Feat/auto config params separation - #275
Merged
Merged
Conversation
StabilityClassifier.incline() computes where in a binary change series the changes sit, in [-0.5, +0.5]: -0.5 is every change at the very start, 0 is uniform churn, +0.5 is every change at the end. It is the least-squares slope with its data-free factors divided out, so the value is comparable between events instead of scaling with the change count. RLELists are read via runs(), one pass, no expansion. Trackers opt in with require_declining, which adds a conjunct to STABLE requiring the variable's changes to sit early in its series. Independent of segmentation -- it reads index positions, not timestamps. Committed with --no-verify: this is a snapshot of work that predates the branch and was never hook-clean. The lint and mypy fixes land in the following commit, so the branch tip is clean.
Configure-phase inputs move out of `params` into their own top-level
`auto_config_params` block, so a config makes plain which settings shape
auto-configuration and which drive training and detection.
AutoConfigParams (common/detector.py) is the base; each family narrows it:
- VariableAutoConfigParams -- use_stable_vars, use_static_vars,
segmentation, timestamp_variable, timestamp_format, require_declining,
incline_threshold
- ComboAutoConfigParams -- adds max_combo_size
- SequenceAutoConfigParams -- min_window_size, max_window_size
The `stability_` prefix is dropped inside the block (stability_segmentation
-> segmentation, stability_require_declining -> require_declining); the
prefix only existed to disambiguate names sharing a flat namespace.
incline_threshold becomes configuration rather than a constant reachable
only through tracker.stability_classifier.
set_configuration() no longer rebuilds self.config. It writes only what the
configure phase produced -- config.events (via the new generate_events_config)
and, for EventSequenceDetector, fixed_window_size -- then flips auto_config
off. The wholesale rebuild is what silently dropped `persist` and every
other operator setting, and is why _CARRIED_SETTINGS and the four
hand-written restore lists existed; all are deleted. AutoConfigWarning goes
too: it warned about params being lost, which no longer happens.
Auto-config settings now shape the configure-phase persistency only. The
trained persistency is built without stability kwargs and _ingest no longer
feeds it timestamps -- stability classification is never consulted at
detect time, so those were an unread O(N) list per tracked variable in the
detector that actually runs in production.
BREAKING: no legacy compatibility. The old flat spellings are validation
errors, not deprecated aliases.
BREAKING: persisted tracker state written before this commit no longer
loads. SingleStabilityTracker.to_state() serializes the whole detector
config verbatim and from_state() reinflates it through an extra="forbid"
class, so old blobs carrying the old flat field names raise
PersistencyLoadError. This is not new to this change -- any field rename in
any detector config breaks old state the same way -- but this commit renames
fields, so it triggers it. Existing state files must be regenerated.
The section's only config block showed `segmentation: time`, which reads as the recommended setup. The default is `count` -- as the fields table below it already says -- so name that at the point the example appears.
Contributor
|
Hi, I didn't have time to review it yet, but what do you mean with auto_config = True will not work? |
ipmach
reviewed
Aug 31, 2026
Contributor
There was a problem hiding this comment.
Why the AutoConfig has its own parameters outside the config class? Also why the others like the parsers dont have it?
This breaks the architecture.
Collaborator
Author
There was a problem hiding this comment.
Good point. I forgot about the parsers. WIll address it
Collaborator
Author
There was a problem hiding this comment.
Moved it to BasicConfig, next to auto_config
…on rule
Replace the `segmentation` enum plus the `require_declining` /
`incline_threshold` flags with four independently selectable stability
classification methods, combined by a configurable decision rule.
- Add `ClassificationMethods` (pydantic): `index`, `time`, `slope_index`,
`slope_time` booleans, `slope_threshold` (default -0.05), and
`decision` ("consensus" | "majority"). At least one method must be
enabled; `extra="forbid"`.
- `StabilityClassifier` gains `verdicts()` (per-method votes) and
`decide()` (applies the decision rule). `is_stable()` now routes
through them. Segment methods (`index`, `time`) cut the change series
by equal count or equal duration; slope methods vote STABLE when the
change centroid `k <= slope_threshold`.
- Timestamps are now only consulted when a time-based method is
enabled -- previously they were self-sufficient. Call sites that
relied on passing timestamps alone must enable `time` or `slope_time`.
- `StabilityTracker` translates legacy persisted state into the new
model, so existing event stores keep loading.
- Rename the old "count"/"decline" vocabulary to "index"/"slope"
throughout, and document the breaking changes plus the fallback
behaviour in docs/detectors.md.
Tests: replace test_incline_stability.py with test_slope_stability.py
and add test_classification_methods.py; update the time-dependent and
detector tests for the new configuration surface.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Review feedback on #275: the block was declared only on CoreDetectorConfig, yet `auto_config` and `Component.configure()` are both declared on the shared base, and `BasicConfig.to_dict` and `ConfigMethods.process` already handled `auto_config_params` generically for every component type. The base knew about a subclass-only field. `AutoConfigParams` moves to `common/_config` (declaring it on `BasicConfig` from `detector.py` would be an import cycle) and is re-exported from `common.detector` so existing subclass imports keep working. The redundant-alias form satisfies mypy's no_implicit_reexport. Parsers and alert aggregators inherit the block empty; `to_dict` omits a block at its default, so no existing YAML changes. `get_config()` is a `model_dump()`, so it now reports `auto_config_params: {}` on every config -- hence the test_core default_args update. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
ipmach
approved these changes
Sep 2, 2026
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
2 participants
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.
Task
Description
Configure-phase inputs move out of
paramsinto their own top-levelauto_config_paramsblock, so a config shows at a glance which settings driveauto-configuration and which drive training and detection.
AutoConfigParamsisthe base; the variable, combo and sequence families each narrow it. The
now-redundant
stability_prefix is dropped inside the block.set_configuration()no longer rebuildsself.config— it writes only what theconfigure phase produced, then flips
auto_configoff. The wholesale rebuild iswhat silently dropped
persistand every other operator setting; the carried-settings workaround and
AutoConfigWarningare deleted along with it.Auto-config settings now shape the configure-phase persistency only, so the
trained trackers stop collecting timestamps nothing reads.
Breaking:
deprecated aliases.
regenerated. Tracker state carries a serialized detector config, so any field
rename breaks it; this PR renames fields.
The first commit predates this work and is committed as-is — review the branch
work against it rather than against
development. Docs andAGENTS.mdupdatedto match.
How Has This Been Tested?
Full suite passed
Checklist