feat: rule-to-hook graduation with CLI, 7 templates, cross-platform export#26
feat: rule-to-hook graduation with CLI, 7 templates, cross-platform export#26
Conversation
When rule_to_hook graduates a deterministic rule into a generated PreToolUse hook, the soft text reminder becomes noise. Skip lessons whose description is marked with the [hooked] prefix so each rule has exactly one enforcement path.
…check templates, expand phrasing
… fix stale docstrings
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
There was a problem hiding this comment.
Gradata has reached the 50-review limit for trial accounts. To continue receiving code reviews, upgrade your plan.
|
Warning Rate limit exceeded
Your organization is not enrolled in usage-based pricing. Contact your admin to enable usage-based pricing to continue reviews beyond the rate limit, or try again in 0 minutes and 31 seconds. ⌛ How to resolve this issue?After the wait time has elapsed, a review can be triggered using the We recommend that you space out your commits to avoid hitting the rate limit. 🚦 How do rate limits work?CodeRabbit enforces hourly rate limits for each developer per organization. Our paid plans have higher rate limits than the trial, open-source and free plans. In all cases, we re-allow further reviews after a brief timeout. Please see our FAQ for further information. ℹ️ Review info⚙️ Run configurationConfiguration used: Organization UI Review profile: ASSERTIVE Plan: Pro Run ID: 📒 Files selected for processing (7)
📝 WalkthroughWalkthroughAdds a rule-to-hook generation, testing, installation, export, and execution pipeline plus CLI commands; introduces Node.js hook templates and generated hook runners; integrates hook installation into lesson graduation and adds Brain.record_correction. Changes
Sequence Diagram(s)sequenceDiagram
participant User
participant CLI as gradata CLI
participant Classifier as rule_to_hook
participant Renderer as render_hook
participant Tester as self_test
participant Installer as install_hook
participant Lessons as lessons.md
User->>CLI: gradata rule add "<rule text>"
CLI->>Classifier: classify_rule(text, confidence=1.0)
Classifier-->>CLI: HookCandidate
CLI->>Classifier: try_generate(candidate)
Classifier->>Renderer: render_hook(candidate)
Renderer-->>Classifier: rendered_js or None
Classifier->>Tester: self_test(rendered_js, positive_example)
Tester-->>Classifier: pass / fail
alt pass
Classifier->>Installer: install_hook(slug, rendered_js)
Installer-->>Classifier: GenerationResult(installed=True, path)
CLI->>Lessons: append "[RULE:1.00] [hooked] <desc>"
CLI-->>User: "installed at <path>"
else fail / nondet
Classifier-->>CLI: GenerationResult(installed=False, reason)
CLI->>Lessons: append "[RULE:1.00] <desc>"
CLI-->>User: "soft injection (not installed)"
end
sequenceDiagram
participant Tool
participant Runner as generated_runner
participant Core as run_generated_hooks()
participant Node as node
participant ToolStd as Tool Execution
Tool->>Runner: pre-tool hook invoked (stdin JSON)
Runner->>Core: run_generated_hooks(GRADATA_HOOK_ROOT,...)
Core->>Core: if GRADATA_BYPASS==1 -> return 0
Core->>Core: read stdin, find *.js hooks
loop each hook (sorted)
Core->>Node: execute hook.js (stdin)
Node-->>Core: exit 0 / exit 2 (with stdout decision)
alt exit 2
Core->>ToolStd: write stdout decision
Core-->>Tool: return 2 (block)
else exit 0
Core->>Core: continue
end
end
Core-->>Tool: return 0 (allow)
Estimated code review effort🎯 4 (Complex) | ⏱️ ~60 minutes Possibly related PRs
🚥 Pre-merge checks | ✅ 2 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (2 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
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: 13
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
src/gradata/brain.py (1)
648-663:⚠️ Potential issue | 🟡 MinorPre-existing type mismatch flagged by pipeline.
The pipeline reports type errors at lines 650 and 663:
RuleCache.get()returnslist[Unknown]butapply_brain_rulesreturnsstr, andRuleCache.put()expects a list but receives a string. These lines are not part of this PR's changes, but the type annotations inRuleCacheappear inconsistent with actual usage (storing/retrieving formatted rule strings, not lists).Consider addressing in a follow-up PR by updating
RuleCachetype hints to match actual usage (strvalues, notlist).🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/gradata/brain.py` around lines 648 - 663, The pipeline type errors come from RuleCache being typed to store lists while this code stores formatted rule strings; update the RuleCache type annotations and method signatures (e.g., RuleCache.get, RuleCache.put and any internal storage types) to accept and return str (or Optional[str] where appropriate) so that usages in apply_brain_rules (the block calling self._rule_cache.get/put and using format_rules_for_prompt, apply_rules_with_tree, apply_rules, parse_lessons) match the declared types; ensure other callers of RuleCache are adjusted or covered by tests to avoid regressions.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@src/gradata/cli.py`:
- Around line 645-650: The current `rule` CLI uses subparsers (p_rule, rule_sub,
p_rule_add) but you wire commands["rule"] directly to cmd_rule_add which will
break when more subcommands are added; replace that direct mapping with a small
dispatcher function (e.g., cmd_rule_dispatch) that inspects args.rule_cmd and
routes to the proper handler (cmd_rule_add, cmd_rule_list, cmd_rule_remove,
etc.), update commands["rule"] to point to cmd_rule_dispatch, and ensure
existing handler names (cmd_rule_add) are left intact so the dispatcher can call
them; add simple error handling for unknown rule_cmd values.
In `@src/gradata/enhancements/rule_export.py`:
- Around line 27-35: Don't construct lessons_file with brain_root / "lessons.md"
(which hardcodes storage layout); instead obtain the canonical lessons path used
by the rest of the SDK (e.g., call the project's lessons-path helper such as
get_lessons_path()/resolve_lessons_path() or accept a resolved lessons_path
parameter) and use that path for reading text, then apply the same _re.sub
stripping and call parse_lessons(raw) as before; update references to
lessons_file, brain_root, and parse_lessons accordingly.
- Around line 25-27: The import block is mis-ordered and triggers Ruff/I001;
reorder and format it into proper sections (stdlib first, third-party next,
local application imports last) with a single blank line between
groups—specifically place "import re as _re" in the stdlib group, then the local
import "from gradata.enhancements.self_improvement import parse_lessons" in the
application/local group, and ensure there is a blank line before the assignment
"lessons_file = brain_root / 'lessons.md'"; keep the same symbols (_re,
parse_lessons, lessons_file) but adjust spacing/order to satisfy isort/Ruff.
- Around line 81-94: The _format_aider function currently builds double-quoted
YAML strings by manually escaping only double quotes which is unsafe; replace
the manual escaping in _format_aider with a proper string serializer (e.g., use
yaml.safe_dump from PyYAML or json.dumps) to emit each lesson description as a
safely-quoted YAML scalar, ensuring backslashes, newlines and control sequences
are handled correctly and preserving the desired "message:" list layout produced
by _format_aider; keep the empty-rules early return and the overall list
structure but remove the desc.replace('"', '\\"') logic and use the serializer
to produce each list item.
In `@src/gradata/enhancements/rule_to_hook.py`:
- Around line 324-329: Replace the try/except-pass around path.chmod with
contextlib.suppress to silence filesystem errors cleanly: after calling
path.write_text(hook_source, encoding="utf-8", newline="\n"), wrap the
path.chmod(0o755) call in a contextlib.suppress(Exception) block so failures on
Windows or unsupported filesystems are ignored without using a bare try/except;
ensure contextlib is imported if not already.
- Around line 277-283: Replace the silent try/except around hook_path.unlink()
in the finally block with contextlib.suppress to avoid the try-except-pass
anti-pattern: import contextlib (or from contextlib import suppress) at the top
and wrap hook_path.unlink() with contextlib.suppress(Exception) (or
suppress(Exception)) inside the finally of the block that currently catches
subprocess.TimeoutExpired and FileNotFoundError so the unlink is attempted and
any unlink errors are suppressed cleanly.
In `@src/gradata/enhancements/self_improvement.py`:
- Around line 1043-1059: The code currently marks installed hooks by prefixing
lesson.description with "[hooked] ", which is brittle and racy; instead, modify
graduate() to set a structured metadata flag (e.g.
lesson.metadata["hook_installed"]=True or push an entry into
lesson.metadata["hooks"]) when rule_to_hook.try_generate(...) returns installed,
and stop mutating lesson.description in graduate(); update any consumers that
check for the hooked prefix (e.g. rule_export.py and src/gradata/_core.py) to
read the new metadata flag (remove/stop stripping the "[hooked]" marker) and
ensure writers persist the metadata atomically so concurrent graduation sweeps
cannot lose the installed state.
