feat(dataset): add CodingContentGenerator for pseudo-realistic coding-trace replay#968
Open
ajcasagrande wants to merge 2 commits into
Open
feat(dataset): add CodingContentGenerator for pseudo-realistic coding-trace replay#968ajcasagrande wants to merge 2 commits into
ajcasagrande wants to merge 2 commits into
Conversation
…-trace replay Adds a token-pool prompt generator that produces structurally plausible, template-filled coding content (code, bash output, JSON, error tracebacks, git diffs, CI output, configs, markdown, test output, user prompts). The output is not real code — it's filler whose token distribution approximates real coding-agent traffic closely enough to drive realistic expert-routing patterns on Mixture-of-Experts models. PromptGenerator's Shakespeare corpus routes to a narrow subset of experts (English-prose experts dominate), which is a poor benchmark proxy for production coding workloads — coding-flavored content activates the broader set of experts that real agentic-coding traffic hits, so MoE serving metrics (per-expert load, hot/cold imbalance, throughput under realistic expert pressure) measured under this corpus are representative of agentic-coding production rather than literature. This completes #839/#943: main already ships the agentic_code_gen session SYNTHESIZER (hash_ids + input/output lengths + delays in Mooncake-style JSONL), but has no coding-text generator on the runtime side. Without this PR, synthesized sessions executed against a real model use Shakespeare filler and the resulting MoE expert-routing pattern is unrepresentative. Added: - src/aiperf/dataset/generator/coding_content.py: CodingContentGenerator (4082 lines; ~673 vocab tuples + ~1386 lines inside triple-quoted code template literals + ~1.3k lines of generator scaffolding) - src/aiperf/common/hash_id_random_generator.py: deterministic per-hash-id RNG used by the generator's tool-pool sampling - tests/unit/dataset/generator/test_coding_content_generator.py: 76 unit tests covering init, generate(), token sequence building, sampling, and per-language template families - PromptCorpus enum (sonnet/coding) in aiperf.common.enums - corpus field on PromptConfig with CLI flag --prompt-corpus (the Pydantic field drops the parent-prefix to match the existing isl/osl/block_size/ batch_size convention; the flat CLI field stays prompt_corpus to match prompt_batch_size / prompt_sequence_distribution) - default_prompt_corpus field on Custom/Public DatasetLoaderMetadata (declared as Literal["sonnet", "coding"] in schemas.py to avoid the aiperf.common.enums -> aiperf.plugin chain circular import) - sample_tokens_from_corpus() module-level helper on prompt.py (used by CodingContentGenerator to build its tool/text pools without sharing PromptGenerator state) - composer/custom.py wiring: when an is_trace loader runs and the effective corpus is CODING, swap in CodingContentGenerator instead of the default Shakespeare PromptGenerator. Reads corpus from the synthetic-dataset prompts config when present and falls back to the loader plugin's default_prompt_corpus. Not included (out of scope for this PR, can land separately): - composer/public.py trace-loader wiring (main's PublicDatasetLoaderMetadata has no is_trace field; branch's _inject_trace_kwargs depends on it) - Weka/semianalysis trace loaders that would set default_prompt_corpus=coding - The Weka-specific BPE-stable-terminator + hash-id RNG threading work on PromptGenerator (Weka trace replay concern, not coding-content) Verified: - pytest tests/unit/ -n auto: 12749 passed, 79 skipped - make validate-plugin-schemas: 33 categories, 219 plugins clean - make generate-all-plugin-files: schemas regenerated - pre-commit run: all hooks pass (ergonomics baseline regenerated to accept the corpus-as-data file/function sizes) Signed-off-by: Anthony Casagrande <[email protected]>
Try out this PRQuick install: pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@7e8edc5215229d2b8cccc73800d22854b4e6ba9eRecommended with virtual environment (using uv): uv venv --python 3.12 && source .venv/bin/activate
uv pip install --upgrade --force-reinstall git+https://github.com/ai-dynamo/aiperf.git@7e8edc5215229d2b8cccc73800d22854b4e6ba9eLast updated for commit: |
WalkthroughThis PR introduces a ChangesPrompt Corpus Selection and Dataset Generation
🎯 2 (Simple) | ⏱️ ~12 minutes
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 4
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@src/aiperf/common/hash_id_random_generator.py`:
- Line 28: Add missing type hints: annotate the __getattr__ method signature as
def __getattr__(self, name: str) -> Any (import Any from typing if not already),
and annotate the constructor as def __init__(self, ...) -> None (add -> None to
its signature). Update imports to include Any when adding the __getattr__ return
type.
