just-vibe 0.7.0

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

@@ -0,0 +1,62 @@
+---
+name: ml-train
+description: "Implement training with checkpoints and recorded configuration Use for bounded training implementation/execution; ml-debug-training diagnoses a failed optimization process."
+---
+
+# ml-train
+
+Implement training with checkpoints and recorded configuration
+
+## Choose this workflow
+
+Use for bounded training implementation/execution; ml-debug-training diagnoses a failed optimization process.
+
+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 when planning is requested or execution details are unresolved. A request to implement training code selects apply for code and bounded local checks; launching training requires the requested environment and resource limits. Do not treat implementation as permission to provision compute.
+
+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
+
+Reproducible training with checkpoint/resume and declared stopping criteria.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Validate shapes and pipeline, run a small smoke test, record configuration/environment, train within bounds, checkpoint, and evaluate only the permitted validation protocol.
+- For resumable training inventory model, optimizer, scheduler, scaler when used, step, RNG and sampler/data position; checkpoint atomically and compare interrupted versus uninterrupted continuation under declared tolerances.
+
+## Read when relevant
+
+- Selecting classical, tensor or distributed/resumable methods: [Training scenarios](../../references/scenarios/training.md).
+
+## Decision branches
+
+- **When only weights were saved:** Treat loading as initialization unless all required continuation state is available; label the run a restart rather than exact resume.
+- **When implementation is requested but a training run is not:** Write the pipeline and bounded checks without provisioning or launching a long job; report unmeasured model quality.
+- **When distributed training changes workers or accumulation boundaries:** Verify sampler/metric aggregation and checkpoint ownership; declare approximate continuation if exact state cannot be restored.
+
+## Deliver and verify
+
+- Training code/run record, checkpoints, metrics, and resume instructions.
+- Training configuration, resource cap, artifact/state manifest, metrics and demonstrated resume conditions.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Interrupted training can resume with identified state; NaNs or budget exhaustion stop clearly without a false completed result.
+
+## Stop and recover
+
+- No paid hardware provisioning implicitly. Do not label the best validation checkpoint as independently test-validated.
+
+## Example requests
+
+- **Normal (plan):** Plan checkpointed training on one GPU for at most six hours; do not provision compute.
+- **edge (plan):** Implement checkpointed training that resumes mid-epoch without silently changing sample order.
+- **blocked (inspect):** Plan training with no authorized hardware budget; do not provision or launch a run.

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

@@ -0,0 +1,56 @@
+---
+name: ml-training-cost
+description: "Profile training time, memory, and resource bottlenecks Use to analyze training resource use; ml-inference-perf measures deployed prediction work."
+---
+
+# ml-training-cost
+
+Profile training time, memory, and resource bottlenecks
+
+## Choose this workflow
+
+Use to analyze training resource use; ml-inference-perf measures deployed prediction work.
+
+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; training traces, hardware, workload, budget, and cost-rate evidence.
+
+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
+
+Time, memory, utilization, data loading, and cost bottlenecks.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Separate startup/loading/compute/checkpoint time, inspect batch/resource utilization, identify bottlenecks, and propose measured optimizations or bounded profiling.
+- Separate data loading, host-to-device transfer, compute, synchronization and checkpoint time; relate utilization to the same quality target and workload.
+
+## Decision branches
+
+- **When throughput improves by changing effective batch or precision:** Compare convergence/quality and total time-to-target before claiming a useful speedup.
+
+## Deliver and verify
+
+- Resource breakdown, unit-cost assumptions, and prioritized improvement experiments.
+- Timing/resource breakdown, verified price basis if used and bounded optimization plan.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Idle accelerator time caused by input loading is recognized; local elapsed time is not converted into a fabricated cloud bill.
+
+## Stop and recover
+
+- New profiling/training requires execution authorization. Do not claim speedup without equivalent model quality and workload comparison.
+
+## Example requests
+
+- **Normal (inspect):** Analyze the supplied profile for loading, compute, and checkpoint bottlenecks.
+- **edge (inspect):** Diagnose an idle accelerator stalled by data loading.
+- **blocked (inspect):** Estimate training effort from sparse logs without inventing cloud prices.

