chaos-study-run — the only phase that changes Azure resources

Everything before this is analysis. This skill executes a Chaos Studio scenario against your Azure resources, so it is deliberately the most conservative script in the suite.

Principles

Dry run is the default. -DryRun defaults to $true. A dry run walks the entire sequence — plan verification, the configuration it would create, the consent phrase — and executes nothing. Read it before arming anything.

Consent is typed, not flagged. Arming requires -DryRun:$false and -Consent matching the phrase the plan prints, exactly. A wrong phrase exits 11. There is no -Yes, no -Force that bypasses it, and no environment variable that pre-approves it.

Granting access is a second, separate consent. Injecting a fault is reversible and bounded; a role assignment is neither. So -ApprovePermissions takes its own phrase, derived from the grants the service actually previewed, and one consent can never stand in for the other.

The plan is verified, not trusted. The plan hash is recomputed before execution. Any drift since scoping exits 12. If the blast radius changed after you approved it, this refuses rather than running something you did not review.

Validation is a gate, not advice. The scenario configuration is created and validated before any evidence is collected or anything is injected, and the run is refused unless the service reports Succeeded. If validation fails, the missing role assignments are previewed with --what-if and shown — and then the study stops. Granting them is not covered by the injection consent: a role assignment outlives the study and changes who can reach those resources afterwards, whereas the injection ends when the run ends. So it takes its own phrase, bound to the exact grant set the service previewed. Neither consent implies the other, and a preview that widens invalidates the phrase.

Evidence is collected around the run, not just during it. Baseline, during, and post windows are all collected. Without baseline there is no "normal" to compare to; without post there is no way to tell degradation from damage.

Missing evidence stays missing. A collector that fails records null and a reason. Nothing is interpolated, averaged, or inferred to fill a gap — the report would rather say not measured than mislead.

Cleanup is unconditional, and removal is observed rather than assumed. The scenario configuration is deleted in a finally block, so a crash, a Ctrl-C, or a failed collection still tears it down. A run still in flight is cancelled. The delete returning 0 only means the control plane accepted the request, so each removal is followed by a read-back: only an observed absence is reported as removed, and anything still present stays in the residue ledger with the exact command to finish the job. Use -KeepConfiguration only when you intend to inspect it afterwards.

Sealed studies are immutable. Re-running a sealed study exits 13. Re-test by scoping a new one, so history stays comparable.

Usage

Preview (this is the default, and executes nothing):

./scripts/Invoke-ChaosStudyRun.ps1 -StudyId latest

Arm it, using the phrase the preview printed:

./scripts/Invoke-ChaosStudyRun.ps1 -StudyId latest -DryRun:$false `
    -Consent '<phrase from the preview>'

Switches worth knowing:

• -SignalSource — add evidence sources beyond the plan's (metrics: / logs:)
• -PollSeconds — scenario run status poll interval
• -KeepConfiguration — leave the configuration for inspection

What actually happens

1. Load the plan and re-verify its hash
2. Render the consent phrase and require it back verbatim
3. Create the scenario configuration on the workspace, carrying the plan's frozen parameters, filters and exclusions
4. Validate it; on failure, preview the missing grants with --what-if and stop with exit 19 unless -ApprovePermissions '<phrase>' matches that preview — only then apply the grants, record them, and validate again
5. Refuse to continue unless validation now says Succeeded
6. Collect the baseline window
7. Execute the configuration; poll the scenario run until it reaches a terminal state or the injection window closes
8. Collect the during window while the run is live
9. Wait out recovery, then collect the post window
10. Cancel any in-flight run and delete the configuration (always), then write run-record.v3.json and evidence

Steps 3–5 run before step 6 deliberately. A configuration that cannot validate is a run that fails within seconds, and finding that out after a baseline window has already elapsed wastes the window and the study. Start-ChaosStudyScenarioRun takes the validation result as a required argument and asserts it again as the last thing before run start, so the guarantee belongs to the function that starts the run rather than to the order of statements in one caller.

The configuration is built from what the plan captured live at scope time: the scenario, its parameters as the {key,value} pairs the service takes, and the blast radius as filters and exclusions. An empty blast-radius member is omitted rather than sent — {"locations":[]} means no locations, which would silently turn a real study into a no-op that still reports success.

Implementation note

Execution uses Chaos Studio V2 exclusively: workspaces are the lifecycle root, resource selection comes from workspace scopes and the resources the service discovered inside them, and the run is a scenario run against a validated scenario configuration. There is no V1 path and no fallback to one. Api-versions are pinned centrally so a service-side default change cannot silently alter behaviour.

Exit codes

18 is a pause, not a failure, and it is the one exit where the temptation to improvise is most dangerous. Under -Adapter external this phase cannot call Azure itself, so it wrote the exact request — including the scenario run start or status poll — into the study's operations/ directory. Execute that call with your own authenticated tooling, write the response back as the matching result file, and re-run the identical command. The run resumes from where it stopped; consent and the frozen plan hash still hold, so nothing is re-armed and no fault is injected twice.

Do not fabricate a run outcome or hand-write evidence to get past this. Results are hashed against their request and recorded in the study's provenance, and a report built on unprovenanced evidence cannot be sealed as compliant.

If it fails partway

The study state on disk reflects exactly how far it got, and whatever evidence was collected is kept. Because reporting is a pure read of that evidence, a partial run can still be reported — the report will simply carry the gaps as explicit limitations rather than hiding them.

Next

chaos-study-report -StudyId <studyId>

See ../chaos-study/references/study-method.md for how the windows are chosen. Per-action blast radius is not documented locally: it comes from the live action metadata frozen into the plan, which the dry run prints in full.