- Around line 48-49: from_base_rng incorrectly treats base_rng.seed == 0 as
falsy; change the logic in from_base_rng to check explicitly for None (e.g., use
"if base_rng.seed is not None then base_seed = base_rng.seed else base_seed =
base_rng.randrange(0, 2**64)") so a legitimate seed of 0 is preserved; update
the return path that constructs the instance (reference: class method
from_base_rng and attribute base_rng.seed) to use that explicit None check
instead of "or".
In `@src/aiperf/config/flags/_converter_dataset.py`:
- Around line 82-83: The CLI `--prompt-corpus` value is being added to the local
`prompts` dict (when `cli.prompt_corpus` is set) but later `_apply_dataset_type`
removes `prompts` for file datasets, making `--prompt-corpus` a no-op; update
`_apply_dataset_type` (or the dataset conversion flow that strips `prompts`) to
either preserve `prompts["corpus"]` for supported file/trace loaders (check the
loader type/format and pass `prompts` through to the file/trace loader
constructors) or validate early and raise a clear error when a provided
`cli.prompt_corpus` is incompatible with the chosen dataset type; locate
references to `prompts`, `cli.prompt_corpus`, and `_apply_dataset_type` in
_converter_dataset.py and implement the preservation or explicit rejection
accordingly.
In `@src/aiperf/dataset/generator/prompt.py`:
- Around line 53-60: The wrap logic may return fewer than num_tokens when
num_tokens > corpus_len because it only wraps once; change the logic in the
token-selection section (around start = rng_to_use.randrange(corpus_len), end =
start + num_tokens and the tokens.extend(...) branches) to guarantee exactly
num_tokens elements by iterating or using modulo arithmetic over corpus indices
(e.g., append corpus[(start + i) % corpus_len] for i in range(num_tokens) or
loop extending slices until len(tokens) == num_tokens) so tokens always reaches
the requested length regardless of num_tokens relative to corpus_len.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
Run ID: 23f3a421-fa90-4ad9-aa86-e9e3f3a24bb3
📒 Files selected for processing (16)
docs/cli-options.mdsrc/aiperf/common/enums/__init__.pysrc/aiperf/common/enums/enums.pysrc/aiperf/common/hash_id_random_generator.pysrc/aiperf/config/dataset/content.pysrc/aiperf/config/flags/_converter_dataset.pysrc/aiperf/config/flags/_section_fields.pysrc/aiperf/config/flags/cli_config.pysrc/aiperf/config/schema/aiperf-config.schema.jsonsrc/aiperf/dataset/composer/custom.pysrc/aiperf/dataset/generator/coding_content.pysrc/aiperf/dataset/generator/prompt.pysrc/aiperf/plugin/schema/plugins.schema.jsonsrc/aiperf/plugin/schema/schemas.pytests/unit/dataset/generator/test_coding_content_generator.pytools/ergonomics_baseline.json
| class _DisabledNumpyRNG: | ||
| """Raises on any attribute access to prevent NumPy RNG usage.""" | ||
|
|
||
| def __getattr__(self, name): |
There was a problem hiding this comment.
🧩 Analysis chain
🏁 Script executed:
cat -n src/aiperf/common/hash_id_random_generator.pyRepository: ai-dynamo/aiperf
Length of output: 3480
Add missing function type hints.
__getattr__ on line 28 is missing type annotations for the name parameter and return type. __init__ on line 51 is missing the -> None return type annotation. Per coding guidelines, all functions require complete type hints (params and return).
Proposed fix
+from typing import Never
+
class _DisabledNumpyRNG:
"""Raises on any attribute access to prevent NumPy RNG usage."""
- def __getattr__(self, name):
+ def __getattr__(self, name: str) -> Never:
raise RuntimeError(
"HashIdRandomGenerator does not support NumPy RNG operations. "
"Use Python RNG methods (randrange, choice, etc.) instead."
)
class HashIdRandomGenerator(RandomGenerator):
"""RandomGenerator that re-seeds deterministically per (trace_id, hash_id).
Designed for parallel processing where multiple workers need to generate
identical content for the same hash_id within a trace file.
Thread Safety:
NOT thread-safe. Each worker process must have its own instance.
