Diagnose

HA NOVA Diagnose

Scope

Root-cause a concrete failure: "why did X not run", "why did X misbehave", "what broke last night".

• Evidence sources: automation/script traces, HA error log, system log, bounded logbook/history windows, template probes, integration diagnostics.
• Boundary: ha-nova:health reports CURRENT home status (repairs, unavailable entities); ha-nova:history answers plain timeline questions; ha-nova:review audits config quality without a concrete incident. This skill starts from a concrete symptom.
• The ONLY mutation in this skill is a temporary logger.set_level debug escalation (see Flow step 5) — every other fix hands off to the owning skill (ha-nova:write, ha-nova:helper, ha-nova:service-call, ha-nova:integration-setup for credential recovery).

Bootstrap (once per session)

Read and follow ../ha-nova/session-bootstrap.md. Verify relay CLI: ha-nova relay health If this fails: ha-nova setup

Relay Contract

File-based relay requests only:

• ha-nova relay core --method GET --path <PATH> for REST reads (--out <result-file> for large output)
• ha-nova relay core --method POST --path <PATH> --body-file <payload-file> for REST calls with a body (template probes, service calls)
• ha-nova relay ws --data-file <payload-file> for WS commands
• ha-nova trace latest <entity_id> --json / ha-nova trace list <entity_id> --json / ha-nova trace get <entity_id> <run_id> --json for run traces (positional entity, automation.<id> or script.<id>)
• --jq-file <filter-file> for non-trivial filters

Evidence Sources

Source	Call	Notes
Run traces	ha-nova trace latest <entity_id> --json (also trace list <entity_id>, trace get <entity_id> <run_id>)	First stop for automation/script failures: shows the exact step, condition results, and variables
System log	WS {"type":"system_log/list"}	PRIMARY log source: structured warnings/errors with counts and sources, works on every install type
Error log	GET /api/error_log	404 on HA OS/Supervised since 2025.11 (the log file moved to journald) — that is normal, not an error; stay with system_log/list. Where the file exists (Container/Core, or re-enabled), see Reading the error log
Logbook window	GET /api/logbook/<ISO-start>?end_time=<ISO-end>&entity=<entity_id>	Human-readable events around the incident time
State history	GET /api/history/period/<ISO-start>?end_time=<ISO-end>&filter_entity_id=<ids>	What states the involved entities actually had
Template probe	ha-nova relay core --method POST --path /api/template --body-file <payload-file> with {"template":"{{ ... }}"}	Evaluate the exact condition/template against live state; the rendered result is the string in .data.body
Integration diagnostics	WS {"type":"diagnostics/list"}, then GET /api/diagnostics/config_entry/<entry_id> with --out <result-file>	Deep integration state when an integration itself is the suspect
Device diagnostics	GET /api/diagnostics/config_entry/<entry_id>/device/<device_id> with --out <result-file>	One device's slice when a single device misbehaves; resolve the (entry_id, device_id) pair via WS diagnostics/list (only entries with a device-diagnostics handler) — a device can carry several entry ids

Diagnostics redaction is integration-controlled, so downloads are potentially sensitive and NEVER hit bare stdout — save with --out, extract only relevant fields via relay jq before quoting.

Recency honesty: error_log and system_log/list only cover the time since the last Core restart. For older incidents use logbook/history windows, and say explicitly that live logs no longer reach back that far.

Reading the error log

Only where the log file exists (Container/Core installs; on HA OS/Supervised the user can restore it with ha core options --duplicate-log-file=true followed by ha core rebuild and ha core restart, since 2026.1). The upstream body is plain text, but the relay wraps it in the /core envelope — the whole log is one escaped string in .data.body, so searching the saved file line-wise does not work.

1. ha-nova relay core --method GET --path /api/error_log --out <envelope-file>

2. Filter the lines you need with a raw jq read (never dump the whole log):

   ha-nova relay jq -r --file <envelope-file> --jq-file <filter-file>
   

   with <filter-file>:

   .data.body | split("\n")[] | select(test("<entity_or_integration>"; "i"))
   

   -r is required for raw text output — relay core --jq does not support it, so the filtering happens in this second step. Substitute only an entity_id, domain, or integration slug — never a friendly name, which can contain (, ) or | and would break the filter. Even a slug is not literal: the . in sensor.kitchen is a regex wildcard, so escape it (sensor\\.kitchen) or the filter also matches sensor_kitchen and reports unrelated lines as evidence.

   The identifier is the ONLY selector — never OR a severity into it. test("sensor\\.kitchen|ERROR") matches every error line in the log, from every unrelated integration, and each one then looks like evidence for this incident. To narrow by severity, AND it: select(test("sensor\\.kitchen"; "i") and test("ERROR|WARNING"; "i")). And a log with no matching line is a finding in itself — say the log holds nothing about this entity, rather than widening the filter until something appears.

