thump

A bounded SRE loop for Kubernetes with a reasoning plane bolted into the middle. It watches reliability signals, investigates an evidence snapshot with an LLM, and executes an authored, catalog-bound action once governance clears.

What it can act on is determined by the action catalog. In the shipped profiles at config/<profile>/actions/catalog.yaml, that catalog holds seven actions: three Rook/Ceph runbooks (hold-rebalance, accelerate-recovery, throttle-non-critical-paths), three against the OpenTelemetry Astronomy Shop demo (disable-product-catalog-failure, disable-cart-failure, restart-cart-pod), and one synthetic acme domain action (acme-shed-load). Adding remediation capabilities means adding catalog entries; the reasoner and the governor stay untouched. Onboarding a domain in config alone exercises that boundary.

The safety argument rests on deliberate rigidity. The reasoner cannot invent an action outside the catalog, fabricate parameter magnitudes, bypass the readiness gate on a hunch, or execute anything governance has not approved. A fixed catalog bounds blast radius, governance evaluates policy without re-reasoning, the kill switch fails closed, an undo fires on success as well as failure, and the engine declines explicitly when evidence falls short.

Try it, without a cluster

git clone https://github.com/IanEff/thump && cd thump
task doctor
task ci

Three and a half minutes on a cold clone with an empty build cache, requiring neither Kubernetes nor an ANTHROPIC_API_KEY. task doctor checks required CLI tools upfront. The pipeline executes gofmt checks, go vet, golangci-lint, govulncheck, five strict kubeconform schema passes over the Helm chart and custom resources, promtool unit tests against SLO recording rules, profile configuration audits (task validate:config), the test suite under -race, integration tests, and compiles six binaries.

The test that exercises the full composition:

go test ./test/onboarding -v

That drives all five beats across a domain authored in seven YAML files. The fixture domain is called acme so that any accidental requirement for a domain-specific Go discriminator immediately breaks the test.

The full command list is under Building & testing. Setting up a local cluster is covered further down.

Watch it run

A real incident on the thump-test rig (GCE VMs running k3s): a fault is injected, rattle fingerprints the SLO burn, clank gathers evidence and proposes a catalogued action, hiss rules on it, and thump executes then watches for convergence. The rig is public — github.com/IanEff/thump-test — driven by scenarios in chaos/scenarios.yaml.

Two load-bearing mechanics govern actuation. The reversal operation is bound from the catalog before execution begins, guaranteeing an undo path exists. A missed success window fires that undo automatically. On 2026-07-26 that path ran against Ceph, adjusted two OSD tunables mid-incident, settled success, and restored default configuration.

The same loop is worked step by step against an uncompiled service.

Provenance, stated up front

The four-plane architecture, boundary-object discipline, incident-response loop, and belief-formation defenses are David Jambor's, from Agentic Reliability Engineering (O'Reilly). thump exists because the book left its two central mechanisms unspecified: calibrated confidence and the learning loop were treated as tuning details rather than architectural foundations. Building those two mechanisms is what this repository actually did.

Build methodology comes from John Arundel's The Power of Go: Tools and The Power of Go: Tests (Bitfield). Delivery and layout conventions come from Joel Holmes's Shipping Go (Manning). Where a rule traces to one of them, docs/invariants.md documents the constraint; where the implementation departs from Jambor, docs/design-decisions.md records the rationale.

Evaluating this engine for your own systems starts in Authority model & guardrails and Onboard your own domain. Mechanism-level specifications live in docs/.

---

Table of contents

• Try it, without a cluster
• Watch it run
• Why this shape
• Authority model & guardrails
• The five beats
• Repo tour
• A golden path, worked end to end
• Onboard your own domain
• Install
• Standing it up locally
• Invariants (read as law)
• Known-open
• Building & testing
• Contributing
• Documentation

---

Why this shape

Connecting alert feeds directly to an unconstrained LLM shell creates three failure modes simultaneously.

Observation collapses into interpretation. When an alert arrives stating "system degraded", the model reasons over an upstream conclusion without inspecting the underlying telemetry.