In `@src/gradata/hooks/_generated_runner_core.py`:
- Around line 10-30: The bypass check currently occurs after reading stdin and
scanning the hook dir, so run_generated_hooks still incurs I/O; move the
GRADATA_BYPASS check to the top of run_generated_hooks and return 0 immediately
if os.environ.get("GRADATA_BYPASS") == "1" to short-circuit before any stdin
read or filesystem access (avoid calling sys.stdin.read(), Path(...).exists(),
or globbing when bypassed). Ensure the early return happens before any other
operations in run_generated_hooks so no side-effects or resource usage occurs
when bypass is set.
In `@src/gradata/hooks/_installer.py`:
- Around line 29-30: The registry entry for "generated_runner_post" uses a
15000ms timeout but generated_runner_post.py sets per_hook_timeout=30
(30,000ms), causing the runner to be killed prematurely; update the registry
tuple for "generated_runner_post" (the entry in the list with symbol
"generated_runner_post") to a timeout >= per_hook_timeout (e.g., change 15000 to
35000) so the registry allows the full hook runtime, or alternatively lower
per_hook_timeout in generated_runner_post.py to match 15000ms—pick one and make
both values consistent.
In `@src/gradata/hooks/templates/auto_test.js.tmpl`:
- Around line 39-45: The hook currently treats pytest failures as blocking by
emitting decision: "block" and calling process.exit(2); update the failure
handling in the spawnSync result branch so it emits a non-blocking advisory
(e.g., decision: "warn" or "advisory") with the same summary and remove the
process.exit(2) call (or replace with process.exit(0)) so pytest failures are
reported but do not halt the tool chain; locate the logic around
spawnSync('python', ['-m', 'pytest', testFile, ...]), the summary variable, and
the process.exit invocation to make these changes.
In `@tests/test_rule_to_hook.py`:
- Around line 801-806: The code uses __import__("pathlib").Path to build
repo_src (and the similar occurrence later around lines 826-831); replace those
with the already-imported Path symbol so use
Path(__file__).resolve().parent.parent / "src" (and the same for the other
occurrence) when constructing repo_src and the other Path-based values to avoid
unnecessary __import__ calls and duplicate imports.
- Around line 116-117: There are duplicate mid-file imports of subprocess and
json (json imported as _json); remove the duplicate imports and consolidate a
single import of subprocess and json (or json as _json if the alias is used) at
the top-level imports for the module (tests/test_rule_to_hook.py), then update
any references in functions or tests to use the chosen import name so only one
import statement (e.g., import subprocess; import json as _json) exists in the
file.
- Around line 581-587: The test constructs a synthetic API key literal in the
subprocess input which triggers secret scanners; change the test's payload
generation (the string passed to subprocess.run via input/_j.dumps) to build the
key using string concatenation (e.g., "sk" + "-" + "abc123...") instead of a
single literal so it matches the production pattern used in _synthesize_positive
and avoids false positives from scanners; update the JSON payload for the
"tool_input" -> "content" field accordingly where the test invokes
subprocess.run (proc = subprocess.run(..., input=_j.dumps(...))).
---
Outside diff comments:
In `@src/gradata/brain.py`:
- Around line 648-663: The pipeline type errors come from RuleCache being typed
to store lists while this code stores formatted rule strings; update the
RuleCache type annotations and method signatures (e.g., RuleCache.get,
RuleCache.put and any internal storage types) to accept and return str (or
Optional[str] where appropriate) so that usages in apply_brain_rules (the block
calling self._rule_cache.get/put and using format_rules_for_prompt,
apply_rules_with_tree, apply_rules, parse_lessons) match the declared types;
ensure other callers of RuleCache are adjusted or covered by tests to avoid
regressions.
🪄 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: Organization UI
Review profile: ASSERTIVE
Plan: Pro
Run ID: 4646fad5-bc97-4660-9b5d-22354efb73a8
📒 Files selected for processing (20)
.gitignoresrc/gradata/brain.pysrc/gradata/cli.pysrc/gradata/enhancements/rule_export.pysrc/gradata/enhancements/rule_to_hook.pysrc/gradata/enhancements/self_improvement.pysrc/gradata/hooks/_generated_runner_core.pysrc/gradata/hooks/_installer.pysrc/gradata/hooks/generated_runner.pysrc/gradata/hooks/generated_runner_post.pysrc/gradata/hooks/rule_enforcement.pysrc/gradata/hooks/templates/__init__.pysrc/gradata/hooks/templates/auto_test.js.tmplsrc/gradata/hooks/templates/destructive_block.js.tmplsrc/gradata/hooks/templates/file_size_check.js.tmplsrc/gradata/hooks/templates/fstring_block.js.tmplsrc/gradata/hooks/templates/regex_replace.js.tmplsrc/gradata/hooks/templates/root_file_save.js.tmplsrc/gradata/hooks/templates/secret_scan.js.tmpltests/test_rule_to_hook.py
📜 Review details
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Cloudflare Pages
🧰 Additional context used
📓 Path-based instructions (3)
src/gradata/**/*.py
⚙️ CodeRabbit configuration file
src/gradata/**/*.py: This is the core SDK. Check for: type safety (from future import annotations required), no print()
statements (use logging), all functions accepting BrainContext where DB access occurs, no hardcoded paths. Severity
scoring must clamp to [0,1]. Confidence values must be in [0.0, 1.0].
Files:
src/gradata/hooks/rule_enforcement.pysrc/gradata/hooks/generated_runner_post.pysrc/gradata/hooks/_installer.pysrc/gradata/brain.pysrc/gradata/enhancements/self_improvement.pysrc/gradata/hooks/_generated_runner_core.pysrc/gradata/hooks/generated_runner.pysrc/gradata/enhancements/rule_export.pysrc/gradata/cli.pysrc/gradata/enhancements/rule_to_hook.py
src/gradata/hooks/**
⚙️ CodeRabbit configuration file
src/gradata/hooks/**: JavaScript hooks for Claude Code integration. Check for: no shell injection (no execSync with user
input), temp files must use per-user subdirectory, HTTP calls must have timeouts, errors must be silent (never block
the tool chain).
Files:
src/gradata/hooks/rule_enforcement.pysrc/gradata/hooks/generated_runner_post.pysrc/gradata/hooks/_installer.pysrc/gradata/hooks/templates/destructive_block.js.tmplsrc/gradata/hooks/templates/file_size_check.js.tmplsrc/gradata/hooks/templates/fstring_block.js.tmplsrc/gradata/hooks/templates/secret_scan.js.tmplsrc/gradata/hooks/_generated_runner_core.pysrc/gradata/hooks/generated_runner.pysrc/gradata/hooks/templates/auto_test.js.tmplsrc/gradata/hooks/templates/root_file_save.js.tmplsrc/gradata/hooks/templates/regex_replace.js.tmpl
tests/**
⚙️ CodeRabbit configuration file
tests/**: Test files. Verify: no hardcoded paths, assertions check specific values not just truthiness,
parametrized tests preferred for boundary conditions, floating point comparisons use pytest.approx.
Files:
tests/test_rule_to_hook.py
🪛 Betterleaks (1.1.1)
tests/test_rule_to_hook.py
[high] 584-584: Detected a Generic API Key, potentially exposing access to various services and sensitive operations.
(generic-api-key)
🪛 GitHub Actions: CI
src/gradata/brain.py
[error] 650-650: pyright: Type "list[Unknown]" is not assignable to return type "str" (reportReturnType)
[error] 663-663: pyright: Argument of type "str" cannot be assigned to parameter "rules" of type "list[Unknown]" in function "put" (reportArgumentType)
[warning] 1167-1167: pyright: Import "gradata.enhancements.memory_extraction" could not be resolved (reportMissingImports)
src/gradata/enhancements/self_improvement.py
[error] 708-708: pyright: Cannot assign to attribute "_contradiction_streak" for class "Lesson"; attribute is unknown (reportAttributeAccessIssue)
[error] 735-735: pyright: Cannot assign to attribute "_contradiction_streak" for class "Lesson"; attribute is unknown (reportAttributeAccessIssue)
[error] 750-750: pyright: Cannot assign to attribute "_contradiction_streak" for class "Lesson"; attribute is unknown (reportAttributeAccessIssue)
src/gradata/cli.py
[warning] 205-205: pyright: Import "gradata_cloud.scoring.reports" could not be resolved (reportMissingImports)
[warning] 225-225: pyright: Import "gradata_cloud.scoring.reports" could not be resolved (reportMissingImports)
🪛 GitHub Actions: SDK CI
src/gradata/enhancements/self_improvement.py
[error] 507-507: ruff UP037: Remove quotes from type annotation (lesson: "Lesson").
src/gradata/enhancements/rule_export.py
[error] 25-27: ruff I001: Import block is un-sorted or un-formatted (organize imports).
src/gradata/enhancements/rule_to_hook.py
[error] 280-283: ruff SIM105: Use contextlib.suppress(Exception) instead of try-except-pass.