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

@@ -0,0 +1,56 @@
+---
+name: ml-tune
+description: "Design a bounded hyperparameter search with a fixed evaluation protocol Use for a bounded search under a valid protocol; ml-ablation isolates component contribution."
+---
+
+# ml-tune
+
+Design a bounded hyperparameter search with a fixed evaluation protocol
+
+## Choose this workflow
+
+Use for a bounded search under a valid protocol; ml-ablation isolates component contribution.
+
+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; baseline, search space, objective, fixed splits, trial/time/compute caps, and selection rule.
+
+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
+
+Bounded hyperparameter search; execution only within authorized resources.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Validate comparable trials, select search strategy, define pruning/failure behavior, log every trial, and choose by the predeclared validation criterion.
+- Freeze search space, split, objective, trial/resource caps and selection rule; keep failure/pruning records and compare candidates under equal evaluation conditions.
+
+## Decision branches
+
+- **When tuning repeatedly consults held-out test results:** Stop that selection loop and define fresh independent confirmation before reporting generalization.
+
+## Deliver and verify
+
+- Search configuration, trial ledger, selected candidate, and cost/selection caveats.
+- Trial ledger, budgets consumed, selection rationale and untouched confirmation set.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Trial failures remain visible; test labels never influence parameter selection.
+
+## Stop and recover
+
+- Stop at any resource cap and preserve partial results. Do not enlarge the search merely because no improvement appears.
+
+## Example requests
+
+- **Normal (plan):** Plan at most 20 trials on validation PR-AUC; never tune on the test set.
+- **edge (plan):** Tune with failed trials and a strict GPU-hour cap.
+- **blocked (inspect):** Design tuning when compute is unavailable; do not fabricate winning hyperparameters.

package/plugins/just-vibe/skills/ops-alerts/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: ops-alerts
+description: "Design actionable alerts with ownership and response guidance Use to design response-worthy alerts; ops-observability supplies reliable signals."
+---
+
+# ops-alerts
+
+Design actionable alerts with ownership and response guidance
+
+## Choose this workflow
+
+Use to design response-worthy alerts; ops-observability supplies reliable signals.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [Operations methods](../../references/packs/operations.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; service objectives, telemetry, response ownership, notification destination, and noise tolerance.
+
+exact service/environment, time window, revision/configuration identity, authorized logs/metrics, and operational constraints. Prefer observation before intervention; live restarts, traffic changes, restores, and notifications require the requested target/action. Redact sensitive telemetry.
+
+Declared evidence requirements: `project.read`. Use actual host discovery or adequate supplied artifacts; unavailable evidence remains blocked/unknown.
+
+## Scope
+
+Actionable alert rules and response guidance; enabling/sending notifications requires an explicit request.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Tie signals to impact, define windows/thresholds and missing-data behavior, test historical/fixture events, include recovery and suppression rules, and identify the responder action.
+- Tie each alert to impact and responder action, define evaluation/recovery windows, missing-data behavior and noise suppression, then test historical or synthetic cases.
+
+## Decision branches
+
+- **When signal quality cannot distinguish failure from missing telemetry:** Add an explicit missing-data state instead of silently treating absence as healthy.
+
+## Deliver and verify
+
+- Alert definitions, rationale, runbook links, and false-positive/missed-event checks.
+- Alert contract, trigger/recovery examples, owner/destination assumptions and runbook action.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Sustained harmful failure triggers; expected brief transients do not page indiscriminately.
+
+## Stop and recover
+
+- Do not invent owners or send test pages implicitly. Missing signal quality is a prerequisite, not something thresholds can hide.
+
+## Example requests
+
+- **Normal (plan):** Design actionable error-rate alerts with recovery behavior; do not enable notifications.
+- **edge (plan):** Design alerts that ignore brief spikes but detect sustained customer failures.
+- **blocked (inspect):** Draft alerts without sending pages or inventing response ownership.

