AgentRun events.jsonl records type=usage_recorded which is missing from AGENT_RUN_EVENT_TYPES → strict decode fails, desktop app won't start

[songlairui]

Summary

The Maka Desktop app (v0.1.3, macOS) fails to start with:

AgentRun 99272103-6b74-4dd2-b367-b231bbe0b6c7 has a corrupt JSONL record at line 11: Invalid AgentRun event schema

Root cause

The run event ledger at workspaces/<ws>/sessions/<session>/runs/<run>/events.jsonl contains events whose type is usage_recorded, but usage_recorded is not in the AGENT_RUN_EVENT_TYPES enum in @maka/core (src/agent-run.ts). The strict read path (readEvents with strict=true in @maka/storage agent-run-store.ts) throws on any event whose type is not in the enum, which blocks app startup.

Verified locally:

• Every event in the ledger is valid JSON — the failure is purely schema-level.
• All 4 runs on this machine contain a usage_recorded event (at line 11 or 21), i.e. every run that completed a turn with token usage writes it. This means the app fails to start again after any new run — not a one-off corruption.
• The string usage_recorded / "Token usage recorded" does not appear anywhere in the installed app bundle (app.asar + app.asar.unpacked + workers/), so the writer (likely the AI-SDK runtime or a recently updated agent runtime) and the validator (compiled schema enum) are out of sync.

Event shape that fails

Line 11 of the affected run (valid JSON, rejected by schema):

{"type":"usage_recorded","id":"ac9072a7-...","runId":"99272103-...","sessionId":"78d21d15-...","turnId":"300c9dab-...","ts":1785736878682,"message":"Token usage recorded","data":{...inputTokens/outputTokens/cacheHitInputTokens/cacheMissInputTokens/cacheMissInputSource/reasoningTokens/totalTokens/rawFinishReason/systemPromptHash/prefixHash/requestShapeHash/toolAvailability...}}

The data payload is the same token-usage accounting that runtime-events.jsonl records as a tokenUsage action on a system message (that file passes validation fine).

Impact

• App is bricked at startup once any run has written a usage_recorded event (i.e. virtually every real run).
• The event_corrupt tolerance path exists for non-strict reads, but startup uses strict mode and hard-fails.

Workaround (local)

Removing the usage_recorded line from each events.jsonl lets the app start again; remaining events form a valid ledger (run_created → … → run_completed).

Suggested fix

• Add usage_recorded to AGENT_RUN_EVENT_TYPES (and its projection handling), or
• Align the runtime writer so it emits an already-supported event (e.g. model_call_attempt_recorded), and/or
• Make the startup path non-strict / quarantine unknown event types as event_corrupt instead of aborting.

Possibly related: #1679 (model-call accounting RFC) and #1418 (one undecodable file breaking startup).