Blast radius remains unbounded. An action space defined by shell access makes containment depend on model judgment, which cannot be audited ahead of time.

Policy becomes invisible. Rules such as freeze windows turn into conditional branches inside prompts or reasoner loops, where they cannot be audited, tested, or tightened independently.

thump enforces structural separation at every boundary. Detection cannot interpret facts. Reasoning cannot mutate infrastructure. Governance evaluates policy without re-reasoning. Execution applies approved contracts without deciding policy.

---

Authority model & guardrails

Remediation magnitudes are declared in catalog configuration (config/<profile>/actions/catalog.yaml). In the shipped catalog, throttle-non-critical-paths specifies a severityReductionPct of 0.7, accelerate-recovery specifies 0.8, and restart-cart-pod specifies 0.1. Restarting the cart pod is a valid, low-blast operation that passes readiness, but the 0.1 forecast separates it from disable-cart-failure (0.9) because the fault is flagd flag state rather than container state. The LLM selects which action fits the diagnosis; it cannot declare that a throttle will reduce severity by 73%. An unforecasted action carries nil, which renders as unmeasured so absence stays distinct from zero effect.

Hypothesis confidence is computed deterministically. The output score is the product of rattle's signal-strength confidence, a grounding multiplier based on live citations across distinct telemetry backends, a case-base alignment term (active once the case base reaches its two-vote floor), and causal likelihood scores from change events. The model's self-reported confidence acts strictly as a ceiling via min(). A high self-score without corroborating evidence is capped; it cannot inflate itself.

Blast tiers, a kill switch, and deduplication windows bound runtime risk. Every action carries an authored blastTier (low/med/high). accelerate-recovery is marked high because trading client I/O for recovery concurrency requires explicit human approval. hiss evaluates tier and reversibility against config/<profile>/hiss/policy.yaml, placing actions exceeding the auto-fire ceiling into a hold state.

Underneath governance sits a file-based kill switch (THUMP_KILLSWITCH, internal/thump/killswitch.go) that fails closed on missing, unreadable, or invalid files. A reload failure clears armed state immediately. Disarmed executions record {mode: live, result: blocked} to ensure visibility. A DedupeWindow (default 1h, configured via DEDUPE_WINDOW) prevents recurring alert firings from launching overlapping actions for the same fingerprint.

Reversal operations are exempt from the kill switch. Halting an active remediation mid-flight risks stranding infrastructure in a partial state, so an approved undo is allowed to complete.

Reversals trigger on two distinct conditions. A missed convergence window initiates an undo. Additionally, an action declaring restoreOnSuccess: true (such as accelerate-recovery or hold-rebalance) reverts its temporary tunables once convergence succeeds, settling the outcome as success. This prevents temporary tuning adjustments from persisting indefinitely.

Declining to act is a first-class result. Emitting no_action with a cited reason (ProposalSet.Status.Reason) is an expected execution path. It triggers when evidence is insufficient, when a proposed action does not match the diagnosed failure class (errClassMismatch), when cited telemetry was never collected, or when the readiness gate vetoes a candidate.

---

The five beats

SignalDetection        ProposalSet             Decision                Outcome
      │                     │                       │                      │
  ┌───▼───┐   detects   ┌───▼───┐   selects     ┌───▼───┐   permits    ┌───▼───┐
  │rattle │────────────▶│ clank │──────────────▶│ hiss  │─────────────▶│ thump │
  └───────┘             └───────┘               └───────┘              └───┬───┘
   Signal                Reasoning              Governance                 │ acts
                                                                           ▼
                                                                    (the cluster)
                                                                           │
                                                     click reads Outcome back into
                                                     clank's case base — the return
                                                     edge, not a sixth box

clank observes three strict boundaries: it trusts rattle's signal and fingerprint without re-evaluating divergence, its entire output is a document, and its candidate authority levels are requests that only hiss can convert into permits.

These boundaries are formalized in CHARTER.md. Detailed component interactions are documented in docs/architecture.md.

