feat(tasks): add BFCL v3 — single-turn function calling, both protocols - #143

Merged

ethan-scitix merged 28 commits into

mainfrom

worktree-bfcl-v3

Sep 11, 2026

Commits on Sep 11, 2026

feat(core): let a task declare that it requires native function calling

ethan-scitix
and
claude
committed
9ccf7e3
feat(community): vendor the BFCL v3 AST checker from gorilla v1.3

ethan-scitix
and
claude
committed
aae9199
fix(community): correct three misstated BFCL v3 provenance docstrings

ethan-scitix
and
claude
committed
db73ebb
feat(community): vendor BFCL's Java and JavaScript source parsers

ethan-scitix
and
claude
committed
c037c47
feat(datasets): add BFCL v3 single-turn non-live and live datasets

ethan-scitix
and
claude
committed
4901912
feat(core): carry native tool calls through to the legacy model output

ethan-scitix
and
claude
committed
c4f5189
fix(datasets): stop the bfcl-v3 extra from taking out the dataset registry

ethan-scitix
and
claude
committed
7c2c640
test(datasets): cover the cross-category and per-file BFCL v3 guards

ethan-scitix
and
claude
committed
0b324c2
fix(community): keep `ast_parse` statically visible behind the lazy import

ethan-scitix
and
claude
committed
95e9add
fix(datasets): assert the BFCL v3 gold count and name the rows that collide

ethan-scitix
and
claude
committed
445fad1
feat(tasks): add the BFCL v3 shared stages for both protocols

ethan-scitix
and
claude
committed
3cc0676
docs(deps): correct when the `bfcl-v3` extra is actually reached

ethan-scitix
and
claude
committed
f999135
test(tasks): pin that a missing gold fails the row rather than scoring it

ethan-scitix
and
claude
committed
f5635fc
feat(tasks): roll BFCL v3 up the way upstream scores it

ethan-scitix
and
claude
committed
7392109
feat(tasks): register the four BFCL v3 single-turn tasks

ethan-scitix
and
claude
committed
f3d6a92
fix(tasks): declare the tree-sitter extra on the two BFCL v3 Prompt leaves

ethan-scitix
and
claude
committed
aa85279
test(tasks): replay BFCL v3 gold as the reply to check the grader wiring

ethan-scitix
and
claude
committed
e2a1c5f
docs(tasks): add the _fc variant row to the naming vocabulary

ethan-scitix
and
claude
committed
82c93cb
fix(scripts): exempt vendored relative imports, and count violations not lines

ethan-scitix
and
claude
committed
95127e5
docs(scripts): say which rule the community exemption actually protects

ethan-scitix
and
claude
committed
838d6ce
fix(tasks): apply upstream's model-facing preprocessing on both BFCL protocols

ethan-scitix
and
claude
committed
9841907
fix(community): stop the BFCL v3 Prompt decoder from executing model output

ethan-scitix
and
claude
committed
48a8603
fix(tasks): stop the BFCL v3 decode guard from scoring a broken environment

ethan-scitix
and
claude
committed
985c73c
fix(tasks): stop BFCL v3's shipped notes from claiming an unmodified grader

ethan-scitix
and
claude
committed
3246159
fix(tasks): reword BFCL v3's excluded-category list so it stops reading as a path

ethan-scitix
and
claude
committed
6b83864
fix(community): spell BFCL v3's cross-package import absolutely, and bound string `%`

ethan-scitix
and
claude
committed
a8a7bd7
feat(tasks): promote BFCL v3 to stable, anchored on upstream's released rollouts

ethan-scitix
and
claude
committed
e167624
docs(tasks): cut the BFCL v3 prose back to what the code does not already say

ethan-scitix
and
claude
committed
56ecec3

Navigation Menu

feat(tasks): add BFCL v3 — single-turn function calling, both protocols - #143

feat(tasks): add BFCL v3 — single-turn function calling, both protocols#143
ethan-scitix merged 28 commits into
mainfrom
worktree-bfcl-v3

Commits on Sep 11, 2026

feat(core): let a task declare that it requires native function calling

feat(community): vendor the BFCL v3 AST checker from gorilla v1.3

fix(community): correct three misstated BFCL v3 provenance docstrings

feat(community): vendor BFCL's Java and JavaScript source parsers

feat(datasets): add BFCL v3 single-turn non-live and live datasets

feat(core): carry native tool calls through to the legacy model output

fix(datasets): stop the bfcl-v3 extra from taking out the dataset registry

test(datasets): cover the cross-category and per-file BFCL v3 guards

fix(community): keep `ast_parse` statically visible behind the lazy import

fix(datasets): assert the BFCL v3 gold count and name the rows that collide

feat(tasks): add the BFCL v3 shared stages for both protocols

docs(deps): correct when the `bfcl-v3` extra is actually reached

test(tasks): pin that a missing gold fails the row rather than scoring it

feat(tasks): roll BFCL v3 up the way upstream scores it

feat(tasks): register the four BFCL v3 single-turn tasks

fix(tasks): declare the tree-sitter extra on the two BFCL v3 Prompt leaves

test(tasks): replay BFCL v3 gold as the reply to check the grader wiring

docs(tasks): add the _fc variant row to the naming vocabulary

fix(scripts): exempt vendored relative imports, and count violations not lines

docs(scripts): say which rule the community exemption actually protects

fix(tasks): apply upstream's model-facing preprocessing on both BFCL protocols

fix(community): stop the BFCL v3 Prompt decoder from executing model output

fix(tasks): stop the BFCL v3 decode guard from scoring a broken environment

fix(tasks): stop BFCL v3's shipped notes from claiming an unmodified grader

fix(tasks): reword BFCL v3's excluded-category list so it stops reading as a path

fix(community): spell BFCL v3's cross-package import absolutely, and bound string `%`

feat(tasks): promote BFCL v3 to stable, anchored on upstream's released rollouts

docs(tasks): cut the BFCL v3 prose back to what the code does not already say