package/plugins/just-vibe/skills/ops-container/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: ops-container
+description: "Diagnose container builds, runtime failures, and configuration differences Use for image/build/runtime diagnosis; ops-restore handles recovery of persisted state."
+---
+
+# ops-container
+
+Diagnose container builds, runtime failures, and configuration differences
+
+## Choose this workflow
+
+Use for image/build/runtime diagnosis; ops-restore handles recovery of persisted state.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [Operations methods](../../references/packs/operations.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; Dockerfile/image/runtime configuration, logs, and failing build/start behavior.
+
+exact service/environment, time window, revision/configuration identity, authorized logs/metrics, and operational constraints. Prefer observation before intervention; live restarts, traffic changes, restores, and notifications require the requested target/action. Redact sensitive telemetry.
+
+Declared evidence requirements: `container.context`. Use actual host discovery or adequate supplied artifacts; unavailable evidence remains blocked/unknown.
+
+## Scope
+
+Container build context, dependencies, permissions, entrypoint, networking, resources, and health checks.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Compare build/runtime stages and host assumptions, inspect image metadata/logs, reproduce in isolation when authorized, and propose or apply a focused fix.
+- Compare build context, multi-stage copy paths, runtime user, working directory, ports and volume permissions with logs from the intended image digest.
+
+## Decision branches
+
+- **When local build and deployed digest differ:** Establish artifact identity before patching source or diagnosing runtime configuration.
+
+## Deliver and verify
+
+- Diagnosis or patch with build/start evidence and remaining environment gaps.
+- Build/runtime boundary, artifact identity and isolated reproduction or verification gaps.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Required runtime files survive multi-stage builds; a non-root process can access only intended paths.
+
+## Stop and recover
+
+- No privileged host mounts or deployment changes by default. Building/running untrusted images requires evaluating their execution effects first.
+
+## Example requests
+
+- **Normal (inspect):** Diagnose why this multi-stage image lacks its runtime files.
+- **edge (inspect):** Diagnose a multi-stage image missing a required runtime file under a non-root user.
+- **blocked (inspect):** Inspect a Dockerfile without building untrusted images or granting privileged mounts.

package/plugins/just-vibe/skills/ops-incident/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: ops-incident
+description: "Organize symptoms, evidence, impact, hypotheses, and immediate actions Use for current operational triage; ops-postmortem reconstructs the completed incident."
+---
+
+# ops-incident
+
+Organize symptoms, evidence, impact, hypotheses, and immediate actions
+
+## Choose this workflow
+
+Use for current operational triage; ops-postmortem reconstructs the completed incident.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [Operations methods](../../references/packs/operations.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; symptoms, affected service/environment, incident window, and known impact.
+
+exact service/environment, time window, revision/configuration identity, authorized logs/metrics, and operational constraints. Prefer observation before intervention; live restarts, traffic changes, restores, and notifications require the requested target/action. Redact sensitive telemetry.
+
+Declared evidence requirements: `telemetry.read`. Use actual host discovery or adequate supplied artifacts; unavailable evidence remains blocked/unknown.
+
+## Scope
+
+Organize triage and recommend immediate actions; remediation follows explicit incident authority.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Establish impact and timeline, separate facts from hypotheses, inspect recent changes and dependencies, prioritize reversible mitigations, and track action/evidence state.
+- Establish impact, time window, affected revision and current changes; keep a timestamped fact/hypothesis/action ledger and prefer reversible mitigations within scope.
+
+## Decision branches
+
+- **When an intervention may erase evidence or duplicate effects:** Capture relevant state and define its observation/abort condition before acting.
+
+## Deliver and verify
+
+- Current incident brief, supported hypotheses, next diagnostic steps, and mitigation options.
+- Impact/timeline, supported hypotheses, mitigation options and observed recovery.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- A coincident deployment is a hypothesis until supported; unknown customer impact is not reported as zero.
+
+## Stop and recover
+
+- No unrequested restarts, failovers, or customer messages. Preserve evidence before interventions that may erase it.
+
+## Example requests
+
+- **Normal (inspect):** Organize this incident's impact, timeline, and next diagnostic actions without restarting services.
+- **edge (inspect):** Triage rising errors after a deployment with an unrelated provider incident.
+- **blocked (inspect):** Inspect supplied incident evidence without restarting services or sending customer messages.

