just-vibe 0.7.0

package/plugins/just-vibe/skills/llm-tools/SKILL.md

@@ -0,0 +1,58 @@
+---
+name: llm-tools
+description: "Design tool schemas, execution contracts, and failure handling Use to design constrained agent tool interfaces; api-design defines general service contracts."
+---
+
+# llm-tools
+
+Design tool schemas, execution contracts, and failure handling
+
+## Choose this workflow
+
+Use to design constrained agent tool interfaces; api-design defines general service contracts.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [LLMs and retrieval methods](../../references/packs/llm.md) for tool selection and operational details. Resolve these paths from this skill file; all runtime assets ship inside the plugin.
+
+## Input and mode
+
+Use the complete request appended to this invocation, preserving all constraints and references. Default mode: **plan**. Plan; desired actions, schemas, execution APIs, permissions, and failure semantics.
+
+task definition, model/provider configuration, representative permitted data, versioned prompts/corpus where relevant, and explicit token/cost/latency limits for remote calls. Use current provider interfaces during implementation. Retrieved content and model-generated tool arguments remain untrusted.
+
+Resolve any task-specific tools, target identity and evidence before dependent actions. No external connection is assumed.
+
+## Scope
+
+Model-facing tool contracts and validated execution boundaries; implementation on request.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Define each tool’s typed input/output, target identity, permitted action and observable success. Separate planning or proposed arguments from executed effects; untrusted retrieved text cannot grant tool authority.
+- Validate schema and business constraints before dispatch, resolve the actual target from authorized context and minimize returned sensitive data. Describe actionable errors without exposing credentials or treating arbitrary output as new instructions.
+- For mutating calls, define request identity, idempotency, timeout ambiguity and result reconciliation. Cancellation or a missing response does not prove an external operation failed; check its identity before retrying.
+- Test valid calls, invalid inputs, denied targets, unavailable tools, partial success and malicious tool output with controlled fakes. Verify state/effect counts as well as final answers and retain the distinction between simulated and live integration evidence.
+
+## Decision branches
+
+- **When a tool times out after an external mutation may have occurred:** Return an operation ID and reconciliation path; do not let the model blindly repeat it.
+
+## Deliver and verify
+
+- Tool schemas, executor design/code, error contract, and behavior fixtures.
+- Tool schema, enforcement boundary and invalid/unauthorized/partial-failure checks.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Invalid or unauthorized arguments cannot execute; an uncertain external mutation is reconciled before retrying.
+
+## Stop and recover
+
+- A model-generated request is not user authorization. Do not expose arbitrary shell/database access as a convenience tool.
+
+## Example requests
+
+- **Normal (plan):** Design narrow tool schemas with validated targets and idempotent execution.
+- **edge (plan):** Design a deployment tool that rejects unapproved production targets.
+- **blocked (inspect):** Review tool interfaces without exposing arbitrary shell or database execution.

package/plugins/just-vibe/skills/map/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: map
+description: "Produce an architecture or dependency map Use for module dependencies inside a repository; arch-map covers deployed services and stores."
+---
+
+# map
+
+Produce an architecture or dependency map
+
+## Choose this workflow
+
+Use for module dependencies inside a repository; arch-map covers deployed services and stores.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [General methods](../../references/packs/general.md) for tool selection and operational details. Resolve these paths from this skill file; all runtime assets ship inside the plugin.
+
+## Input and mode
+
+Use the complete request appended to this invocation, preserving all constraints and references. Default mode: **inspect**. Inspect; repository/subsystem and desired map depth. Requires source and manifests.
+
+Resolve the user brief and inspect the relevant project or supplied evidence. External capabilities are optional unless the selected action actually needs them.
+
+Declared evidence requirements: `project.read`. Use actual host discovery or adequate supplied artifacts; unavailable evidence remains blocked/unknown.
+
+## Scope
+
+Existing modules and dependencies; deeper distributed architecture questions belong to `arch-map`.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Identify nodes and imports, trace representative relationships, separate declared from observed dependencies, and select a readable level of detail.
+- Identify public entry points and dependency direction; collapse generated/vendor code and distinguish imports, calls and data sharing.
+
+## Decision branches
+
+- **When static analysis cannot resolve dynamic loading:** Mark the edge inferred and inspect registration/configuration sites instead of inventing a dependency.
+
+## Deliver and verify
+
+- Diagram or adjacency table with a legend, entry points, evidence links, and uncertain edges.
+- Diagram legend, representative dependency paths, cycles and unresolved dynamic edges.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Detects a real dependency cycle; generated/vendor directories do not overwhelm the map.
+
+## Stop and recover
+
+- Cap graph expansion at the requested boundary and summarize external nodes. Do not present the diagram as an approved future architecture.
+
+## Example requests
+
+- **Normal (inspect):** Map dependencies between the billing modules; include cycles.
+- **edge (inspect):** Map modules including a plugin loaded from configuration.
+- **blocked (inspect):** Map this partial source snapshot without claiming complete dependency coverage.