"""
`@classmethod`
def from_base_rng(cls, base_rng: RandomGenerator) -> "HashIdRandomGenerator":
"""Create from a base RandomGenerator (typically from rng.derive())."""
base_seed = base_rng.seed or base_rng.randrange(0, 2**64)
return cls(base_seed, _internal=True)
- def __init__(self, base_seed: int, *, _internal: bool = False):
+ def __init__(self, base_seed: int, *, _internal: bool = False) -> None:
super().__init__(base_seed, _internal=_internal)
self._numpy_rng = _DisabledNumpyRNG()
self._trace_id: str = ""🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@src/aiperf/common/hash_id_random_generator.py` at line 28, Add missing type
hints: annotate the __getattr__ method signature as def __getattr__(self, name:
str) -> Any (import Any from typing if not already), and annotate the
constructor as def __init__(self, ...) -> None (add -> None to its signature).
Update imports to include Any when adding the __getattr__ return type.
Comment on lines
+48
to
+49
| base_seed = base_rng.seed or base_rng.randrange(0, 2**64) | ||
| return cls(base_seed, _internal=True) |
There was a problem hiding this comment.
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
python - <<'PY'
seed = 0
fallback = 123
print(seed or fallback) # demonstrates current behavior picks fallback instead of 0
PYRepository: ai-dynamo/aiperf
Length of output: 64
🏁 Script executed:
cat -n src/aiperf/common/hash_id_random_generator.pyRepository: ai-dynamo/aiperf
Length of output: 3480
🏁 Script executed:
rg "class RandomGenerator" -A 20 src/aiperf/Repository: ai-dynamo/aiperf
Length of output: 1811
🏁 Script executed:
rg "self\.seed" src/aiperf/common/random_generator.py -B 2 -A 2Repository: ai-dynamo/aiperf
Length of output: 42
🏁 Script executed:
cat -n src/aiperf/common/random_generator.py | head -80Repository: ai-dynamo/aiperf
Length of output: 3380
🏁 Script executed:
cat -n src/aiperf/common/random_generator.py | sed -n '81,150p'Repository: ai-dynamo/aiperf
Length of output: 3100
🏁 Script executed:
cat -n src/aiperf/common/random_generator.py | sed -n '118,122p'Repository: ai-dynamo/aiperf
Length of output: 232
Fix seed value 0 being treated as falsy in from_base_rng.
Line 48 uses base_rng.seed or base_rng.randrange(0, 2**64), which treats seed value 0 as falsy. Since 0 is documented as a valid seed (per RandomGenerator's docstring: "Optional random seed (0 to 2^64-1)"), this breaks reproducibility when seed is legitimately 0. Replace the or operator with an explicit is not None check.
Proposed fix
- base_seed = base_rng.seed or base_rng.randrange(0, 2**64)
+ base_seed = base_rng.seed if base_rng.seed is not None else base_rng.randrange(0, 2**64)📝 Committable suggestion
‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.
Suggested change
| base_seed = base_rng.seed or base_rng.randrange(0, 2**64) | |
| return cls(base_seed, _internal=True) | |
| base_seed = base_rng.seed if base_rng.seed is not None else base_rng.randrange(0, 2**64) | |
| return cls(base_seed, _internal=True) |
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@src/aiperf/common/hash_id_random_generator.py` around lines 48 - 49,
from_base_rng incorrectly treats base_rng.seed == 0 as falsy; change the logic
in from_base_rng to check explicitly for None (e.g., use "if base_rng.seed is
not None then base_seed = base_rng.seed else base_seed = base_rng.randrange(0,
2**64)") so a legitimate seed of 0 is preserved; update the return path that
constructs the instance (reference: class method from_base_rng and attribute
base_rng.seed) to use that explicit None check instead of "or".
Comment on lines
+82
to
+83
| if "prompt_corpus" in s and cli.prompt_corpus is not None: | ||
| prompts["corpus"] = cli.prompt_corpus |
There was a problem hiding this comment.
--prompt-corpus is silently discarded for file-dataset paths
Line [82] stores prompt_corpus under prompts, but file datasets later remove prompts in _apply_dataset_type (Lines [330]-[338]). That makes --prompt-corpus a no-op for --input-file flows instead of applying or failing fast. Please preserve corpus for supported file/trace loaders (or explicitly reject it for unsupported file formats) rather than dropping it silently.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@src/aiperf/config/flags/_converter_dataset.py` around lines 82 - 83, The CLI
`--prompt-corpus` value is being added to the local `prompts` dict (when
`cli.prompt_corpus` is set) but later `_apply_dataset_type` removes `prompts`
for file datasets, making `--prompt-corpus` a no-op; update
`_apply_dataset_type` (or the dataset conversion flow that strips `prompts`) to
either preserve `prompts["corpus"]` for supported file/trace loaders (check the
loader type/format and pass `prompts` through to the file/trace loader
constructors) or validate early and raise a clear error when a provided
`cli.prompt_corpus` is incompatible with the chosen dataset type; locate
references to `prompts`, `cli.prompt_corpus`, and `_apply_dataset_type` in
_converter_dataset.py and implement the preservation or explicit rejection
accordingly.
