dorgu report

Print a local summary of how the loop is doing on this cluster, as text or as JSON (schema dorgu.report/v1)

​

Overview

dorgu report answers “is this working here?” from the cluster’s own records. It reads RemediationAction, IncidentMemory, persona and Event resources, computes the numbers below for a time window, and prints them. It sends nothing anywhere. If you are a design partner, the output is what you paste to the Dorgu team.

It is read-only and needs CLI v0.14.0. The heal and refusal sections need operator v0.13.0 or newer, which is the release that writes the records they are computed from.

It never prints a zero for something it could not measure. A figure with no sample behind it reads no data (null with no_data: true in JSON). A cluster read that fails is named in the output, and every section that depends on it says no data with the reason. A zero in this report is a measurement.

​

Synopsis

dorgu report [flags]

​

Flags

Flag	Type	Default	Description
--since	string	7d	Window to report on: 7d, 24h, 2w, 90m, or any Go duration
-n, --namespace	string	all namespaces	Limit the report to one namespace
--kubeconfig	string	kubectl’s own resolution: $KUBECONFIG, then ~/.kube/config	Path to the kubeconfig
--json	bool	false	Print JSON, schema dorgu.report/v1 (inherited global flag)

​

Examples

dorgu report
dorgu report --since 24h
dorgu report --since 30d --json
dorgu report -n production

​

What it reports

Section	What it tells you
North star: safe heals	Clean heals whose approval-time gate run passed. A heal is clean only when the RemediationAction completed, HealCredited is True, and the incident outcome is resolved.
Heal accounting	How every RemediationAction created in the window ended, and why each one that is not a clean heal is not one: partial, self_resolved, heal_not_credited, no_credit_record, outcome_unconfirmed, rolled_back, approved_not_healed, failed, rejected, expired, acknowledged, refused_by_gate, in_flight. Every action lands in exactly one bucket. approved_not_healed is an approval that never became a heal: it was approved with --no-heal (or never patched), so the workload kept its problem and the operator rolled the persona back. It is not counted as a rollback of a heal.
Active cluster, review depth, approval rate	Whether any proposal was decided in the window, decisions per week, and approved against rejected. Below 5 decisions it prints the counts and says too few decisions instead of a rate. Beyond that, an approval rate under 30% reads as noisy proposals and over 80% as possible rubber-stamping.
Time to heal	p50 and p90 for detect to proposal, proposal to decision (human time), decision to applied, and applied to verified. Detect to proposal is measured once per incident, to its first proposal.
Proposal precision	The share of decided proposals rejected with the reason wrong-diagnosis, and a breakdown of every other reason. See remediation reject.
Refusal mix	What the gate refused, by phase (proposal, approval, apply), stage and reason, with each reason’s share of proposals. A reason over 30% of proposals, on at least 10 proposals, is flagged. An outstanding proposal is shown separately as deferred, because it is waiting for a decision and is not a refusal.
First-run clamps	How many proposals had a value narrowed by a guardrail before you saw the plan, by guardrail.
CLI declines	Times approve or heal declined, by reason. Kubernetes keeps Events for about an hour, so declines older than that are not counted.
Not measurable from a cluster	The measures a cluster cannot report (external installs, design partners, time to first value, retention, and similar), listed so their absence is not read as zero.

The text output labels some lines with short IDs (A2, A5, A6, Q1, Q2, Q3, Q6). They name the measure, so two pasted reports can be compared line by line.

​

Example output

This is real output from the report’s own test fixture, with the cluster name changed. Your numbers will differ. Long lines are shortened with ....

Dorgu report (dorgu.report/v1)
  Window:    2026-11-01T12:00:00Z to 2026-11-08T12:00:00Z (last 7d)
  Cluster:   staging-eks, namespace scope: all
  Operator:  v0.13.0
  Read:      RemediationActions 12, IncidentMemories 9, personas 12, ClusterPersonas 1, Events 7
  Basis:     RemediationActions and IncidentMemories count by when they were created, a proposal-time refusal by when it
             was recorded, Events by their last timestamp. A proposal made before the window and decided inside it is not in it.