package/plugins/just-vibe/skills/match/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: match
+description: "Compare an implementation against a visual reference and close gaps Use when a supplied visual reference is the acceptance target; design creates a direction when no exact target exists."
+---
+
+# match
+
+Compare an implementation against a visual reference and close gaps
+
+## Choose this workflow
+
+Use when a supplied visual reference is the acceptance target; design creates a direction when no exact target exists.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [General methods](../../references/packs/general.md) for tool selection and operational details. Resolve these paths from this skill file; all runtime assets ship inside the plugin.
+
+## Input and mode
+
+Use the complete request appended to this invocation, preserving all constraints and references. Default mode: **apply**. Apply; reference image/design, target implementation, viewport, and assets. Requires readable reference and render capability for visual proof.
+
+Resolve the user brief and inspect the relevant project or supplied evidence. External capabilities are optional unless the selected action actually needs them.
+
+Declared evidence requirements: `project.read`. Use actual host discovery or adequate supplied artifacts; unavailable evidence remains blocked/unknown.
+
+## Scope
+
+Match layout, typography, imagery, and states within the given reference; no invented unseen product behavior.
+
+Only the requested local changes; external actions require their exact action and target in session authorization.
+
+## Execute
+
+- Analyze reference geometry, compare an equivalent render, prioritize largest discrepancies, implement changes, and repeat bounded comparisons.
+- Match viewport, content and fonts before comparing geometry; adjust large layout discrepancies before decorative details and record legitimate responsive differences.
+
+## Decision branches
+
+- **When the reference omits a required state:** Preserve established interaction behavior and label the added state as an interpretation.
+
+## Deliver and verify
+
+- Updated UI plus matched-state comparison and remaining differences.
+- Reference conditions, matched regions, remaining differences and responsive checks.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Layout aligns at the reference viewport; responsive behavior is retained outside that viewport.
+
+## Stop and recover
+
+- Missing fonts/assets are identified explicitly. Do not call a match pixel-perfect without a controlled comparison.
+
+## Example requests
+
+- **Normal (apply):** Match this supplied pricing-page reference at its captured viewport.
+- **edge (apply):** Match this desktop reference while retaining a usable narrow layout.
+- **blocked (inspect):** Analyze a reference with unavailable brand assets; distinguish substitutes from exact matches.

package/plugins/just-vibe/skills/migrate/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: migrate
+description: "Plan and apply a version, schema, or implementation migration Use for coordinated version or platform transitions; db-migrate handles database-specific mechanics."
+---
+
+# migrate
+
+Plan and apply a version, schema, or implementation migration
+
+## Choose this workflow
+
+Use for coordinated version or platform transitions; db-migrate handles database-specific mechanics.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [General methods](../../references/packs/general.md) for tool selection and operational details. Resolve these paths from this skill file; all runtime assets ship inside the plugin.
+
+## Input and mode
+
+Use the complete request appended to this invocation, preserving all constraints and references. Default mode: **plan**. Plan; source/target version or implementation, compatibility needs, and rollout environment. Apply only when migration execution is requested.
+
+Resolve the user brief and inspect the relevant project or supplied evidence. External capabilities are optional unless the selected action actually needs them.
+
+Declared evidence requirements: `project.read`. Use actual host discovery or adequate supplied artifacts; unavailable evidence remains blocked/unknown.
+
+## Scope
+
+Ordered migration across affected code/config/data contracts; no unrelated upgrades.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Inventory dependents, read version-specific changes, design transitional compatibility, prepare edits/checks, and define recovery before execution.
+- Inventory old/new consumers and persisted formats; identify the last reversible point and validate coexistence before removing compatibility code.
+
+## Decision branches
+
+- **When the target rejects an old persisted format:** Add an explicit conversion and recovery path before changing readers.
+
+## Deliver and verify
+
+- Migration sequence or authorized patch, compatibility matrix, verification, and recovery limitations.
+- Compatibility matrix, ordered transitions, recovery point and verified/unverified stages.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Old/new overlap behaves as specified; interrupted execution has a documented restart or recovery path.
+
+## Stop and recover
+
+- Irreversible data loss, unsupported targets, or ambiguous production scope must be resolved before dependent mutations.
+
+## Example requests
+
+- **Normal (plan):** Plan upgrading the job library while old workers remain active.
+- **edge (plan):** Migrate a library while older workers continue reading stored jobs.
+- **blocked (inspect):** Plan a migration with missing legacy fixtures; identify the compatibility evidence still needed.