---

Repo tour

Implementation entry points and packages for each beat:

Supporting commands in cmd/ include bootstrap (in-cluster setup Job) and calipers, the operator CLI. The thirteen calipers subcommands — incidents, approve, force, unseal, corpus, rca, tune, replay, harvest, probe, transcript, scorecard, validate — are dispatched by verb in internal/calipers/calipers.go. Investigation tools (metrics, loki, kube) live in internal/evidence, LLM provider clients in internal/anthropic and internal/gemini, and coordinate resolution in internal/subjects.

internal/clank/doc.go is the primary reference for reasoning loop invariants and belief-formation defenses (docs/invariants.md, I-6). rattle and hiss handle signal detection and policy evaluation; clank implements bounded LLM investigation.

Test utilities sit alongside their respective packages: natstest (embedded NATS server), s3test, configtest, leaftest (dependency leaf validation), and tlsxtest.

internal/tlsx and internal/tlsxtest centralize TLS configuration. tlsx.go is the sole location where *tls.Config instances are constructed. tlsxtest.go generates ephemeral ECDSA certificate authorities per test, enabling deterministic testing of negative handshake paths (TestClient_ServerLeafFromDifferentCA_HandshakeRefused, TestClient_ExpiredServerLeaf_HandshakeRefused, TestServer_ClientPresentsNoCertificate_HandshakeRefused, TestServer_RotatedKeypair_PickedUpWithoutRestart) without expiring static fixtures.

---

A golden path, worked end to end

The engine operates over arbitrary domains configured via YAML. The walkthrough below traces TestOperator_OnboardsANewDomainInConfigAlone from test/onboarding/onboard_test.go, which runs in task ci without credentials or external infrastructure:

1. rattle detects. acme_api_error_ratio diverges from baseline (observed 0.42 against a 0.001 baseline, severity.DegradationPct: 0.42, trajectory accelerating). rattle assigns fingerprint fp-acme-api-availability-001 and emits a SignalDetection at 0.9 signal confidence.
2. clank reasons. Intake builds the Situational Awareness Object (SAO) with topology from whir/catalog-info.yaml and state queries (acme-db and acme-cache both returning healthy). clank executes the metrics tool for acme_api_error_ratio and the loki tool for namespace: acme, citing both (acme_api_error_ratio and loki-1). It proposes acme-shed-load (blastTier: low, reversal restore-acme-capacity, scope parameter serving_replicas: 2).
3. The gate evaluates. budgetOK, dedupeOK, and evidenceOK all evaluate to true. Two live citations across distinct backends (Prometheus and Loki) satisfy the grounding requirement, producing a computed confidence of 0.9.
4. hiss governs. The tier-1 confidence floor for service_failure is 0.75; 0.9 clears it. The low blast tier maps to RiskBand: act_reversible, which falls within the auto-fire threshold. The decision records Verdict: approved, policyVersion: acme-v1, and floorApplied: 0.75.
5. thump acts. In default dry-run mode, the order is rendered: Outcome{mode: dry_run, result: rendered}. In live mode, thump scales acme-api to 2 replicas, monitors acme_api_error_ratio over the 5-minute window, and executes restore-acme-capacity if convergence fails.

Declines follow the same audit trail. In internal/clank/testdata/detections/ceph-rgw-saturation.yaml, captured JetStream telemetry shows clank identifying a traffic_shift class for which no catalog action is registered. clank proposes nothing, hiss classifies the input as ReasonUngatedInput, and thump declines execution.

Every step maintains provenance by threading signalRef through the pipeline: Detection.Fingerprint → ProposalSet.SignalRef → Decision.SignalRef → Outcome.SignalRef.

---

Onboard your own domain

Onboarding requires no Go modifications. Signals, evidence queries, topology, failure classes, action catalogs, and execution mechanics are authored entirely in YAML. The layout in test/onboarding/testdata/acme/ provides the reference template:

An action's execution block maps to compiled mechanisms in internal/actuate:

- name: acme-shed-load
  applicableFailureClasses: [service_failure]
  applicableTiers: [tier-1]
  blastTier: low
  reversal: {method: restore-acme-capacity, fallback: page-oncall}
  execution:
    forward: [{verb: scale, namespace: acme, deployment: acme-api, replicas: 2}]
    reverse: [{verb: scale, namespace: acme, deployment: acme-api, replicas: 10}]

The compiled verb set — scale, restart, flagVariant, exec, and maintenanceRelease — is validated at process startup. An unmapped verb, missing target parameter, or forward step lacking a reversal fails validation immediately.