package/plugins/just-vibe/skills/ops-logs/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: ops-logs
+description: "Correlate available logs around a specific failure Use for bounded log correlation; trace follows a particular execution path through source and telemetry."
+---
+
+# ops-logs
+
+Correlate available logs around a specific failure
+
+## Choose this workflow
+
+Use for bounded log correlation; trace follows a particular execution path through source and telemetry.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [Operations methods](../../references/packs/operations.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; service/environment, time window, correlation identifiers, and question.
+
+exact service/environment, time window, revision/configuration identity, authorized logs/metrics, and operational constraints. Prefer observation before intervention; live restarts, traffic changes, restores, and notifications require the requested target/action. Redact sensitive telemetry.
+
+Declared evidence requirements: `telemetry.read`. Use actual host discovery or adequate supplied artifacts; unavailable evidence remains blocked/unknown.
+
+## Scope
+
+Focused log correlation and anomaly explanation, not unrestricted telemetry export.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Normalize timestamps, follow request/job IDs, compare related services, separate repeated symptoms from root events, and redact sensitive fields.
+- Normalize time zones and identify clock skew, follow stable request/job IDs and distinguish original failures from retry cascades and repeated symptoms.
+
+## Decision branches
+
+- **When a relevant span or time range is missing:** Report the gap and query needed; absence is not proof the action never occurred.
+
+## Deliver and verify
+
+- Evidence-linked timeline, likely causal sequence, and gaps requiring metrics/traces.
+- Ordered evidence with timestamps, correlation IDs, causal candidates and redactions.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Clock/time-zone differences are considered; missing logs are not proof that an event never occurred.
+
+## Stop and recover
+
+- Bound query range and volume. Do not dump entire production logs into the conversation or external storage.
+
+## Example requests
+
+- **Normal (inspect):** Correlate sanitized logs for this request ID and time window.
+- **edge (inspect):** Correlate logs across services with different clock offsets.
+- **blocked (inspect):** Analyze a redacted excerpt without querying or dumping full production logs.

package/plugins/just-vibe/skills/ops-observability/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: ops-observability
+description: "Add useful logs, metrics, and traces to unclear execution paths Use to implement requested diagnostic telemetry; ops-alerts turns signals into actionable notifications."
+---
+
+# ops-observability
+
+Add useful logs, metrics, and traces to unclear execution paths
+
+## Choose this workflow
+
+Use to implement requested diagnostic telemetry; ops-alerts turns signals into actionable notifications.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [Operations methods](../../references/packs/operations.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; unclear path, diagnostic goals, telemetry stack, privacy, and overhead constraints.
+
+exact service/environment, time window, revision/configuration identity, authorized logs/metrics, and operational constraints. Prefer observation before intervention; live restarts, traffic changes, restores, and notifications require the requested target/action. Redact sensitive telemetry.
+
+Declared evidence requirements: `project.read`. Use actual host discovery or adequate supplied artifacts; unavailable evidence remains blocked/unknown.
+
+## Scope
+
+Useful logs, metrics, traces, correlation, and error context in the selected path.
+
+Only the requested local changes; external actions require their exact action and target in session authorization.
+
+## Execute
+
+- Identify questions telemetry must answer, choose stable low-cardinality dimensions, propagate correlation, redact data, and verify normal/error instrumentation locally.
+- Start with questions operators must answer, propagate correlation across boundaries and select bounded-cardinality metrics plus redacted structured events.
+
+## Decision branches
+
+- **When proposed labels contain user IDs or arbitrary payload values:** Move needed detail to controlled logs/traces or aggregate dimensions instead of unbounded metric labels.
+
+## Deliver and verify
+
+- Instrumentation changes, field/metric definitions, and validation evidence.
+- Question/signal/location table, privacy/overhead decisions and success/error telemetry checks.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- A failed operation can be traced across the intended boundary; raw user payloads or unbounded IDs do not become unsafe metric labels.
+
+## Stop and recover
+
+- No provider provisioning or production configuration rollout implicitly. Do not add high-volume logging without overhead consideration.
+
+## Example requests
+
+- **Normal (apply):** Add focused tracing without logging raw user payloads or high-cardinality labels.
+- **edge (apply):** Add tracing across a queue while keeping sensitive payloads out of logs.
+- **blocked (inspect):** Design observability without provisioning a vendor or changing production configuration.