package/plugins/just-vibe/skills/ml-ablation/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: ml-ablation
+description: "Measure contributions of features or model components Use to estimate component contribution; ml-tune optimizes parameters."
+---
+
+# ml-ablation
+
+Measure contributions of features or model components
+
+## Choose this workflow
+
+Use to estimate component contribution; ml-tune optimizes parameters.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [ML experimentation methods](../../references/packs/ml-experiments.md) for tool selection and operational details. Resolve these paths from this skill file; all runtime assets ship inside the plugin.
+
+## Input and mode
+
+Use the complete request appended to this invocation, preserving all constraints and references. Default mode: **plan**. Plan; hypothesis, reference model, feature/component variants, fixed evaluation, and budget.
+
+dataset/split manifests, fixed objective/metric, environment/dependencies, baseline where applicable, and explicit compute limits. Record code revision, configuration, seeds, artifact paths, and resource use. Local smoke checks do not imply authorization for paid training. Never optimize on the held-out test set.
+
+Resolve any task-specific tools, target identity and evidence before dependent actions. No external connection is assumed.
+
+## Scope
+
+Isolate the contribution of specified components using controlled comparisons.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Define one meaningful variation at a time or a justified factorial design, hold data/evaluation constant, repeat seeds when needed, and compare uncertainty/cost.
+- State the causal comparison, hold data/protocol constant and repeat seeds or matched runs where variance could overwhelm the effect.
+
+## Decision branches
+
+- **When removing a component changes another required contract:** Redesign the comparison or disclose the confound instead of attributing all change to one component.
+
+## Deliver and verify
+
+- Ablation protocol or run results with attributable differences and caveats.
+- Variant matrix, controlled differences, uncertainty and cost/quality interpretation.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Data changes do not confound the component comparison; noisy score differences are not declared decisive.
+
+## Stop and recover
+
+- Do not remove interdependent components without explaining the changed model contract. Budget limits constrain repeats and confidence.
+
+## Example requests
+
+- **Normal (plan):** Plan a controlled comparison of the new features at fixed data and seed conditions.
+- **edge (plan):** Ablate a feature group while preprocessing depends on those columns.
+- **blocked (inspect):** Plan an ablation with insufficient run budget; state the confidence limitation.

package/plugins/just-vibe/skills/ml-baseline/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: ml-baseline
+description: "Establish simple, reproducible baselines Use for a first comparable benchmark; ml-tune searches hyperparameters after protocol validity."
+---
+
+# ml-baseline
+
+Establish simple, reproducible baselines
+
+## Choose this workflow
+
+Use for a first comparable benchmark; ml-tune searches hyperparameters after protocol validity.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [ML experimentation methods](../../references/packs/ml-experiments.md) for tool selection and operational details. Resolve these paths from this skill file; all runtime assets ship inside the plugin.
+
+## Input and mode
+
+Use the complete request appended to this invocation, preserving all constraints and references. Default mode: **plan**. Plan; task, training/validation split, metric, and run budget. Explicit baseline-run requests select apply.
+
+dataset/split manifests, fixed objective/metric, environment/dependencies, baseline where applicable, and explicit compute limits. Record code revision, configuration, seeds, artifact paths, and resource use. Local smoke checks do not imply authorization for paid training. Never optimize on the held-out test set.
+
+Resolve any task-specific tools, target identity and evidence before dependent actions. No external connection is assumed.
+
+## Scope
+
+Simple reproducible reference models, including a naive/rule-based comparator.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Match baseline to task, build leakage-safe preprocessing, fix evaluation protocol, run bounded training when authorized, and report score plus cost/uncertainty.
+- Include a task-appropriate constant/rule baseline and simple model, use the same splits and preprocessing fit boundaries and record resources alongside quality.
+
+## Decision branches
+
+- **When the baseline wins or matches within uncertainty:** Keep the simpler option and identify the missing signal before expanding model search.
+
+## Deliver and verify
+
+- Baseline configuration/code and results if executed.
+- Data/split identity, baseline/model comparison, metric definitions and compute record.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- A naive comparator is included; all candidates use identical permitted evaluation data.
+
+## Stop and recover
+
+- Do not begin extensive model search. Missing label/split validity blocks trustworthy benchmark claims even if training technically runs.
+
+## Example requests
+
+- **Normal (plan):** Plan reproducible naive and simple-model baselines under the fixed split.
+- **edge (plan):** Establish a baseline when class imbalance makes accuracy misleading.
+- **blocked (inspect):** Plan a baseline with unresolved label maturity; do not claim trustworthy scores.