North star
  Safe heals: 1  (clean heals whose approval-time gate run passed; clean heals in all: 1)

Heal accounting (12 RemediationActions created in the window)
  completed: 4
  clean heals: 1 (HealCredited=True and the incident outcome is resolved)
  everything else, by why it is not a clean heal:
    partial              1
    self_resolved        2
    rolled_back          1
    failed               1
    rejected             3
    refused_by_gate      2
    in_flight            1
  HealCredited=False by reason: NothingPatched 1, ResourcesNotTicket 1

Tier A: value and adoption
  A2 active cluster:    yes (11 proposals decided in the window)
  A5 review depth:      11.0 decisions per week (11 over 7 days)
  A6 approval rate:     72.7% (8 approved, 3 rejected; inside the 30-80% band)
  A7 external agents:   no data (a RemediationAction records how its plan was written, not who proposed it ...)
  note: A2 and A5 count decisions (approved or rejected). Reading a proposal with `dorgu remediation diff` leaves no record, so a decision is the measurable proxy for a review.
  note: A6 is a band, not a target: under 30% means noisy proposals, over 80% means rubber-stamping.

Tier Q: operational quality
  Q1 time to heal (p50 / p90, n):
    detect -> proposal             1h00m / 1h00m (n=7)
    proposal -> decision (human)   1h00m / 1h00m (n=11)
    decision -> applied            1m00s / 1m00s (n=5)
    applied -> verified            59m00s / 59m00s (n=4)
  Q2 proposal precision: 9.1% of decided proposals were called wrong-diagnosis (1 of 11)
    rejections by reason: no-reason-given 1, too-risky 1, wrong-diagnosis 1
  Q3 LLM efficiency:     no data (LLM calls and cost per incident are operator metrics, not stored on any CRD)
    plan sources: ai-anthropic 1, rule-based 11
  Q6 refusal mix (14 proposals; phase / stage / reason):
    apply     Ticket         TicketExpired              1  (7.1%)
    approval  Guardrails     BlastRadius                1  (7.1%)
    proposal  Ground         WorkloadGone               1  (7.1%)
    proposal  ModePolicy     SelfHealingMode            1  (7.1%)
    not refusals: 1 incident(s) with a proposal waiting for a decision (counted neither as refusals nor as proposals)

First-run clamps (spec.steps[].safety and status.gate)
  proposals with a clamped value: 1
  clamps by guardrail: blast-radius 1
  step verdicts: clamped 1, rejected 1

CLI declines (Events on RemediationActions)
  DorguCLIDeclined           2
  DorguCLINoTicket           3
  DorguCLITicketExpired      1
  DorguCLITicketMismatch     1
  note: The CLI records a decline only as a Kubernetes Event, and Kubernetes keeps Events for about an hour ...

Not measurable from a cluster
  A1   External installs: a cluster cannot tell whether it is ours; read from partner reports
  A3   Design partners: comes from the conversation log
  A4   Time to first value: measured on a clean-room or screen-shared install, not from CRDs
  A8   4-week retention: needs reports from earlier weeks; one cluster's report cannot show it
  A9   Would-pay signals: comes from the conversation log
  A10  Inbound: comes from GitHub
  Q4   Operator footprint: not measured by this report: read the operator's CPU and memory from its own metrics ...
  Q5   Clean-room: comes from clean-room runs

Where there is no sample, a line reads no data (reason). In the example, Q3 has no data because model calls and cost per incident are operator metrics (see the metrics reference) and are not stored on any resource.

​

JSON output

--json prints one document. schema is dorgu.report/v1, and it changes only when a field is removed or its meaning changes. New fields can be added without a new version, so read the fields you need and ignore the rest.