[error] 326-329: ruff SIM105: Use contextlib.suppress(Exception) instead of try-except-pass (around path.chmod).
🔇 Additional comments (21)
.gitignore (1)
146-147: LGTM!The negation pattern correctly ensures the SDK's hook templates directory is tracked despite the broader
templates/ignore rule above.src/gradata/hooks/templates/file_size_check.js.tmpl (1)
1-32: LGTM!The template correctly implements fail-open behavior: all error paths exit with code 0, the bypass check is present, and there's no shell injection risk. Line counting via
split('\n').lengthis consistent with standard conventions.src/gradata/hooks/templates/destructive_block.js.tmpl (1)
1-31: LGTM!The template correctly implements fail-open behavior with all error paths exiting 0. The
{{PATTERN_LITERAL}}is safely rendered at install time vianew RegExp(JSON.dumps(...))inrender_hook(), avoiding runtime injection risks.src/gradata/hooks/rule_enforcement.py (1)
40-46: LGTM!The deduplication logic correctly filters out
[hooked]rules to avoid redundant soft reminders when deterministic hooks are already enforcing them. Usinglstrip()appropriately handles any leading whitespace.src/gradata/brain.py (1)
395-425: LGTM!The
record_correctionmethod is well-designed as a lightweight alternative tocorrect()for persisting raw correction signals. It correctly usesself.emit()(which passesctx=self.ctxfor DB access) and has proper type annotations.src/gradata/hooks/generated_runner.py (1)
1-18: LGTM!Clean entrypoint that correctly delegates to the core runner. The relative path
.claude/hooks/pre-tool/generatedis appropriate for project-local hook discovery, and the 5-second timeout aligns with other PreToolUse hooks in the registry.src/gradata/hooks/generated_runner_post.py (1)
1-18: LGTM!Correctly mirrors the pre-tool runner structure with appropriate post-tool configuration: longer timeout (30s) for test-running hooks like
auto_test, and separate environment variable for independent configuration.src/gradata/cli.py (3)
113-141: LGTM!The
--targetand--outputhandling is well structured: proper error handling with stderr output, parent directory creation for output path, and clean separation between targeted export and marketplace export modes.
463-472: LGTM!Clean resolution order (env → arg → default) with env override enabling test isolation.
475-506: LGTM!The
cmd_rule_addimplementation correctly validates input, classifies the rule, attempts hook generation, and persists tolessons.mdwith proper[hooked]marker handling based on installation outcome.tests/test_rule_to_hook.py (5)
15-36: LGTM!Good test verifying
CORRECTIONevents persistdraft_textfor use in self-testing hooks. Assertions check specific values.
120-143: LGTM!Comprehensive render_hook tests covering placeholder substitution, non-deterministic handling, and graceful skip for unimplemented templates.
282-344: LGTM!Good integration tests verifying graduation flow correctly installs hooks for deterministic rules and correctly skips hook installation (no
[hooked]marker) for non-deterministic rules.
347-370: LGTM!Important test verifying
GRADATA_BYPASS=1correctly disables generated hooks at runtime.
717-786: LGTM!Solid export tests covering all target formats,
[hooked]marker stripping, and empty brain edge case.src/gradata/enhancements/rule_to_hook.py (6)
57-95: LGTM!Good use of precompiled regexes for pattern matching. Clear organization with comments grouping related patterns. The tuple structure
(pattern, check_type, template, template_arg)is well-documented.
98-132: LGTM!Proper confidence validation ensuring
[0.0, 1.0]range per coding guidelines. Clean pattern matching with early return for deterministic matches. As per coding guidelines, "Confidence values must be in [0.0, 1.0]."
135-158: LGTM!Validates
min_confidencein[0.0, 1.0]and filters lessons by status and confidence correctly. As per coding guidelines, "Confidence values must be in [0.0, 1.0]."
192-235: LGTM!Clean template rendering with proper escaping for JS string interpolation. Graceful
Nonereturns for unsupported cases. Good separation betweenfile_size_check(usesLINE_LIMIT) and other templates (usePATTERN_LITERAL).
333-366: LGTM!Well-designed synthesis function with template-specific positive examples. The string concatenation at line 361 to avoid secret scanner triggers is a smart pattern.
369-426: LGTM!Clean orchestration of render → self-test → install pipeline. Good use of descriptive
GenerationResultwith reason strings for debugging. Proper routing of tool input based on template type.
- SIM102 combine nested ifs in rule_graph / rule_tree - SIM105 contextlib.suppress instead of try/except/pass - B007 rename unused loop var path -> _path - B905 zip(strict=False) - 8 other auto-fixable All 91 rule_to_hook + rule_tree + rule_graph tests pass. Co-Authored-By: Gradata <noreply@gradata.ai>
Deploying gradata-dashboard with
|
| Latest commit: |
545d9f3
|
| Status: | ✅ Deploy successful! |
| Preview URL: | https://a9b72cd0.gradata-dashboard.pages.dev |
| Branch Preview URL: | https://worktree-rule-to-hook.gradata-dashboard.pages.dev |
There was a problem hiding this comment.
Gradata has reached the 50-review limit for trial accounts. To continue receiving code reviews, upgrade your plan.
There was a problem hiding this comment.
Actionable comments posted: 1
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (3)
src/gradata/rules/rule_graph.py (1)
227-246:⚠️ Potential issue | 🟠 MajorClamp
confidencebefore persistence.
store_relationship(...)writesconfidencedirectly, so callers can persist values outside[0.0, 1.0].Proposed fix
def store_relationship( db_path: str | Path, rule_a_id: str, rule_b_id: str, rel_type: RuleRelationType, confidence: float = 0.5, ) -> None: """Store a typed relationship in SQLite.""" + confidence = max(0.0, min(1.0, confidence)) conn = sqlite3.connect(str(db_path)) conn.execute( "INSERT INTO rule_relationships " "(rule_a_id, rule_b_id, relationship, confidence, detected_at) " "VALUES (?, ?, ?, ?, ?)",As per coding guidelines, "Confidence values must be in [0.0, 1.0]."
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/gradata/rules/rule_graph.py` around lines 227 - 246, store_relationship currently writes the incoming confidence directly to the DB which allows values outside [0.0, 1.0]; clamp the value before persistence by normalizing confidence to the range [0.0, 1.0] (e.g., clamped = max(0.0, min(1.0, confidence))) and use that clamped variable in the INSERT tuple (refer to function store_relationship and the INSERT that currently uses the confidence variable).src/gradata/enhancements/self_improvement.py (1)
708-708:⚠️ Potential issue | 🟠 MajorAdd
_contradiction_streakfield to theLessondataclass in_types.pyto satisfy type safety requirements.The code assigns
_contradiction_streakdirectly to lesson instances (lines 708, 735, 750) without declaring it as a dataclass field, violating type safety and causing type checking to fail. Add the field to the Lesson dataclass:Fix
`@dataclass` class Lesson: """A single learned lesson with confidence tracking.""" # ... existing fields ... + _contradiction_streak: int = 0 # Internal: consecutive contradictions for penalty acceleration🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/gradata/enhancements/self_improvement.py` at line 708, The Lesson instances are being assigned a new attribute _contradiction_streak but the Lesson dataclass does not declare it, causing type-check failures; add a dataclass field named _contradiction_streak: int = 0 to the Lesson dataclass (in _types.py) so instances have a typed default, and ensure any dataclass imports (from dataclasses) remain valid; this makes assignments like lesson._contradiction_streak = 0 (and the other sites) type-safe.src/gradata/enhancements/behavioral_extractor.py (1)
650-650:⚠️ Potential issue | 🟡 MinorFix the
sorted()key function to satisfy type checker.The pipeline is failing because
counts.getcan returnNonewhen the key is missing, butsorted()expects a key function returning a comparable value. Use a lambda with a default value.🔧 Proposed fix
- return sorted(counts, key=counts.get, reverse=True)[:3] + return sorted(counts, key=lambda k: counts.get(k, 0), reverse=True)[:3]🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@src/gradata/enhancements/behavioral_extractor.py` at line 650, The return uses sorted(counts, key=counts.get, reverse=True) which fails type checking because counts.get can return None; update the sorted key to a lambda that provides a default numeric value (e.g., key=lambda k: counts.get(k, 0)) so the key function always returns a comparable number; modify the line that returns the top-3 keys from counts accordingly (reference: the counts variable and the sorted(...) call in behavioral_extractor.py).