package/plugins/just-vibe/skills/ops-postmortem/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: ops-postmortem
+description: "Produce evidence-based timelines and concrete follow-up work Use to reconstruct an actual incident; decision-premortem analyzes hypothetical failure."
+---
+
+# ops-postmortem
+
+Produce evidence-based timelines and concrete follow-up work
+
+## Choose this workflow
+
+Use to reconstruct an actual incident; decision-premortem analyzes hypothetical failure.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [Operations methods](../../references/packs/operations.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; incident evidence, timeline, impacts, actions, and intended audience.
+
+exact service/environment, time window, revision/configuration identity, authorized logs/metrics, and operational constraints. Prefer observation before intervention; live restarts, traffic changes, restores, and notifications require the requested target/action. Redact sensitive telemetry.
+
+Declared evidence requirements: `telemetry.read`. Use actual host discovery or adequate supplied artifacts; unavailable evidence remains blocked/unknown.
+
+## Scope
+
+Evidence-based learning and follow-up design; no blame assignment or external publication.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Reconcile times and observations, distinguish trigger from contributing conditions, document detection/recovery gaps, and define specific preventive/detective actions with measurable outcomes.
+- Reconcile timestamps and impact evidence, separate trigger from contributing conditions and tie each proposed action to a documented detection or recovery gap.
+
+## Decision branches
+
+- **When root cause or impact remains unknown:** Preserve the uncertainty and propose a discriminating follow-up instead of filling the narrative with guesses.
+
+## Deliver and verify
+
+- Postmortem with evidence-linked timeline, impact, causal analysis, and follow-up proposals.
+- Evidence-backed timeline, causal factors, response gaps and measurable follow-up actions.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Unknown root cause remains explicit; each proposed action addresses a documented failure mechanism.
+
+## Stop and recover
+
+- Do not invent owners, impact counts, or consensus. Sending the report or creating external action tickets requires explicit instructions.
+
+## Example requests
+
+- **Normal (plan):** Write an evidence-based postmortem without inventing impact counts or owners.
+- **edge (plan):** Write a postmortem where deployment timing correlates with failure but causation is unproven.
+- **blocked (inspect):** Draft from partial logs without inventing customers affected, owners or consensus.