⚠️ A catalog PR is an execution-surface PR. The exec verb accepts command arguments directly, allowing anyone who can merge config/<profile>/actions/catalog.yaml to execute commands within pods accessible to thump's ServiceAccount. Access is bounded by Kubernetes RBAC (pods/exec per namespace), the kill switch, and hiss policy. ServiceAccounts must be scoped tightly; see CONTRIBUTING.md.

Configuration details and environment variables are documented in docs/onboarding.md.

---

Install

Release archives are published for Linux and macOS (amd64 and arm64), bundling the four services (rattle, clank, hiss, thump) and calipers. bootstrap is an in-cluster Job that ships exclusively as a container image:

gh release download v0.1.0 --repo IanEff/thump --pattern '*_linux_x86_64.tar.gz'
tar xzf thump_0.1.0_linux_x86_64.tar.gz

Multi-arch container images for rattle, clank, hiss, thump, and bootstrap are published with attached SBOMs and SLSA provenance attestations, signed keylessly via Sigstore/cosign. calipers is distributed as a binary only. Building from source via task build outputs all six binaries to bin/.

---

Standing it up locally

The codebase supports two cluster profiles defined in Tiltfile: dev and thump-test.

dev provisions a self-contained local k3d cluster with Cilium, cert-manager, Prometheus, Alertmanager, Loki, Tempo, S3Mock, the OpenTelemetry demo, and the synthetic acme domain. Detailed instructions are in docs/dev-environment.md.

thump-test runs against GCE VMs hosting k3s with Rook/Ceph and the OpenTelemetry Astronomy Shop demo.

task dev:up                        # local k3d dev cluster
tilt up -- --cluster=thump-test    # remote thump-test rig

Dry-run execution is active by default (THUMP_EXECUTOR=dry). Live execution requires setting THUMP_EXECUTOR=live alongside an armed THUMP_KILLSWITCH file. When the kill switch is disarmed, live attempts emit ResultBlocked.

The operator surface provides inspection and approval workflows:

task dev:certs                      # extract NATS client certificates
task dev:incidents                  # list active and held incidents
task dev:approve FP=<fingerprint>   # approve a held action

Held actions can also be released via Kubernetes custom resources (approvalRequests.enabled: true in deploy/chart/thump/values.yaml). When enabled, hiss creates an ApprovalRequest custom resource that an authorized user approves via standard RBAC:

kubectl -n thump get approvalrequests
kubectl -n thump patch approvalrequest ar-8f3c1d2e0a5b7649 \
  --type=merge -p '{"spec":{"decision":"approve"}}'

Under custom resource approval, a MutatingAdmissionPolicy validates the user identity and records the approver into thump.dev/approved-by, producing an independent Kubernetes audit log entry. The controller accepts only spec.decision: approve. Bypassing governance is reserved for calipers force (D-9).

Environment configuration variables are defined in internal/config/config.go.

---

Invariants (read as law)

Architectural rules are numbered for reference in code reviews and design discussions. Full specifications and test mappings are in docs/invariants.md.