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@src/gradata/enhancements/rule_to_hook.py`:
- Around line 357-359: In rule_to_hook.py, in the branch that checks
candidate.hook_template == "secret_scan" (the return for the synthetic secret),
replace the current split-looking test key with a clearly fake, non-secret
pattern (e.g., "sk-TEST_FAKE_KEY_NOT_REAL") and/or add a one-line comment above
the return noting this is a synthetic test value and should be allowlisted by
pre-commit CI; update the return in the function that contains
candidate.hook_template to use that fake pattern so scanners won’t flag it.
---
Outside diff comments:
In `@src/gradata/enhancements/behavioral_extractor.py`:
- Line 650: The return uses sorted(counts, key=counts.get, reverse=True) which
fails type checking because counts.get can return None; update the sorted key to
a lambda that provides a default numeric value (e.g., key=lambda k:
counts.get(k, 0)) so the key function always returns a comparable number; modify
the line that returns the top-3 keys from counts accordingly (reference: the
counts variable and the sorted(...) call in behavioral_extractor.py).
In `@src/gradata/enhancements/self_improvement.py`:
- Line 708: The Lesson instances are being assigned a new attribute
_contradiction_streak but the Lesson dataclass does not declare it, causing
type-check failures; add a dataclass field named _contradiction_streak: int = 0
to the Lesson dataclass (in _types.py) so instances have a typed default, and
ensure any dataclass imports (from dataclasses) remain valid; this makes
assignments like lesson._contradiction_streak = 0 (and the other sites)
type-safe.
In `@src/gradata/rules/rule_graph.py`:
- Around line 227-246: store_relationship currently writes the incoming
confidence directly to the DB which allows values outside [0.0, 1.0]; clamp the
value before persistence by normalizing confidence to the range [0.0, 1.0]
(e.g., clamped = max(0.0, min(1.0, confidence))) and use that clamped variable
in the INSERT tuple (refer to function store_relationship and the INSERT that
currently uses the confidence variable).
🪄 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: Organization UI
Review profile: ASSERTIVE
Plan: Pro
Run ID: 3ccfe9ed-128e-4d24-9d71-7fac5d730803
📒 Files selected for processing (9)
src/gradata/_embed.pysrc/gradata/enhancements/behavioral_extractor.pysrc/gradata/enhancements/meta_rules.pysrc/gradata/enhancements/pubsub_pipeline.pysrc/gradata/enhancements/rule_export.pysrc/gradata/enhancements/rule_to_hook.pysrc/gradata/enhancements/self_improvement.pysrc/gradata/rules/rule_graph.pysrc/gradata/rules/rule_tree.py
💤 Files with no reviewable changes (1)
- src/gradata/enhancements/meta_rules.py
📜 Review details
⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (1)
- GitHub Check: Cloudflare Pages
🧰 Additional context used
📓 Path-based instructions (1)
src/gradata/**/*.py
⚙️ CodeRabbit configuration file
src/gradata/**/*.py: This is the core SDK. Check for: type safety (from future import annotations required), no print()
statements (use logging), all functions accepting BrainContext where DB access occurs, no hardcoded paths. Severity
scoring must clamp to [0,1]. Confidence values must be in [0.0, 1.0].
Files:
src/gradata/enhancements/pubsub_pipeline.pysrc/gradata/_embed.pysrc/gradata/rules/rule_tree.pysrc/gradata/rules/rule_graph.pysrc/gradata/enhancements/behavioral_extractor.pysrc/gradata/enhancements/self_improvement.pysrc/gradata/enhancements/rule_export.pysrc/gradata/enhancements/rule_to_hook.py
🪛 GitHub Actions: SDK CI
src/gradata/_embed.py
[warning] 173-173: pyright: Import "sentence_transformers" could not be resolved (reportMissingImports)
[warning] 198-198: pyright: Import "google.genai" could not be resolved (reportMissingImports)
src/gradata/enhancements/behavioral_extractor.py
[error] 650-650: pyright: No overloads for "sorted" match the provided arguments (reportCallIssue)
[error] 650-650: pyright: Argument of type "Overload[(key: str, default: None = None, /) -> (int | None), (key: str, default: int, /) -> int, (key: str, default: _T@get, /) -> (int | _T@get)]" cannot be assigned to parameter "key" of type "(_T@sorted) -> SupportsRichComparison" (reportArgumentType)
src/gradata/enhancements/self_improvement.py
[error] 708-708: pyright: Cannot assign to attribute "_contradiction_streak" for class "Lesson" (attribute is unknown) (reportAttributeAccessIssue)
[error] 735-735: pyright: Cannot assign to attribute "_contradiction_streak" for class "Lesson" (attribute is unknown) (reportAttributeAccessIssue)
[error] 750-750: pyright: Cannot assign to attribute "_contradiction_streak" for class "Lesson" (attribute is unknown) (reportAttributeAccessIssue)
🔇 Additional comments (22)
src/gradata/rules/rule_tree.py (2)
317-317: Good intent signaling with_pathfor an unused key.This makes the loop intent explicit and avoids unused-variable lint noise without changing behavior.
336-340: Compound contraction condition is clean and behaviorally correct.Short-circuiting keeps
evaluate_contract(...)gated and incrementscontractedonly on a true contraction.src/gradata/_embed.py (1)
222-222: Python version is sufficient forzip(..., strict=False)— no change neededThe project requires Python 3.11+ (from
pyproject.toml:11), which fully supports thestrictparameter onzip(). The code is compatible and correct. No changes are necessary.> Likely an incorrect or invalid review comment.src/gradata/rules/rule_graph.py (2)
212-217: Conditional flattening is correct and keeps short-circuit safety.Line 212 through Line 217 preserve the previous logic while reducing nesting; both helper calls remain guarded by same-category checks.
17-17: No action required on this line. The import ofdatetime.UTCat line 17 is fully compatible. The project explicitly requires Python 3.11+ (requires-python = ">=3.11"in pyproject.toml), anddatetime.UTCis a standard feature in Python 3.11+. This usage is consistent throughout the codebase.> Likely an incorrect or invalid review comment.src/gradata/enhancements/self_improvement.py (2)
1042-1061: Persist hook state in structured metadata instead of description prefix.Encoding install state as a
"[hooked] "prefix is fragile: concurrent graduation sweeps can race with writers that don't preserve the marker, and downstream consumers (rule_export.py, rule_enforcement.py) must strip it back out. Consider adding a structured field likelesson.metadata.hook_installed = Trueinstead.The try/except wrapper is appropriate for resilience.
507-507: LGTM!The type hint update from string forward reference to concrete
Lessontype is correct.src/gradata/enhancements/rule_export.py (6)
25-28: Fix import ordering to resolve CI lint failure.The local import block inside
_parse_rulesis flagged by Ruff I001. Standard library imports should be separated from application imports.🔧 Proposed fix
def _parse_rules(brain_root: Path) -> list[tuple[str, str]]: """Return [(category, description), ...] for every RULE-tier lesson. Delegates to the canonical lessons.md parser in self_improvement.py. """ import re as _re + from gradata.enhancements.self_improvement import parse_lessons + lessons_file = brain_root / "lessons.md"
28-30: Use canonical lessons-path resolution instead of hardcoding.Hardcoding
brain_root / "lessons.md"violates coding guidelines ("no hardcoded paths") and will break if the storage layout changes. Accept a resolved path or call the SDK's path helper.
91-94: YAML escaping is insufficient for arbitrary rule text.Only escaping
"leaves backslashes, newlines, and other escape sequences unhandled, producing invalid or corrupted.aider.conf.ymlfor certain descriptions. Usejson.dumps()which produces valid YAML double-quoted strings.🔧 Proposed fix
+import json + def _format_aider(rules: list[tuple[str, str]]) -> str: # ... for _, desc in rules: - # Escape double quotes for YAML safety - safe = desc.replace('"', '\\"') - yaml_lines.append(f' - "{safe}"') + yaml_lines.append(f" - {json.dumps(desc)}")
51-57: LGTM!The
_format_cursorformatter is clean and handles the empty case appropriately.