Comment on lines
+53
to
+60
| start = rng_to_use.randrange(corpus_len) | ||
| end = start + num_tokens | ||
|
|
||
| if end <= corpus_len: | ||
| tokens.extend(corpus[start:end]) | ||
| else: | ||
| tokens.extend(corpus[start:]) | ||
| tokens.extend(corpus[: end - corpus_len]) |
There was a problem hiding this comment.
Fix wrap logic to always return the requested token count.
The current branch only wraps once. For larger num_tokens, it can return fewer tokens than requested, which breaks the helper’s “exact length” contract.
Suggested fix
- start = rng_to_use.randrange(corpus_len)
- end = start + num_tokens
-
- if end <= corpus_len:
- tokens.extend(corpus[start:end])
- else:
- tokens.extend(corpus[start:])
- tokens.extend(corpus[: end - corpus_len])
+ if num_tokens <= 0:
+ return tokens
+ if corpus_len == 0:
+ raise ValueError("corpus must be non-empty when num_tokens > 0")
+
+ start = rng_to_use.randrange(corpus_len)
+ tokens.extend(
+ corpus[(start + i) % corpus_len]
+ for i in range(num_tokens)
+ )🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@src/aiperf/dataset/generator/prompt.py` around lines 53 - 60, The wrap logic
may return fewer than num_tokens when num_tokens > corpus_len because it only
wraps once; change the logic in the token-selection section (around start =
rng_to_use.randrange(corpus_len), end = start + num_tokens and the
tokens.extend(...) branches) to guarantee exactly num_tokens elements by
iterating or using modulo arithmetic over corpus indices (e.g., append
corpus[(start + i) % corpus_len] for i in range(num_tokens) or loop extending
slices until len(tokens) == num_tokens) so tokens always reaches the requested
length regardless of num_tokens relative to corpus_len.
Codecov Report❌ Patch coverage is 📢 Thoughts on this report? Let us know! |
Route synthetic non-trace prompt generation through CodingContentGenerator when requested so MoE-oriented coding content applies beyond trace replay. Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]> Signed-off-by: Anthony Casagrande <[email protected]>
There was a problem hiding this comment.
Actionable comments posted: 1
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In `@tests/unit/dataset/composer/test_synthetic_composer.py`:
- Line 33: The test names don't follow the required
test_<function>_<scenario>_<expected> convention; rename the function
test_initialization_with_coding_corpus and the other test mentioned at line 46
to follow that pattern (for example
test_initialization_with_coding_corpus_success or
test_initialization_with_coding_corpus_validates_tokenization) so they
explicitly state the function under test, the scenario, and the expected
outcome; update any references/imports in the test module (e.g., the class
containing these test methods and usages of mock_tokenizer) to use the new
names.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
Run ID: dfe021d5-aa5d-473c-b2b9-5f40ebb304ad
📒 Files selected for processing (3)
src/aiperf/dataset/composer/base.pysrc/aiperf/dataset/generator/coding_content.pytests/unit/dataset/composer/test_synthetic_composer.py
| assert composer.include_image is False | ||
| assert composer.include_audio is False | ||
|
|
||
| def test_initialization_with_coding_corpus(self, mock_tokenizer): |
There was a problem hiding this comment.
Rename these tests to match the required naming contract.
Line 33 and Line 46 do not follow the test_<function>_<scenario>_<expected> convention.
Proposed rename
- def test_initialization_with_coding_corpus(self, mock_tokenizer):
+ def test_initialization_coding_corpus_uses_coding_content_generator(self, mock_tokenizer):
@@
- def test_coding_corpus_generates_context_prompts(self, mock_tokenizer):
+ def test_create_dataset_coding_corpus_populates_context_prompts(self, mock_tokenizer):As per coding guidelines: "Test naming convention: test_<function>_<scenario>_<expected>".
Also applies to: 46-46
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@tests/unit/dataset/composer/test_synthetic_composer.py` at line 33, The test
names don't follow the required test_<function>_<scenario>_<expected>
convention; rename the function test_initialization_with_coding_corpus and the
other test mentioned at line 46 to follow that pattern (for example
test_initialization_with_coding_corpus_success or
test_initialization_with_coding_corpus_validates_tokenization) so they
explicitly state the function under test, the scenario, and the expected
outcome; update any references/imports in the test module (e.g., the class
containing these test methods and usages of mock_tokenizer) to use the new
names.