package/plugins/just-vibe/skills/ops-restore/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: ops-restore
+description: "Prepare or validate backup restoration in an appropriate environment Use for a scoped backup recovery plan or rehearsal; db-migrate changes schema/data intentionally."
+---
+
+# ops-restore
+
+Prepare or validate backup restoration in an appropriate environment
+
+## Choose this workflow
+
+Use for a scoped backup recovery plan or rehearsal; db-migrate changes schema/data intentionally.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [Operations methods](../../references/packs/operations.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; backup identity, source system, isolated destination, recovery objectives, and encryption/access prerequisites.
+
+exact service/environment, time window, revision/configuration identity, authorized logs/metrics, and operational constraints. Prefer observation before intervention; live restarts, traffic changes, restores, and notifications require the requested target/action. Redact sensitive telemetry.
+
+Declared evidence requirements: `project.read`. Use actual host discovery or adequate supplied artifacts; unavailable evidence remains blocked/unknown.
+
+## Scope
+
+Validate restoration and readiness; production replacement requires exact explicit authorization.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Verify backup provenance/completeness, plan target isolation, execute authorized restore, check schema/counts/integrity and application behavior, and record recovery duration/data loss window.
+- Verify backup identity, completeness, keys and destination isolation, then reconcile schema, counts, integrity and application behavior after authorized restoration.
+
+## Decision branches
+
+- **When backup reads successfully but application checks fail:** Treat recovery as incomplete and preserve the isolated target for diagnosis; do not overwrite the live source.
+
+## Deliver and verify
+
+- Restore procedure or exercise report with verified recovery evidence and limitations.
+- Source/target identities, recovery timing/data-loss window and integrity/application results.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- Restore succeeds into the intended isolated destination; a readable backup alone does not count as demonstrated recoverability.
+
+## Stop and recover
+
+- Never overwrite live data implicitly. Missing keys, integrity checks, or target identity stops execution before destructive steps.
+
+## Example requests
+
+- **Normal (plan):** Plan restoring the specified backup into an isolated target, not production.
+- **edge (plan):** Rehearse restore into an isolated database with missing recent transactions.
+- **blocked (inspect):** Plan restore with missing decryption keys or an ambiguous destination; do not execute.

package/plugins/just-vibe/skills/ops-runbook/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: ops-runbook
+description: "Write operational procedures from verified commands and behavior Use to write an operational procedure; ops-incident executes a scoped response."
+---
+
+# ops-runbook
+
+Write operational procedures from verified commands and behavior
+
+## Choose this workflow
+
+Use to write an operational procedure; ops-incident executes a scoped response.
+
+Read [shared execution](../../references/execution.md) for context/mode/authority handling and [Operations methods](../../references/packs/operations.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; operational scenario, environment, existing procedures, and verified commands.
+
+exact service/environment, time window, revision/configuration identity, authorized logs/metrics, and operational constraints. Prefer observation before intervention; live restarts, traffic changes, restores, and notifications require the requested target/action. Redact sensitive telemetry.
+
+Declared evidence requirements: `project.read`. Use actual host discovery or adequate supplied artifacts; unavailable evidence remains blocked/unknown.
+
+## Scope
+
+Diagnosis, intervention, validation, and recovery instructions for a defined event.
+
+None by default. Plan artifacts may be saved when requested.
+
+## Execute
+
+- Inspect actual tooling/configuration, document prerequisites and target checks, order low-risk diagnostics first, mark destructive steps, and specify observable success/abort criteria.
+- Resolve actual environment/tool conventions, order diagnostics before mutation and give each action a target check, expected observation and abort/recovery path.
+
+## Decision branches
+
+- **When a command cannot be exercised safely:** Mark it unverified and state its prerequisites rather than presenting it as rehearsed.
+
+## Deliver and verify
+
+- Runbook with exact contextualized commands, expected outcomes, escalation conditions, and recovery steps.
+- Runnable steps with target checks, expected outputs and escalation/recovery conditions.
+
+Verify these observable conditions when applicable to the actual task; do not claim they were exercised from merely reading this file:
+
+- A responder can distinguish the intended environment; each mutating step has a verification and failure path.
+
+## Stop and recover
+
+- Do not execute the incident procedure merely to write it. Mark commands not exercised in a safe environment as unverified.
+
+## Example requests
+
+- **Normal (plan):** Write a restore runbook with exact target checks and verification steps.
+- **edge (plan):** Write a runbook for restoring queue processing without replaying completed charges.
+- **blocked (inspect):** Draft a runbook without executing incident operations or fabricating terminal output.