1. Signals describe state, never interpretation. "p99 412ms vs 38ms baseline" is a signal; "system degraded" is an interpretation. rattle emits raw measurements.
2. Two confidence numbers, never one field. Signal-strength confidence belongs to rattle (signal.Divergence.Confidence). Hypothesis confidence belongs to clank (proposal.Candidate.Confidence), computed from evidence corroboration and historical alignment.
3. Policy lives only in Governance. Criticality tiers, error budgets, and confidence thresholds belong exclusively in hiss. clank and rattle contain zero policy rules.
4. The catalog is the autonomy boundary. Blast radius is bounded by declared action contracts, not model judgment. Actions outside the catalog produce immediate errors.
5. Gate ≠ shaper. The readiness gate evaluates a conjunction of minimums (budget ∧ dedup ∧ evidence). Risk shaping evaluates execution autonomy separately.
6. The five belief-formation defenses are not optional. Includes a >=2-source corroboration floor, freshness decay on historical alignment, decrements for predicted-but-absent indicators, representable partial_non_converging outcomes, and mandatory live telemetry citations.
7. Reasoning selects, Governance permits. clank requests an authority level; hiss grants or denies permission without re-ranking or modifying proposals.
8. Learn is a return edge, not a module. click is the return path delivering Outcome records to clank's case base; it is not a standalone service.
9. The signal contract owns the if. Freshness constraints and significance thresholds reside in rattle's contract. Degraded trust attenuates confidence without dropping the signal.
10. Nothing executes ungoverned. Execution defaults to dry-run, requires an armed kill switch, and binds an executable reversal path.
11. The log is the system of record. Detections, proposals, decisions, and outcomes are persisted to an S3 write-ahead log. etcd stores only human-authored configuration.
12. The Trust Ceiling. Autonomous write access requires four operational pillars: runtime governance, automatic reversals, declared signal contracts, and calibrated confidence.
13. Every wave stays red→green. Unverified seams do not cross into subsequent components.
14. Delivery is at-least-once; identity is the fingerprint. Consumers are idempotent and deduplicate using producer-assigned fingerprints rather than transport metadata.
15. The operator surface never disposes. User interfaces read state and emit approval events; they do not write decisions. The sole exception is break-glass calipers force.
16. Every leg is TLS-negotiated by this process, or declared plaintext. TLS configs are constructed exclusively in internal/tlsx. InsecureSkipVerify is disallowed. Plaintext exceptions are explicitly cataloged.
17. Every client this engine dials has an authored retry policy, timeout, and a failure that surfaces. Network clients require explicit timeouts and error propagation.

---

Known-open

• Operator tooling is CLI-focused. calipers incidents, calipers approve, and calipers force provide core hold and approval workflows, but drill-down telemetry views and web interfaces are not built.
• Topology-aware prior adjustment is planned. Actions carry static baseline severityReductionPct values; dynamic adjustment based on situational awareness object topology is not yet implemented.
• Risk shaping uses a two-factor lattice. RiskBand is derived from reversibility and blast tier. A continuous composite risk score is deferred (docs/design-decisions.md, D-3).
• accelerate-recovery requires reconciler coordination. Pausing rook-ceph-operator during recovery requires GitOps tools (e.g., ArgoCD) to ignore replica mutations on that deployment (docs/design-decisions.md, D-10).
• Chaos-mesh v2.8.3 IOChaos limitation. toda injector panics when targeting Ceph OSD FUSE layers, preventing end-to-end chaos testing for that specific fault class.
• ServiceMonitor configuration. Missing scrape configurations can obscure operational metrics; verify Prometheus scrape targets when diagnosing pipeline status.

---

Building & testing

Build automation uses go-task (Taskfile.yaml). Run task --list-all for the full target catalog.

task ci passing is the requirement for completion. Pull requests must pass local task ci before submission.

---

Contributing

Reviews focus primarily on invariant preservation: verifying that policy does not leak into the reasoner, raw payloads remain outside conversation contexts, fingerprints are preserved without recomputation, and vocabulary constraints are respected.

CONTRIBUTING.md provides detailed workflow guidelines. Go conventions, comment formatting, and testing standards are defined in AGENTS.md. Security vulnerabilities must be submitted via SECURITY.md.

---

Documentation

Provenance citations

• David Jambor, Agentic Reliability Engineering (O'Reilly) — Four-plane architecture, boundary objects, incident-response lifecycle, and belief-formation defenses. Sourced vocabulary: SignalDetection, ProposalSet, ActionContract, Situational Awareness Object, and Trust Ceiling. Divergence rationale is documented in docs/design-decisions.md.
• John Arundel, The Power of Go: Tools and The Power of Go: Tests (Bitfield) — Build architecture, functional core/imperative shell separation in actuation, and test naming standards.
• Joel Holmes, Shipping Go (Manning) — Release pipeline, binary packaging, and repository layout.