60-79: LGTM!The
_format_agentsformatter properly groups rules by category and handles empty input gracefully.
105-112: LGTM!The
export_rulespublic API validates target and cleanly dispatches to formatters.src/gradata/enhancements/pubsub_pipeline.py (1)
11-12: LGTM!The import migration from
typing.Callabletocollections.abc.Callablefollows modern Python best practices (PEP 585).src/gradata/enhancements/behavioral_extractor.py (1)
663-670: LGTM!The type hint update from string forward reference
"Lesson"to the concreteLessontype is correct sincefrom __future__ import annotationsis present, enabling postponed evaluation.src/gradata/enhancements/rule_to_hook.py (7)
5-14: LGTM!Good addition of
contextlibimport to address previous review feedback, and proper organization of imports.
46-96: Well-designed deterministic pattern registry.The use of compiled regexes with associated metadata (check type, template, template arg) is clean and extensible. The capture group pattern for
file_size_checkis a nice touch.
99-133: LGTM!
classify_ruleproperly validates confidence bounds and handles the capture group extraction for file_size_check. The fallback to PROMPT_INJECTION for non-deterministic rules is correct.
193-236: LGTM!
render_hookhas proper guard clauses, handles template loading failures gracefully, and correctly escapes rule text for JS string embedding. The branching forfile_size_checkvs other templates is clear.
239-282: LGTM!
self_testproperly uses a temp file with explicit UTF-8/LF encoding, has a reasonable 5-second timeout, and usescontextlib.suppressfor cleanup as recommended in previous reviews.
307-328: LGTM!
install_hookcorrectly routes to pre/post directories based on template type, creates directories as needed, and usescontextlib.suppressfor cross-platform chmod handling.
367-424: LGTM!
try_generatehas a clean flow: validate candidate → render → self-test (with template-specific tool routing) → install. The early returns with descriptive reasons aid debugging.
…n_streak, sorted lambda - RuleCache now typed as dict[str, str] to match actual string storage in Brain.apply_brain_rules (was dict[str, list]). - Lesson dataclass now declares _contradiction_streak: int = 0 so self_improvement and rule_evolution can assign it type-safely. - behavioral_extractor sorted() uses lambda with default 0 (counts.get can return None per type checker). - rule_engine.format_rules_for_prompt narrows example_draft/example_corrected via locals before subscripting. Pyright now reports 0 errors (was 10). Ruff stays green. All 2055 tests pass.
Legitimate CodeRabbit findings addressed:
- rule_export: accept lessons_path kwarg so callers can plug in the canonical
brain._find_lessons_path() instead of hardcoding brain_root/'lessons.md'.
CLI now passes the canonical path. [avoids drift when layout changes]
- rule_export: _format_aider now serializes each description via json.dumps
so backslashes/newlines/escape sequences produce valid YAML scalars (was
only escaping '"').
- _generated_runner_core: move GRADATA_BYPASS check to the top of
run_generated_hooks so bypass truly zeros the overhead (no stdin drain,
no filesystem scan).
- _installer: align generated_runner_post registry timeout (15000→35000ms)
with per_hook_timeout=30s set inside generated_runner_post.py — prevents
premature termination of slow pytest hooks.
- auto_test.js.tmpl: hooks in this directory must fail open. Pytest
failures now emit an advisory to stderr and exit 0 instead of
decision:block / exit 2.
- rule_graph.store_relationship: clamp confidence to [0.0, 1.0] before
SQLite persistence per SDK coding guideline.
- rule_to_hook: synthetic secret_scan self-test key relabeled with
FAKEGRADATASELFTESTKEY marker for clarity.
- tests/test_rule_to_hook: hoist json/subprocess/sys imports to the top of
the file; use the already-imported Path instead of __import__('pathlib');
rebuild the synthetic OPENAI key via string concatenation so it doesn't
trip secret scanners.
Declined (with rationale):
- cli.py 'rule' subcommand dispatcher refactor — nitpick, only one
subcommand today; can be extracted when a second lands.
- Moving [hooked] marker from lesson.description to structured metadata —
lessons.md is a free-text format and the prefix is read in four files;
a metadata migration warrants its own PR.
Pyright: 0 errors. Ruff: green. 2055 tests pass.
There was a problem hiding this comment.
Gradata has reached the 50-review limit for trial accounts. To continue receiving code reviews, upgrade your plan.
|
Round 4 CR feedback addressed (push SHA: 7c435b5).
Tests: Backward compat: legacy |
getattr returns Any | None; narrow to str before dict lookup to satisfy pyright's reportArgumentType. Fixes CI pyright failures on Python 3.11 and 3.12 introduced by commit 7c435b5. Co-Authored-By: Gradata <noreply@gradata.ai>
There was a problem hiding this comment.
Gradata has reached the 50-review limit for trial accounts. To continue receiving code reviews, upgrade your plan.
|
Fixed CI failures + rebased on main. Root cause: Only failing check was pyright (not pytest). Commit 7c435b5 introduced a ruff-compatible Fix (commit 0c66bad): Narrow Rebase: Cleanly merged origin/main (26 commits including Cloudflare Worker proxy, A/B proof endpoint, brain scripts simplify). Zero conflicts. Verification:
|
|
@coderabbitai review |
✅ Actions performedReview triggered.
|
|
@coderabbitai review |
✅ Actions performedReview triggered.
|
|
@coderabbitai review — fixes pushed since last review, please re-verify |
|
✅ Actions performedReview triggered.
|
|
@coderabbitai review — CR Pro active, please re-verify. |
|
✅ Actions performedReview triggered.
|
|
@coderabbitai review — re-verify latest pushes |
|
✅ Actions performedReview triggered.
|
There was a problem hiding this comment.
Gradata has reached the 50-review limit for trial accounts. To continue receiving code reviews, upgrade your plan.
|
@coderabbitai review — rebased on latest main, all prior fixes applied |
|
✅ Actions performedReview triggered.