package/plugins/just-vibe/skills/ml-batch/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: ml-batch
+description: "Build resumable batch inference and output tracking Use for resumable batch inference; ml-serving handles request/response service behavior."
+---
+
+# ml-batch
+
+Build resumable batch inference and output tracking
+
+## Choose this workflow
+
+Use for resumable batch inference; ml-serving handles request/response service behavior.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [ML deployment methods](../../references/packs/ml-deployment.md) for tool selection and operational details. Resolve these paths from this skill file; all runtime assets ship inside the plugin.
+
+## Input and mode
+
+Use the complete request appended to this invocation, preserving all constraints and references. Default mode: **apply**. Apply for implementation; model/data versions, partitioning, output keys, checkpoint destination, and run budget.
+
+versioned model and preprocessing artifacts, input/output schema, runtime/dependencies, operating targets, and authorized environment. Validate artifact trust before loading formats that can execute code. Packaging or writing monitoring configuration does not deploy a model or enable a hosted service.
+
+Resolve any task-specific tools, target identity and evidence before dependent actions. No external connection is assumed.
+
+## Scope
+
+Resumable batch inference; real dataset execution must be part of the request.
+
+Only the requested local changes; external actions require their exact action and target in session authorization.
+
+## Execute
+
+- Validate schemas, create stable row/partition identities, implement checkpointed writes, track failures and model versions, and test resume/replay on controlled input.
+- Freeze model and input snapshot identity, partition by stable keys, stage outputs and commit a manifest/checkpoint only after durable complete partitions.
+
+## Decision branches
+
+- **When a restart finds partial output or a different model version:** Reconcile or isolate it before resuming; never silently mix incompatible predictions.
+
+## Deliver and verify
+
+- Batch job, progress/output manifest, error policy, and resume evidence.
+- Partition/model manifest, failure counts, output reconciliation and resume steps.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Restart does not duplicate completed outputs; invalid records are accounted for rather than silently lost.
+
+## Stop and recover
+
+- No unbounded full-dataset inference. Do not mix predictions from incompatible model versions in one unlabeled output.
+
+## Example requests
+
+- **Normal (apply):** Implement resumable batch inference with stable output keys and model-version tracking.
+- **edge (apply):** Resume inference after output writes succeeded but checkpointing failed.
+- **blocked (inspect):** Plan batch prediction without scanning the full dataset or launching unbounded compute.