Flow

1. Pin the symptom: which entity/automation/script, what expected vs observed behavior, and WHEN (ask one question if the incident time is unknown — every later window depends on it).
2. For automation/script symptoms, read the trace first (ha-nova trace latest <entity_id> --json; older runs via trace list <entity_id> + trace get <entity_id> <run_id>). The trace usually answers it: triggered or not, which condition stopped it, which step errored.
3. Correlate logs: search system_log/list (and the error log where it exists) for the involved entities, integrations, and the incident window. Quote only the relevant lines. Align timestamps before claiming causation: the log line must sit within minutes of the trace's failing step — an error far outside that window is context, not the cause; say which it is.
4. Reconstruct context: bounded logbook + history windows (default ±30 minutes around the incident) for the involved entities; probe suspect conditions/templates against live state (relay core --method POST --path /api/template --body-file <payload-file>, body {"template":"{{ ... }}"}; the rendered value comes back in .data.body).
5. Debug escalation (optional, the only mutation): if evidence is insufficient and the user agrees, raise one integration's log level.
   • FIRST read the current levels: WS {"type":"logger/log_info"} → [{"domain":..., "level": <numeric>}]. logger/log_info reports NUMBERS but logger.set_level only accepts NAMES — map the recorded number back before building any payload: 10 → debug, 20 → info, 30 → warning, 40 → error, 50 → critical. A numeric level in a set_level payload is rejected by Home Assistant, which would leave debug logging raised until the next restart. Record the target logger's mapped level — that is the state you must restore. If it has no entry, the effective level is the configured default (HA's own default is warning, but configuration.yaml may set another).
   • Escalate: service logger.set_level with {"homeassistant.components.<domain>":"debug"}. Preview the exact payload and get confirmation.
   • ALWAYS pair it with the reset: after reproducing, restore the recorded level as its NAME (mapped above) — never the raw number, and never a hard-coded warning, which would silently overwrite a level the user set on purpose. Home Assistant has no "remove override" service: an override persists until the next restart, so restoring means setting the recorded level back explicitly (or restarting HA for a truly clean state). Say this plainly before escalating.
6. Conclude with Claim-Evidence Binding: name the root cause only when the evidence shows it (trace step, log line, state sequence). Otherwise present the ranked hypotheses, each with its evidence and the one probe that would decide it.
7. Hand off the fix: config changes -> ha-nova:write / ha-nova:helper; a one-off corrective action -> ha-nova:service-call; invalid credentials or a pending reauthentication -> ha-nova:integration-setup; systemic issues (repairs, unavailable entities) -> ha-nova:health.

Error Handling

Full relay/upstream error taxonomy: skills/ha-nova/relay-api.md -> Error Handling. Diagnose specifics:

• /api/error_log 404: normal on HA OS/Supervised since 2025.11 — continue with system_log/list and mention the --duplicate-log-file re-enable only if the user asks for the raw file.
• Missing traces (empty trace list): traces are kept per entity with a small cap and reset on reload/restart — say so instead of guessing.
• system_log/list empty after a restart is normal; fall back to logbook/history windows.

Output Format

Apply skills/ha-nova/output-rules.md to all user-facing output. Write previews, delete confirmations, and results render as the Cards defined there.

Render the Report shape, Diagnosis specialization (output-rules.md → Report Shape): root cause first, then the evidence chain, then the recommended fix with the owning skill. Quote log lines sparingly — never dump raw logs.

Safety

• Preview before write: nothing is saved until the user confirms the shown preview.

• Confirmation binds to the displayed preview and expires on any change to target, payload, endpoint, or scope (context skill → Active Preview Confirmation).

• Pre-preview phrases ("do it", "go ahead", "implement the plan") authorize drafting and preview only — never the write itself.

• Delete and destructive operations require the typed confirmation code confirm:<token> verbatim; "yes" or any natural-language reply is invalid.

• Never guess entity, service, or config IDs — resolve them or ask.

• Home Assistant is reached exclusively through ha-nova relay.

• For any HA write this skill does not cover, STOP and invoke ha-nova:fallback first — never probe unfamiliar write endpoints.

• Drafts follow skills/ha-nova/smallest-solution.md: the complete requested outcome in the simplest safe design, nothing for hypothetical future needs.

• The debug escalation (logger.set_level) is this skill's single declared mutation: read the current level via logger/log_info first, preview + natural confirmation, and always schedule the restore of THAT recorded level (never a hard-coded default) in the same interaction.

• Never "fix" anything from here — fixes hand off to the owning skill after the diagnosis.

Guardrails

• Bounded reads only: logbook/history windows default to ±30 minutes; widen stepwise on request, never unbounded.
• Search the error log natively from the saved file; never print more than the relevant lines.
• Do not run diagnostics proactively — this skill starts from a user-reported symptom.
• Never leave a raised log level behind: no recorded pre-level and no restore plan means no escalation.