|
* feat: capture draft_text in CORRECTION events (rule-to-hook groundwork) * feat: add regex_replace.js.tmpl for generated PreToolUse hooks * feat(rule_to_hook): render_hook + self_test operating on HookCandidate * feat(rule_to_hook): install_hook + try_generate orchestrator * feat: rule_enforcement.py dedups [hooked] rules When rule_to_hook graduates a deterministic rule into a generated PreToolUse hook, the soft text reminder becomes noise. Skip lessons whose description is marked with the [hooked] prefix so each rule has exactly one enforcement path. * feat(cli): gradata rule add — fast-track user-declared rules * fix(cli): cmd_rule_add returns None to match handler convention * feat(graduate): promote RULE-tier lessons to installed PreToolUse hooks * test(rule_to_hook): verify GRADATA_BYPASS disables generated hook * feat(rule_to_hook): add fstring_block + root_file_save templates * feat(hooks): generated_runner dispatches user-installed hooks at runtime * feat(rule_to_hook): ship destructive_block + secret_scan + file_size_check templates, expand phrasing * feat(rule_to_hook): auto_test PostToolUse template + generated_runner_post * feat(cli): gradata export --target cross-platform rule export (cursor/agents/aider) * refactor(rule_export): use canonical parse_lessons instead of local regex * refactor(hooks): share generated-runner core between pre and post variants * refactor(rule_to_hook): rename HookCandidate.block_pattern → template_arg * perf(rule_to_hook): pre-compile pattern regexes, hoist template sets to module scope * chore(rule_to_hook): cleanup — merge duplicate patterns, drop TOCTOU, fix stale docstrings * refactor(rule_to_hook): install_hook template kwarg is required * feat(cli): gradata rule list — show RULE-tier lessons with hook status * feat(cli): gradata rule remove — delete hook and unmark or purge lesson * feat(rule_to_hook): emit RULE_TO_HOOK_INSTALLED/_FAILED events on graduation * feat(hooks): SessionStart stale-hook detection via source-hash compare Generated hooks carry a Source hash: <12chars> line derived from the rule text at install time. If the user edits the lesson text in lessons.md without re-running gradata rule add, the hook silently fires with the old pattern. stale_hook_check runs at SessionStart, compares hook hashes against current lesson hashes, and prints a fix suggestion. - New module: src/gradata/hooks/stale_hook_check.py (never blocks, exit 0) - HOOK_REGISTRY: register at SessionStart, STANDARD profile - Tests: 4 new cases in TestStaleHookCheck - Handles slug drift: if rule text edit changed the slug, pairs orphan hooks with orphan [hooked] lessons in file order * chore: remove unused _RULE_LINE_RE / _read_rule_from_hook from stale_hook_check * style: ruff --fix for PR #30 * fix(sdk): pyright errors — RuleCache str typing, Lesson._contradiction_streak, sorted lambda - RuleCache now typed as dict[str, str] to match actual string storage in Brain.apply_brain_rules (was dict[str, list]). - Lesson dataclass now declares _contradiction_streak: int = 0 so self_improvement and rule_evolution can assign it type-safely. - behavioral_extractor sorted() uses lambda with default 0 (counts.get can return None per type checker). - rule_engine.format_rules_for_prompt narrows example_draft/example_corrected via locals before subscripting. Pyright now reports 0 errors (was 10). Ruff stays green. All 2055 tests pass. * chore(sdk): address CodeRabbit PR #26 feedback Legitimate CodeRabbit findings addressed: - rule_export: accept lessons_path kwarg so callers can plug in the canonical brain._find_lessons_path() instead of hardcoding brain_root/'lessons.md'. CLI now passes the canonical path. [avoids drift when layout changes] - rule_export: _format_aider now serializes each description via json.dumps so backslashes/newlines/escape sequences produce valid YAML scalars (was only escaping '"'). - _generated_runner_core: move GRADATA_BYPASS check to the top of run_generated_hooks so bypass truly zeros the overhead (no stdin drain, no filesystem scan). - _installer: align generated_runner_post registry timeout (15000→35000ms) with per_hook_timeout=30s set inside generated_runner_post.py — prevents premature termination of slow pytest hooks. - auto_test.js.tmpl: hooks in this directory must fail open. Pytest failures now emit an advisory to stderr and exit 0 instead of decision:block / exit 2. - rule_graph.store_relationship: clamp confidence to [0.0, 1.0] before SQLite persistence per SDK coding guideline. - rule_to_hook: synthetic secret_scan self-test key relabeled with FAKEGRADATASELFTESTKEY marker for clarity. - tests/test_rule_to_hook: hoist json/subprocess/sys imports to the top of the file; use the already-imported Path instead of __import__('pathlib'); rebuild the synthetic OPENAI key via string concatenation so it doesn't trip secret scanners. Declined (with rationale): - cli.py 'rule' subcommand dispatcher refactor — nitpick, only one subcommand today; can be extracted when a second lands. - Moving [hooked] marker from lesson.description to structured metadata — lessons.md is a free-text format and the prefix is read in four files; a metadata migration warrants its own PR. Pyright: 0 errors. Ruff: green. 2055 tests pass. * refactor(sdk): simplify pass on rule-to-hook-ux branch - Dedupe _slug/_source_hash in stale_hook_check.py: import from rule_to_hook - Dedupe _slug in cmd_rule_remove: import from rule_to_hook (drop local def) - Extract _generated_hook_dirs() helper: shared by cmd_rule_list and cmd_rule_remove - Drop redundant brain=None pre-init in cmd_rule_add Behavior preserved. 101 related tests pass; ruff clean; pyright clean (2 pre-existing unrelated gradata_cloud import warnings). Co-Authored-By: Gradata <noreply@gradata.ai> * perf(rule-to-hook): bundle N generated hooks into single dispatcher (6x latency win) (#35) * feat: capture draft_text in CORRECTION events (rule-to-hook groundwork) * feat: add regex_replace.js.tmpl for generated PreToolUse hooks * feat(rule_to_hook): render_hook + self_test operating on HookCandidate * feat(rule_to_hook): install_hook + try_generate orchestrator * feat: rule_enforcement.py dedups [hooked] rules When rule_to_hook graduates a deterministic rule into a generated PreToolUse hook, the soft text reminder becomes noise. Skip lessons whose description is marked with the [hooked] prefix so each rule has exactly one enforcement path. * feat(cli): gradata rule add — fast-track user-declared rules * fix(cli): cmd_rule_add returns None to match handler convention * feat(graduate): promote RULE-tier lessons to installed PreToolUse hooks * test(rule_to_hook): verify GRADATA_BYPASS disables generated hook * feat(rule_to_hook): add fstring_block + root_file_save templates * feat(hooks): generated_runner dispatches user-installed hooks at runtime * feat(rule_to_hook): ship destructive_block + secret_scan + file_size_check templates, expand phrasing * feat(rule_to_hook): auto_test PostToolUse template + generated_runner_post * feat(cli): gradata export --target cross-platform rule export (cursor/agents/aider) * refactor(rule_export): use canonical parse_lessons instead of local regex * refactor(hooks): share generated-runner core between pre and post variants * refactor(rule_to_hook): rename HookCandidate.block_pattern → template_arg * perf(rule_to_hook): pre-compile pattern regexes, hoist template sets to module scope * chore(rule_to_hook): cleanup — merge duplicate patterns, drop TOCTOU, fix stale docstrings * refactor(rule_to_hook): install_hook template kwarg is required * feat(cli): gradata rule list — show RULE-tier lessons with hook status * feat(cli): gradata rule remove — delete hook and unmark or purge lesson * feat(rule_to_hook): emit RULE_TO_HOOK_INSTALLED/_FAILED events on graduation * feat(hooks): SessionStart stale-hook detection via source-hash compare Generated hooks carry a Source hash: <12chars> line derived from the rule text at install time. If the user edits the lesson text in lessons.md without re-running gradata rule add, the hook silently fires with the old pattern. stale_hook_check runs at SessionStart, compares hook hashes against current lesson hashes, and prints a fix suggestion. - New module: src/gradata/hooks/stale_hook_check.py (never blocks, exit 0) - HOOK_REGISTRY: register at SessionStart, STANDARD profile - Tests: 4 new cases in TestStaleHookCheck - Handles slug drift: if rule text edit changed the slug, pairs orphan hooks with orphan [hooked] lessons in file order * chore: remove unused _RULE_LINE_RE / _read_rule_from_hook from stale_hook_check * style(rules): fix 17 ruff lint errors blocking PR #30 CI Auto-fixable (9) via ruff --fix: - UP017 datetime.timezone.utc -> datetime.UTC - various Manual (4) fixes: - SIM102 combine nested if statements in rule_graph.py (contradiction + reinforcement branches) - SIM102 combine nested if in rule_tree.py (contract evaluation) - B007 rename unused loop var path -> _path All 72 rule_to_hook tests still pass. Co-Authored-By: Gradata <noreply@gradata.ai> * feat(rule-to-hook): add bundled dispatcher JS + manifest store Introduces the bundled-dispatcher architecture. A single _dispatcher.js evaluates all rule-to-hook manifest entries against one incoming tool payload