package/plugins/just-vibe/skills/ml-calibrate/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: ml-calibrate
+description: "Assess predicted probabilities against observed outcomes Use to assess or fit probability calibration; ml-threshold maps scores to decisions."
+---
+
+# ml-calibrate
+
+Assess predicted probabilities against observed outcomes
+
+## Choose this workflow
+
+Use to assess or fit probability calibration; ml-threshold maps scores to decisions.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [ML evaluation methods](../../references/packs/ml-evaluation.md) for tool selection and operational details. Resolve these paths from this skill file; all runtime assets ship inside the plugin.
+
+## Input and mode
+
+Use the complete request appended to this invocation, preserving all constraints and references. Default mode: **inspect**. Inspect; predicted probabilities, labels, sampling/prevalence context, and intended use.
+
+frozen model/artifact, evaluation dataset identity, labels where needed, metric definitions, and task/operating context. Report sample counts and uncertainty appropriate to dependencies; avoid repeated test-set tuning. Exploratory findings need fresh confirmation before strong generalization claims.
+
+Declared evidence requirements: `ml.artifacts`. Use actual host discovery or adequate supplied artifacts; unavailable evidence remains blocked/unknown.
+
+## Scope
+
+Probability reliability assessment; fitting calibration requires a separate held-out calibration protocol and explicit apply request.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Validate probability semantics, inspect reliability by range/cohort, use appropriate scoring measures, and compare any authorized calibrator on untouched evaluation data.
+- Check probability semantics, reliability and proper scoring measures; fit any calibrator on permitted data separate from final evaluation and compare by cohort.
+
+## Decision branches
+
+- **When prevalence or sampling changed since calibration:** Assess transfer assumptions and recalibration evidence without claiming ranking quality proves reliable probabilities.
+
+## Deliver and verify
+
+- Calibration report or calibrated artifact with split provenance and checks.
+- Calibration protocol, reliability evidence and independent comparison.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Good ranking is not mistaken for calibrated probability; fitting and evaluating a calibrator on identical records is rejected.
+
+## Stop and recover
+
+- Small samples and prevalence shift limit conclusions. Do not alter deployed probabilities without a rollout request.
+
+## Example requests
+
+- **Normal (inspect):** Assess probability calibration from the supplied held-out scores and outcomes.
+- **edge (inspect):** Calibrate a model trained on oversampled positives.
+- **blocked (inspect):** Review probability outputs with too few outcomes to fit a reliable calibrator.

package/plugins/just-vibe/skills/ml-dataset/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: ml-dataset
+description: "Audit whether data can support the modeling task Use to assess whether data supports a task; data-profile summarizes its columns."
+---
+
+# ml-dataset
+
+Audit whether data can support the modeling task
+
+## Choose this workflow
+
+Use to assess whether data supports a task; data-profile summarizes its columns.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [ML data methods](../../references/packs/ml-data.md) for tool selection and operational details. Resolve these paths from this skill file; all runtime assets ship inside the plugin.
+
+## Input and mode
+
+Use the complete request appended to this invocation, preserving all constraints and references. Default mode: **inspect**. Inspect; framed task, dataset version, collection process, and sampling budget.
+
+task definition, dataset identity, field semantics, entity/time keys, and permission to inspect bounded data. Record prediction moment, label horizon, sampling, and provenance. Preserve held-out evaluation boundaries; no data upload, label alteration, or feature fitting across splits implicitly.
+
+Declared evidence requirements: `data.read`. Use actual host discovery or adequate supplied artifacts; unavailable evidence remains blocked/unknown.
+
+## Scope
+
+Task suitability, population coverage, missingness, dependencies, and collection bias.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Compare available fields/outcomes with task needs, inspect cohort/time coverage, assess missing-label patterns and selection processes, and identify unsupported deployment populations.
+- Compare collection/selection and follow-up windows with deployment population, inspect coverage by cohort/time and identify censored or missing outcomes.
+
+## Decision branches
+
+- **When training data excludes the intended deployment cohort:** Limit generalization claims and propose evidence collection before model complexity.
+
+## Deliver and verify
+
+- Dataset readiness report with exclusions, risks, and needed collection/validation work.
+- Population/coverage table, missingness/selection risks and supported deployment claims.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- A dataset missing outcome follow-up is not treated as fully labeled; a deployment cohort absent from training is flagged.
+
+## Stop and recover
+
+- Do not infer representativeness from sample size alone. Unsupported collection semantics remain unknown.
+
+## Example requests
+
+- **Normal (inspect):** Assess whether this dataset covers the intended deployment population.
+- **edge (inspect):** Assess a dataset with many rows but no labels for recently enrolled users.
+- **blocked (inspect):** Review dataset metadata without row access; avoid inferring representative coverage.