{
  "schema": "dorgu.report/v1",
  "generated_at": "2026-11-08T12:00:00Z",
  "window": {
    "since": "7d",
    "start": "2026-11-01T12:00:00Z",
    "end": "2026-11-08T12:00:00Z"
  },
  "cluster": {
    "context": "staging-eks",
    "namespace": "all",
    "operator_version": "v0.13.0",
    "application_personas": 12,
    "self_healing_modes": [
      "propose"
    ]
  },
  "sources": {
    "remediation_actions": {
      "ok": true,
      "count": 12
    },
    "incidents": {
      "ok": true,
      "count": 9
    },
    "personas": {
      "ok": true,
      "count": 12
    },
    "cluster_personas": {
      "ok": true,
      "count": 1
    },
    "events": {
      "ok": true,
      "count": 7
    }
  },
  "input_errors": [],
  "heal_accounting": {
    "total": 12,
    "completed": 4,
    "clean": 1,
    "safe_heals": 1,
    "no_data": false,
    "excluded": {
      "acknowledged": 0,
      "approved_not_healed": 0,
      "expired": 0,
      "failed": 1,
      "heal_not_credited": 0,
      "in_flight": 1,
      "no_credit_record": 0,
      "outcome_unconfirmed": 0,
      "partial": 1,
      "refused_by_gate": 2,
      "rejected": 3,
      "rolled_back": 1,
      "self_resolved": 2
    },
    "heal_not_credited_by_reason": {
      "NothingPatched": 1,
      "ResourcesNotTicket": 1
    }
  },
  "tier_a": {
    "ns_safe_heals": {
      "safe_heals": 1,
      "clean_heals": 1,
      "no_data": false
    },
    "a2_active_cluster": {
      "active": true,
      "proposals_decided": 11
    },
    "a5_review_depth": {
      "proposals_decided": 11,
      "window_days": 7,
      "per_week": 11,
      "no_data": false
    },
    "a6_approval_rate": {
      "approved": 8,
      "rejected": 3,
      "rate": 0.727,
      "no_data": false,
      "too_few_decisions": false,
      "band": "in"
    },
    "a7_external_agent_share": {
      "no_data": true,
      "reason": "..."
    },
    "decisions": {
      "approved": 8,
      "rejected": 3
    },
    "notes": [
      "..."
    ]
  },
  "tier_q": {
    "q1_time_to_heal": {
      "detect_to_proposal": {
        "n": 7,
        "p50_seconds": 3600,
        "p90_seconds": 3600,
        "no_data": false
      },
      "proposal_to_decision": {
        "n": 11,
        "p50_seconds": 3600,
        "p90_seconds": 3600,
        "no_data": false
      },
      "decision_to_applied": {
        "n": 5,
        "p50_seconds": 60,
        "p90_seconds": 60,
        "no_data": false
      },
      "applied_to_verified": {
        "n": 4,
        "p50_seconds": 3540,
        "p90_seconds": 3540,
        "no_data": false
      }
    },
    "q2_proposal_precision": {
      "decided": 11,
      "wrong": 1,
      "rate": 0.091,
      "no_data": false,
      "too_few_decisions": false,
      "rejections_by_reason": {
        "duplicate": 0,
        "no-reason-given": 1,
        "not-needed": 0,
        "not-now": 0,
        "other": 0,
        "too-risky": 1,
        "wrong-diagnosis": 1
      }
    },
    "q3_llm_efficiency": {
      "no_data": true,
      "reason": "LLM calls and cost per incident are operator metrics (dorgu_llm_requests_total, dorgu_llm_tokens_total), not stored on any CRD",
      "plan_sources": {
        "ai-anthropic": 1,
        "rule-based": 11
      }
    },
    "q6_refusal_mix": {
      "proposals": 14,
      "no_data": false,
      "rows": [
        {
          "phase": "apply",
          "stage": "Ticket",
          "reason": "TicketExpired",
          "count": 1,
          "share_of_proposals": 0.071,
          "flagged": false
        },
        {
          "phase": "approval",
          "stage": "Guardrails",
          "reason": "BlastRadius",
          "count": 1,
          "share_of_proposals": 0.071,
          "flagged": false
        },
        {
          "phase": "proposal",
          "stage": "Ground",
          "reason": "WorkloadGone",
          "count": 1,
          "share_of_proposals": 0.071,
          "flagged": false
        },
        {
          "phase": "proposal",
          "stage": "ModePolicy",
          "reason": "SelfHealingMode",
          "count": 1,
          "share_of_proposals": 0.071,
          "flagged": false
        }
      ],
      "deferred_proposals": 1
    }
  },
  "first_run_clamps": {
    "proposals_with_clamp": 1,
    "step_verdicts": {
      "clamped": 1,
      "rejected": 1
    },
    "clamps_by_rule": {
      "blast-radius": 1
    },
    "gate_clamps_by_guardrail": {},
    "no_data": false
  },
  "cli_declines": {
    "no_data": false,
    "by_reason": {
      "DorguCLIDeclined": 2,
      "DorguCLINoTicket": 3,
      "DorguCLITicketExpired": 1,
      "DorguCLITicketMismatch": 1
    },
    "note": "..."
  },
  "not_measurable": [
    {
      "id": "A1",
      "name": "External installs",
      "reason": "a cluster cannot tell whether it is ours; read from partner reports"
    },
    {
      "id": "A3",
      "name": "Design partners",
      "reason": "comes from the conversation log"
    }
  ]
}

