkpoint is rebuilt from its fully materialized CDSL through the
cdsl_only runtime. Checkpoint GLB files are preview-only; STEP, CDSL, and
reports are available only after the task reaches COMPLETED.
The backend assigns feature and sketch IDs, appends causal dependencies and expands only opaque current-snapshot selector tokens. It does not compile geometry templates or correct workplanes, profiles, sizes, directions or boolean semantics authored by the model. Failed candidates remain auditable but never become revisions.