package/plugins/just-vibe/skills/ml-dataset-version/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: ml-dataset-version
+description: "Record dataset identity, transformations, and provenance Use to identify reproducible data/splits; data-lineage explains transformations."
+---
+
+# ml-dataset-version
+
+Record dataset identity, transformations, and provenance
+
+## Choose this workflow
+
+Use to identify reproducible data/splits; data-lineage explains transformations.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [ML data methods](../../references/packs/ml-data.md) for tool selection and operational details. Resolve these paths from this skill file; all runtime assets ship inside the plugin.
+
+## Input and mode
+
+Use the complete request appended to this invocation, preserving all constraints and references. Default mode: **apply**. Apply; dataset snapshot, transforms, source identifiers, and approved manifest location.
+
+task definition, dataset identity, field semantics, entity/time keys, and permission to inspect bounded data. Record prediction moment, label horizon, sampling, and provenance. Preserve held-out evaluation boundaries; no data upload, label alteration, or feature fitting across splits implicitly.
+
+Declared evidence requirements: `data.read`. Use actual host discovery or adequate supplied artifacts; unavailable evidence remains blocked/unknown.
+
+## Scope
+
+Reproducible identity/provenance metadata; no automatic copying or committing of raw data.
+
+Only the requested local changes; external actions require their exact action and target in session authorization.
+
+## Execute
+
+- Record immutable references or hashes where feasible, schema, transformation/code versions, split identity, and creation parameters; verify referential accessibility.
+- Record source snapshot or content identity, transformation revision, schema, split membership and immutable access references without storing secrets or raw private data.
+
+## Decision branches
+
+- **When only a mutable source URL exists:** Mark identity provisional and define the snapshot/hash mechanism needed for reproducibility.
+
+## Deliver and verify
+
+- Dataset manifest with lineage, reconstruction instructions, and known reproducibility limits.
+- Version manifest, provenance chain, accessibility check and mutable boundaries.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Changed source content changes identity or is detected; a mutable URL alone is not presented as an immutable version.
+
+## Stop and recover
+
+- Avoid expensive full hashing without a budget. Do not place confidential data or credentials in Git/manifests.
+
+## Example requests
+
+- **Normal (apply):** Write a provenance manifest for the identified dataset without copying raw records.
+- **edge (apply):** Version a dataset whose remote contents can change at the same path.
+- **blocked (inspect):** Plan versioning without permission for a full expensive hash scan.

package/plugins/just-vibe/skills/ml-debug-training/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: ml-debug-training
+description: "Investigate exploding loss, unstable gradients, NaNs, or failure to learn Use for NaNs, shape/device errors or non-learning; ml-error-analysis studies generalization failures."
+---
+
+# ml-debug-training
+
+Investigate exploding loss, unstable gradients, NaNs, or failure to learn
+
+## Choose this workflow
+
+Use for NaNs, shape/device errors or non-learning; ml-error-analysis studies generalization failures.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [ML experimentation methods](../../references/packs/ml-experiments.md) for tool selection and operational details. Resolve these paths from this skill file; all runtime assets ship inside the plugin.
+
+## Input and mode
+
+Use the complete request appended to this invocation, preserving all constraints and references. Default mode: **inspect**. Inspect; failing run logs/configuration, batches, model, and training symptom.
+
+dataset/split manifests, fixed objective/metric, environment/dependencies, baseline where applicable, and explicit compute limits. Record code revision, configuration, seeds, artifact paths, and resource use. Local smoke checks do not imply authorization for paid training. Never optimize on the held-out test set.
+
+Declared evidence requirements: `ml.artifacts`. Use actual host discovery or adequate supplied artifacts; unavailable evidence remains blocked/unknown.
+
+## Scope
+
+Numerical instability, shape/device issues, data/target mismatch, gradients, and failure to learn.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Check inputs/loss/optimizer state, compare expected scales, isolate a small batch, propose or run authorized overfit/gradient probes, and test the leading cause.
+- Inspect one batch's shapes, labels, scale, loss and gradients, locate the first non-finite value and compare optimizer updates with the intended objective.
+
+## Decision branches
+
+- **When a tiny-batch overfit probe fails:** Investigate data/loss/gradient/update plumbing before larger architectures or more epochs.
+
+## Deliver and verify
+
+- Diagnosis, minimal corrective change when requested, and controlled evidence.
+- First divergent tensor/step, hypothesis evidence and bounded corrective probe.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- NaNs are localized to their first source; an inability to overfit a tiny batch prompts pipeline investigation rather than a larger model.
+
+## Stop and recover
+
+- No unbounded retraining or random parameter changes. Preserve failed-run evidence and distinguish numerical repair from generalization improvement.
+
+## Example requests
+
+- **Normal (inspect):** Diagnose NaNs from this run's first failing batch and gradient logs.
+- **edge (inspect):** Debug a loss that becomes NaN only after mixed-precision updates.
+- **blocked (inspect):** Inspect saved training logs without retraining or guessing a learning-rate cure.