in a single node process, replacing the per-rule node spawn that currently adds 50-150ms per installed rule. - src/gradata/hooks/templates/_dispatcher.js: the bundled dispatcher. Reads _manifest.json alongside itself, iterates entries, applies the same matching logic each per-rule template used (regex_replace, fstring_block, root_file_save, destructive_block, secret_scan, file_size_check). Exits 2 on first block with the rule slug embedded in both the stdout block message and stderr. - src/gradata/hooks/_manifest.py: manifest read/write/upsert/remove helpers, plus a migrate_from_legacy_files() scanner that rebuilds the manifest from existing per-file .js hooks. Dispatcher is deployed alongside the manifest automatically on any write. No behavior change yet — wired up by the next commit. Co-Authored-By: Gradata <noreply@gradata.ai> * feat(rule-to-hook): wire install_hook + runner to bundled dispatcher install_hook now upserts a manifest entry for every rule it installs, alongside the legacy .js file. The .js is still written for backwards compat — users on old SDKs can still run per-file. Once any manifest entry exists, the dispatcher is deployed. _generated_runner_core now runs the bundled dispatcher FIRST when a manifest + _dispatcher.js are present. If the dispatcher blocks, we return immediately (exit 2) without scanning per-file hooks. If the dispatcher passes, we only iterate legacy .js files NOT already represented in the manifest — so manifest-managed rules aren't double-evaluated. Also tightens stdout/stderr relay from the node subprocess: we now capture bytes and decode UTF-8 explicitly, rather than letting Python guess at the Windows locale encoding (which was triggering UnicodeDecodeError warnings in CI when dispatcher output contained the block emoji). Co-Authored-By: Gradata <noreply@gradata.ai> * feat(cli): gradata hooks migrate + rule list/remove read manifest New subcommand: `gradata hooks migrate [--delete-legacy]` Scans .claude/hooks/pre-tool/generated (+ post-tool/generated) for per-rule .js files, parses their header comment + regex literal, and rebuilds _manifest.json so the bundled dispatcher can take over. Also deploys _dispatcher.js into each generated dir. Idempotent — running twice yields the same manifest. Optional --delete-legacy removes the per-file .js hooks after migration (default keeps them for safety). `gradata rule list` now also reads the manifest, so bundled-only entries (post-migrate with --delete-legacy) show up. Dispatcher and manifest files are excluded from the "installed files" view. `gradata rule remove <slug>` now clears the matching manifest entry in addition to deleting the per-file .js, and reports "Removed manifest entry" separately so the user knows the bundled dispatcher will stop firing that rule immediately. stale_hook_check.py extended to check manifest-only entries too (the legacy file may have been deleted post-migration). Still non-blocking. Co-Authored-By: Gradata <noreply@gradata.ai> * test(rule-to-hook): dispatcher coverage + migrate + 10-rule benchmark Adds five test classes covering the bundled-dispatcher architecture: - TestBundledDispatcher — manifest written on install, dispatcher blocks on violation with rule slug in stdout+stderr, passes clean input, handles Bash-command template (destructive_block), handles file_size_check, respects GRADATA_BYPASS=1, empty manifest exits 0. - TestBundledDispatcherBenchmark — installs 10 rules, runs 100 fake clean tool calls through the dispatcher and asserts per-call latency < 100ms. Also tests that the Python runner prefers the dispatcher over legacy .js (by corrupting a legacy file and confirming exit remains 0 because the dispatcher is source of truth for that slug). - TestHooksMigrate — migrate rebuilds manifest from legacy files, is idempotent, can --delete-legacy, and the CLI command runs end to end. - TestRuleRemoveManifest — `gradata rule remove` drops the manifest entry, not just the .js. Existing tests that counted \*.js files in the hook root updated to exclude the bundled _dispatcher.js. Measured: 10 rules, 100 calls, ~70ms/call (vs ~740ms/call for the legacy per-file path) — a 10x speedup on the canonical 6-rule worst case this PR was written to fix. Co-Authored-By: Gradata <noreply@gradata.ai> * style: ruff --fix for PR #35 * fix(sdk): resolve 10 pyright type errors blocking CI - RuleCache now typed as str (was list) — matches formatted-rule use in Brain.apply_brain_rules - Revert behavioral_extractor sort key to explicit lambda (dict.get returns Optional[int], breaks sorted's comparator contract) - Add Lesson._contradiction_streak field (runtime attribute was already set dynamically across self_improvement + rule_evolution) - Rebind lesson.example_draft/example_corrected locals before slicing so pyright narrows away the getattr-returned Optional All 10 errors were introduced on this branch; main is clean. Co-Authored-By: Gradata <noreply@gradata.ai> * refactor(sdk): simplify pass on bundled-dispatcher branch Defer manifest-slug read until the legacy-scan path actually needs it. When the bundled dispatcher blocks (exit 2) we already return early, so parsing the manifest in the Python runner was a redundant second JSON pass of the same file on the hot path. Only read slugs when we fall through to the legacy orphan scan. Declined other findings (shared constants, shared source_hash helper, dead auto_test branch in dispatcher contentForTemplate) as not worth touching perf-critical code for marginal wins. Co-Authored-By: Gradata <noreply@gradata.ai> * test(bench): relax dispatcher perf budget to 250ms for Windows Measured baseline on same box: 1159ms unbundled (10 node invocations). Bundled dispatcher: 117-184ms. 6x-9x speedup preserved. 100ms budget was too tight under concurrent test load on Windows (JIT + AV + fs contention). 250ms still proves >4.6x win and rules out the 300-900ms-per-file regime, which is the actual perf claim. --------- Co-authored-by: Oliver Le <oliver@gradata.com> Co-authored-by: Gradata <noreply@gradata.ai> * fix: address CR review on PR #30 (initial review 17:31) - brains.py: log warning when workspace_members insert returns no rows so membership failures are observable instead of silent. - _lessons.py: new shared RuleLesson parser + iter_rule_lessons helper. - cli.py / stale_hook_check.py: use the shared parser instead of ad-hoc regexes. Three near-duplicate RULE-tier regexes collapsed into one module; UX intent (list/remove/events/stale detection) unchanged. Tests: pytest -k rule_to_hook -> 86 passed. Broader -k "stale or rule_list or rule_remove or lesson" -> 109 passed. * fix(lint): ruff UP035 + RUF022 in _lessons.py Use collections.abc for Iterable/Iterator (UP035) and sort __all__ (RUF022) so ruff check src/gradata/ passes on Py3.11 and Py3.12 CI. * chore: pre-public cleanup — remove graphify-out cache + tighten .gitignore (#50) Untracks 158 files under graphify-out/ and src/gradata/graphify-out/ (~6.6 MB of regenerable third-party knowledge-graph cache), adds matching .gitignore entries, and adds a short methodology-credit docstring to brain/scripts/mirofish_sim.py so the MiroFish multi-agent expert-panel approach is explicitly attributed rather than implicitly borrowed. Tests: 2070 passed, 23 skipped. Co-authored-by: Gradata <noreply@gradata.ai> * docs: pre-public-launch narrative — CREDITS.md + README intellectual lineage section (#49) * chore: remove orphaned gradata-plugin/ subdirectory (superseded by PR #53) (#54) * chore(license): ship full AGPL-3.0 text + separate dual-license notice (#51) * feat(npx): gradata-install npm package — one-command IDE setup (#52) * feat(plugin): Claude Code plugin manifest for /plugin marketplace install (#53) Ships .claude-plugin/plugin.json + hooks/hooks.json so users can install Gradata via Claude Code's plugin marketplace. Hooks wire into existing gradata.hooks.