A workspace for human developers, engineers and AI tools creating *import conversion extensions. Desktop, command line and Python share the same authoring services.
Startup searches known Python environments. Locate PyCForge accepts an environment, interpreter or installed landscape folder. Engine connection includes Find installed, This Python and the interpreter path. Windows virtual environments use Scripts/python.exe; Linux/macOS use bin/python. PyCForge shows its path in About → Distribution → SDK Python. Failed discovery preserves the last connection. Continue offline opens local projects and cached reference.
Refresh reads standardized files from the selected installation: index.json, api.xml/api.xsd, authoring.json/authoring.schema.json, workbench.json/workbench.schema.json and inventoried Help. Size and SHA-256 checks bind the snapshot. Compatible operations, forms and starters arrive as data without an SDK release. Unknown required protocols are refused before cache replacement. Optional metadata is preserved.
Choose Author-defined measurement formulas in New Project. The starter provides calibration, interval checks and guarded ratios. In Functions, edit Expression to define the result and Bindings for ordered intermediate values. These fields use scalar-declaration/0.4 with implementation expression; keep the native symbol empty. Named parameters and numeric defaults use the existing parameter controls.
bindings: [{"name": "corrected", "expression": "value - offset"}]
expression: corrected * gain
Use +, -, *, /, unary signs, single comparisons, Boolean and/or/not, conditional expressions and abs(x), sqrt(x), min(x,y), max(x,y). Inputs convert to finite binary64; results are float64 or bool. Bindings evaluate once in order. Only selected Boolean/conditional branches execute; conditions require bool and branch types must agree. Each declaration is limited to four parameters, sixteen bindings, 256 expression nodes, depth sixteen and 4096 characters across expressions. Calls to other authored functions, loops, objects and containers are outside this profile.
Nonfinite inputs and negative square roots raise ValueError; zero division raises ZeroDivisionError; nonfinite intermediate results raise OverflowError. Invalid value types raise TypeError. Round-to-nearest and gradual underflow are required. Work-budget exhaustion is fatal. Helpers allocate no memory and retain no inputs. Large integer inputs may round when converted to binary64. Read the engine's composition contract for exact numerical and resource rules.
After changing a formula, update its explanation and independently chosen fixture expectations, then save and explicitly verify conversions again. The SDK preserves expectations without testing runtime output. PyCForge must separately compile and execute relevant programs. Packages containing formulas require an engine supporting scalar-native/0.4.
The SDK never selects, probes or invokes a C compiler, directly or indirectly, and never runs generated programs. PyCForge owns Crucible's compatibility with C compilation. Verification checks conversion results only. Legacy fixture kind run retains author-declared stdout, stderr and exit expectations; these are not evaluated. A passing conversion check or registered package does not establish runtime correctness.
New reports use .pycforge-sdk/conversion-verification.json. Earlier verification.json reports, retained failure inputs and existing packages remain intact. Fixture results can display historical native records when no new report exists. Historical records do not approve new packaging. An engine providing authoring format 1.x is required for conversion verification; older files remain useful for reference and editing.
Save setup stores New Project choices in .pcfsetup. Save project draft stores incomplete fields, Help, expectations, navigation and pending fixture inputs in .pcfdraft. Advanced project data can save incomplete JSON. Open project draft and Review current draft compare original, edited and current data before resuming or saving. A pending fixture needs its own review/apply. Save repair as and Open source repair preserve and compare one fixture source in .pcfrepair.
Keep checkpoints outside the project. They are explicit saves, not autosave or whole-project backups. Later unsaved edits and missing source files need separate recovery. Conflict checks are optimistic; saves are individual file commits, not multi-file transactions. New checkpoint files require hard-link support. Expected results are never replaced automatically by observed output.
Use pycforge-sdk manual for local documentation, --engine-python for a separate installation, and global --json for structured outcomes. Python clients import SDK from pycforge_sdk. Read status and diagnostics before using results; nonzero console exit codes mean the operation did not succeed. Incomplete, canceled and stale reports never approve packaging. Source and field locations retain the engine's diagnostic facts.
Refresh and explicitly verify conversions after project or engine changes. Runtime qualification belongs to PyCForge. The SDK contains no AI model or exemption for automated authors. Use the selected engine's conversion coverage and import manuals for exact supported operations.
Choose the Lazy record pipelines starter to expose chain, islice and pairwise under your namespace. IteratorFunction selects a reviewed kernel; the engine supplies positional call shapes, shared state, retained-input ownership and exhaustion rules. The bundled itertools package demonstrates the same public profile. Verify conversions does not execute iterator behavior; native qualification belongs to PyCForge.