The example shortens the long text fields and lists only two of the not_measurable entries. These are the top-level fields and what they hold:

Field	Holds
schema, generated_at, window	The schema name, when the report ran, and the window as since, start and end
cluster	Context only: no number is computed from it. application_personas is null, never 0, when the personas could not be read, and self_healing_modes is null or empty when the ClusterPersonas could not be read
sources	For each of remediation_actions, incidents, personas, cluster_personas and events: ok, count, and an error class when a read failed
input_errors	One {input, error} entry per failed read. error is one of forbidden, not_found, timeout, unparseable, error
heal_accounting	total, completed, clean and safe_heals (both null when they cannot be known), no_data with reason, excluded (a count per bucket, zeros included), and heal_not_credited_by_reason
tier_a	ns_safe_heals (safe_heals, clean_heals, no_data, reason), a2_active_cluster (active, proposals_decided), a5_review_depth (per_week, window_days), a6_approval_rate (approved, rejected, rate, too_few_decisions, band), a7_external_agent_share, decisions and notes
tier_q	q1_time_to_heal (detect_to_proposal, proposal_to_decision, decision_to_applied, applied_to_verified, each n, p50_seconds, p90_seconds, no_data), q2_proposal_precision (decided, wrong, rate, rejections_by_reason), q3_llm_efficiency (plan_sources), q6_refusal_mix (proposals, rows[] of phase, stage, reason, count, share_of_proposals, flagged, and deferred_proposals)
first_run_clamps	proposals_with_clamp, step_verdicts, clamps_by_rule, gate_clamps_by_guardrail, no_data
cli_declines	by_reason (event reason to count), no_data, and a note on the one-hour Event retention
not_measurable	{id, name, reason} entries

Check no_data before you read a number. clean, safe_heals, a percentile or a rate is null when it cannot be known, and no_data is true with a reason. A script that treats null as 0 will report a failure to read as a fact.

​

The operator line in the header

The header prints the operator version only when the operator’s image tag is a release version. A commit hash, latest or dev reads Operator: unreleased build <tag>, so a build identifier is never shown as if it were a version. cluster.operator_version in the JSON keeps the raw tag.

​

Older operators

Against an operator older than v0.13.0, which writes no tickets and no HealCredited record, the heal sections are no data with the operator version in the reason. The same is true when the operator version cannot be determined (for example an image pinned by digest, or a tag such as latest) and the window has completed actions with no HealCredited record. The report fails toward no data, not toward a measured zero.

​

Permissions

The report reads, with the identity in your kubeconfig, remediationactions, incidentmemories, applicationpersonas, clusterpersonas, Events about RemediationActions, and Deployments (to read the operator’s version). If a read is forbidden, it is listed in input_errors and the sections that depend on it say no data. The command needs kubectl on your PATH, like every dorgu command that talks to the cluster.

​

Related

How Dorgu decides

The pipeline these numbers describe

dorgu remediation

reject --reason feeds proposal precision

Metrics reference

The same events as Prometheus metrics