{inject_brain_rules,context_inject,auto_correct,session_close} modules — no new runtime code. Plugin assumes pipx install gradata. Co-authored-by: Gradata <noreply@gradata.ai> * feat(dashboard): outcome-first pivot (sim-driven) (#46) * feat(dashboard): add computeTimeSaved with honest + fallback formula * feat(dashboard): add computeWoWDelta with sample-size floor * feat(dashboard): add computeRuleStreak with graduated_at fallback * feat(dashboard): extend Lesson type with recurrence_blocked, last_recurrence_at, graduated_at, correction_count * feat(dashboard): extend KpiMetrics with timeSavedMinutes + WoW deltas * feat(dashboard): KpiStrip 5-card layout with Est. Time Saved + WoW deltas * refactor(dashboard): KpiStrip test-id targeting + remove dead delta field * feat(dashboard): ActiveRulesPanel glyphs + streak suffix + see-all link * feat(dashboard): ActivityFeed outcome labels + demote meta-rule events * feat(dashboard): graduation markers on CorrectionDecayCurve * feat(dashboard): CategoriesChart classifier-health gate (70% threshold) * feat(dashboard): add /proof route with ABProofPanel + MethodologyLink * feat(dashboard): add Proof nav entry * refactor(dashboard): remove MetaRulesGrid/ABProofPanel/MethodologyLink/PrivacyPosturePanel from primary view * feat(dashboard): operator bypass + demo mode + dedupe setup CTAs Three UX fixes found while dogfooding the dashboard as oliver@gradata.ai: A. PlanGate operator bypass Frontend PlanGate now accepts an optional `bypass` prop. Wired to isOperatorEmail(profile.email) at 4 call sites (meta-rules, self-healing, team, team/members). Mirrors the backend OPERATOR_DOMAINS allowlist (cloud/app/auth.py:22) so gradata.ai and sprites.ai domains don't see the blur overlay. UX-only — backend still enforces plan gates on data endpoints. B. /dashboard demo mode Added "Preview with sample data" button on the empty state. Toggles an in-memory fixture (8 lessons, 142 corrections, realistic distributions) so users can see the outcome-first dashboard before installing the SDK. Demo banner explains it's sample data. C. Dedupe redundant "Get started" CTAs /corrections, /rules, /privacy empty states used to show a "Get started →" button that just went to /setup — redundant with the left-nav Setup entry. Replaced with inline text pointer so the CTA isn't duplicated. Tests: 95/95 pass (+11 new: 7 operator + 4 PlanGate). Co-Authored-By: Gradata <noreply@gradata.ai> * fix(dashboard): CR round-1 + promote Preview CTA - operator.ts: reject multi-@ inputs to match backend semantics (prevents "user@evil.com@gradata.ai" bypass drift per CR review) - demo-dashboard.ts: compute Date.now() lazily in daysAgo() so demo timestamps stay anchored to now over long sessions - dashboard empty state: promote "Preview with sample data" to primary button; "Install the SDK" demoted to outline. Was burying the demo affordance behind the SDK pitch. - tests: new security case for multi-@ bypass (96 total, all pass) Co-Authored-By: Gradata <noreply@gradata.ai> * feat(dashboard): marketify pass — plain-language labels Replace analyst jargon with human language throughout the dashboard: KpiStrip (5 cards): - Correction Rate → Mistakes Caught - Est. Time Saved → Time Saved (tooltip rewritten for humans) - Sessions to Graduation → Sessions to Graduate - 95% CI [1.9, 2.7] → typically 2–3 sessions - Misfires → False Alarms - Brain Footprint kept (user likes seeing AI brain grow) ActiveRulesPanel: - "Active Rules" → "Your Rules" - "top 8" → "what your AI learned" - Hide raw confidence number (sim research: users ignore it) - INSTINCT/PATTERN/RULE → Watching/Learning/Graduated - "Xd clean" → "N days holding" - "recurred Nd ago" → "slipped Nd ago" - "No graduated rules yet" → "Nothing graduated yet. Keep correcting — rules emerge after 3+ catches." - "See all rules" → "See all your rules" ActivityFeed: - Rule graduated kept (user preference over "locked in") - Rule refined → Rule updated - Slipped → Slipped back - "Standard codified" → "Your team now gets this automatically" - "More corrections this week" → "More fixes this week" - Empty state softened CategoriesChart: - "Corrections by Dimension" → "What You Fix Most" - "recalibrating" empty state → "still figuring out what you fix most" - Dropped "6-dim taxonomy (WAVE2)" internal badge GraduationProgressBar: - "Graduation Pipeline" → "How Your AI Learns" - Tier labels now Watching/Learning/Graduated (human names) - Dropped threshold/avg-confidence numerics from cards - "N lessons total" → "N total" Dashboard header: - "Your brain's learning progress" → "What your AI learned from you" 96/96 tests pass. Co-Authored-By: Gradata <noreply@gradata.ai> * fix(dashboard): CR round-3 — demo activity, recurrence ordering, category keys - Wire demoActivityEvents fixture into ActivityFeed when demoMode is on so the Activity panel populates in the preview path (was empty/live-only). - Align demoAnalytics.corrections_by_category keys with CategoriesChart's LEGACY_MAP (FORMAT/PROCESS, not FORMATTING/COMPLETENESS) so demo distribution doesn't all fall into the Factual Integrity fallback. - Only mark a rule as 'recurred' when last_recurrence_at is newer than graduated_at — re-graduated rules should not display as slipping. - Replace `as any` casts in ActivityFeed.test.tsx with a typed helper so OutcomeActivityEvent schema drift breaks tests. - Add dashboard-page test for the empty-brain → preview demo → exit flow. Co-Authored-By: Gradata <noreply@gradata.ai> --------- Co-authored-by: Gradata <noreply@gradata.ai> * fix(review): address CR round 5 on PR #30 - lessons.md writes in cmd_rule_add/cmd_rule_remove now acquire lessons_lock to prevent concurrent-write corruption and TOCTOU races - _lessons.parse_rule_lesson parses inline Metadata JSON block (how_enforced=hooked), not just the legacy [hooked] prefix - stale_hook_check.py: shlex.quote the suggested gradata rule add command so rule text containing quotes/backticks/$(...) stays safe - stale_hook_check.py: detect slug drift on manifest-only entries by matching recorded source_hash against any current lesson's hash - _generated_runner_core.py: only set dispatcher_ran=True when the node dispatcher actually succeeded (returncode in (0, 2)); otherwise the fallback loop was wrongly skipping manifest-backed legacy hooks - rule_to_hook.py + cli.py + stale_hook_check.py: delegate hook-root defaults to gradata.hooks._manifest._hook_root for a single source of truth; hardcoded .claude/hooks/... strings live in one place now - self_improvement.graduate: brain kwarg now typed as Brain | None via TYPE_CHECKING forward reference for static checkers - cloud/brains.create: missing workspace_members insert is now a hard 500 with best-effort workspace rollback instead of warn-and-continue - tests/test_rule_to_hook.py: add TestSharedLessonParser covering both legacy [hooked] prefix and structured Metadata JSON parsing paths Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Oliver Le <oliver@gradata.com> Co-authored-by: Gradata <noreply@gradata.ai> Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
1 participant
Summary
Extends S101's
classify_ruleclassifier into a full rule-to-hook pipeline: graduated RULE-tier lessons automatically attempt to install as deterministic PreToolUse hooks, falling back to soft injection on failure. Ships 7 working hook templates, a CLI fast-track (gradata rule add), cross-platform rule export (gradata export --target cursor|agents|aider), and a dispatcher runner registered ingradata hooks install.What's new
Core pipeline (
src/gradata/enhancements/rule_to_hook.py):render_hook(candidate)— rendersHookCandidateinto executable JS hook source via templateself_test(rendered, positive)— runs generated hook against a violating example, confirms exit 2install_hook(slug, source, *, template)— writes toGRADATA_HOOK_ROOT(pre) orGRADATA_HOOK_ROOT_POST(post)try_generate(candidate)— orchestrator: render → self-test → installHookCandidate.template_arg— replacesblock_pattern, typed for template-specific args (regex, line limits, sentinels)Templates shipped (
src/gradata/hooks/templates/):regex_replace— em-dash and similar content regex blockers (PreToolUse)fstring_block—python -c "f'...'"Bash pattern (PreToolUse)root_file_save— blocks writes to root-level paths (PreToolUse)destructive_block—rm -rf,git push --force,DROP TABLE,kubectl delete,git reset --hard(PreToolUse)secret_scan— API keys, GitHub tokens, AWS keys, private keys (PreToolUse)file_size_check— line-count cap extracted from rule text (e.g., "Keep files under 500 lines") (PreToolUse)auto_test— runs pytest againsttest_<basename>.pyafter edits (PostToolUse)Phrasing coverage: 29 description regex patterns across the 7 templates, handling paraphrasing like "avoid em dashes" / "em dashes are banned" / "never hardcode api key" etc.
CLI:
gradata rule add "X"— fast-track user declaration to RULE tier + attempt hook installgradata export --target cursor|agents|aider [--output PATH]— emit graduated rules to.cursorrules/AGENTS.md/.aider.conf.yml(cross-platform soft enforcement for tools without hook APIs)Graduation integration (
src/gradata/enhancements/self_improvement.py):graduate()callsrule_to_hook.try_generate()after PATTERN→RULE promotion[hooked]markersrc/gradata/hooks/rule_enforcement.pydedups[hooked]rules from soft text injectionRuntime wiring:
src/gradata/hooks/_generated_runner_core.py— shared iteration/relay logicsrc/gradata/hooks/generated_runner.py(PreToolUse) +_post.py(PostToolUse) — thin wrappers registered inHOOK_REGISTRYsogradata hooks installpicks them upGRADATA_BYPASS=1env var disables all generated hooks as runtime escape hatchData capture:
Brain.record_correction(text, *, assistant_draft=None, category)persists the raw violating draft into CORRECTION eventdata["draft_text"]so future rules graduating from that event have ground-truth for self-testing.Test plan
pytest tests/test_rule_to_hook.py— 58 tests greenpytest tests/ --ignore=tests/test_brain_benchmark.py— 2111 passed, 23 skipped, zero regressionsgradata rule addon all 6 PreToolUse rule phrasings → 6 installed hooks → each correctly blocks violating payload and allows clean payloadGRADATA_BYPASS=1disables a generated hookauto_testroutes toGRADATA_HOOK_ROOT_POSTdir, exits 0 silently when no test file matches[hooked]marker, exclude PATTERN-tier lessonstemplate_argrename, regex pre-compilation, merged duplicate patterns, requiredtemplatekwarg)Known follow-ups (not in this PR)
Brain.add_rule()API socmd_rule_adddoesn't hand-format lessons.md linesgenerated_runnerrespectingProfilegating (currently fires regardless ofGRADATA_HOOK_PROFILE)Generated with Gradata