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
1 participant
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.
Summary
Adds the runtime coding-text content layer that completes the agentic-coding-trace pipeline. Main already ships the session structure synthesizer (
src/aiperf/dataset/agentic_code_gen/, PRs #839/#943) — hash_ids + input/output lengths + delays in Mooncake-style JSONL — but no coding-text generator on the runtime side. Without this PR, sessions produced byagentic_code_genare executed against the default ShakespearePromptGeneratorand produce unrepresentative MoE expert-routing patterns.Why it matters (MoE, not tokenizers)
Mixture-of-Experts models route tokens to different experts based on content. Shakespeare routes to a narrow subset of experts (English-prose experts dominate). Real agentic-coding traffic activates the broader expert set hit by code, bash, JSON, error tracebacks, git diffs, configs, markdown, etc. So MoE serving metrics (per-expert load, hot/cold imbalance, throughput under realistic expert pressure) measured under Shakespeare are systematically biased away from coding-agent production. This PR replaces the corpus when an agentic-coding trace loader runs.
What's in the PR
src/aiperf/dataset/generator/coding_content.py—CodingContentGenerator(4082 lines: ~673 vocab tuples + ~1386 lines inside 81 triple-quoted code template literals + ~1.3k lines of generator scaffolding; the file is mostly corpus data, not logic).src/aiperf/common/hash_id_random_generator.py— deterministic per-hash-id RNG used by the tool-pool sampler.src/aiperf/dataset/generator/prompt.py— module-levelsample_tokens_from_corpus()helper soCodingContentGeneratorcan build its pools without sharingPromptGeneratorstate.PromptCorpusenum (sonnet/coding) inaiperf.common.enums.corpusfield onPromptConfig(drops the parent-prefix to match the existingisl/osl/block_size/batch_sizeconvention onPromptConfig).--prompt-corpus(cli_config.py:prompt_corpus) — stays prefixed to match the existingprompt_batch_size/prompt_sequence_distributionflat-CLI convention.default_prompt_corpusfield onCustom/PublicDatasetLoaderMetadata(declared asLiteral["sonnet", "coding"]inschemas.pyto dodge theaiperf.common.enums→aiperf.pluginchain circular import; mirrorsPromptCorpus.value).composer/custom.pywiring: when anis_traceloader runs and the effective corpus isCODING, swap inCodingContentGeneratorinstead of the default ShakespearePromptGenerator. Readscorpusfromself._synthetic_promptswhen set; otherwise falls back toloader_metadata.default_prompt_corpus(which defaults tosonnet, preserving existing behavior).tests/unit/dataset/generator/test_coding_content_generator.py— 76 unit tests covering init, generate(), token-sequence building, sampling, and per-language template families.What's out of scope (deliberately deferred to follow-up PRs)
composer/public.pytrace-loader wiring — main'sPublicDatasetLoaderMetadatahas nois_tracefield; the branch's_inject_trace_kwargsdepends on it. Addingis_traceto public-loader metadata is a separate, broader change.semianalysis_cc_traces_weka_no_subagentsHF loaders that would setdefault_prompt_corpus: coding. These are the consumers of this generator; they'll land alongside the Weka trace replay work.PromptGeneratorthat live on the source branch — those are Weka trace-replay concerns, not coding-content concerns.Notes
default_prompt_corpusdefaulting tosonneteverywhere and no loader onmainsettingcoding, the new code path is opt-in via--prompt-corpus codingonly. Behavior is unchanged unless the user (or a future loader plugin) requests it.tools/ergonomics_baseline.jsonwas regenerated to accept the corpus-as-data file/function sizes incoding_content.py. The 4082-line file is overwhelmingly hand-written templates and vocabulary tuples; splitting it doesn't make the data smaller and would obscure the per-template-family co-location.Test plan
uv run pytest tests/unit/ -n auto— 12749 passed, 79 skippedmake validate-plugin-schemas— 33 categories, 219 plugins cleanmake generate-all-plugin-files— schemas regenerated, no driftpre-commit run— all hooks passCodingContentGenerator(...).generate_prompt(num_tokens=64)against a realgpt2tokenizer end-to-end_gen_python_*,_gen_go_*,_gen_rust_*,_gen_ts_*,_gen_tool_*)Summary by CodeRabbit
New Features
--prompt-corpusCLI flag to choose synthetic prompt source (sonnetorcoding)Documentation
--prompt-corpusoption and loader-default behaviorSchema
Tests