@adonis-agora/agent 0.83.0 → 0.85.0

package/CHANGELOG.md

@@ -1,5 +1,26 @@
 # @adonis-agora/agent
 
+## 0.85.0
+
+### Minor Changes
+
+- [#453](https://github.com/DavideCarvalho/adonis-agora-agent/pull/453) [`3f8937e`](https://github.com/DavideCarvalho/adonis-agora-agent/commit/3f8937e22ace81cace1b84c7f26706947a9ce66b) Thanks [@DavideCarvalho](https://github.com/DavideCarvalho)! - Text channels: bursts of messages become one turn, and durable jobs run their phases as dispatched steps.
+  
+  - **`coalesce` (opt-in).** `channels.handle(adapter, { coalesce: { quietMs, maxWaitMs?, maxMessages?, when? } })` groups the messages a person sends in a row — "oi" / "tudo bem?" / "queria saber X", or ten PDFs — into ONE turn once the conversation has been quiet for `quietMs` (at most `maxWaitMs`, default 30 s; at most `maxMessages`, default 20): texts joined in order, every file attached, each message's files prepared and refused as if alone. Button presses, text decisions and answers to a question are never grouped and end a burst; messages that arrive while a grouped turn runs go to the next one. Under a durable engine the wait is the engine's durable sleep (resumable on any instance); inline, a timer. Every message keeps its own job and its own dedupe; the first to run leads and claims the others in the channel store, so nothing is answered twice or dropped across retries, recoveries and redelivered webhooks. Hooks see the burst through new optional fields: `ChannelHookContext.coalesced`, `ChannelInbound.parts`, and `coalesced` / `attachments` / `queueWaitMs` on `onTurnStarted` and `onTurnEnded` (set only with `coalesce` on). Off by default: nothing changes for a channel without it.
+  - **Dispatched phases.** The `agora.channel.job` / `agora.channel.timer` workflows are now at version `2`: each phase is a dispatched step (`agora.channel.prepare`, `agora.channel.act`, `agora.channel.read`, `agora.channel.gather`, `agora.channel.timeout`) served on the durable engine's transport by every process that created the channel's handler, so any process runs a phase and any process continues the job. Each carries a liveness timeout (a lost step is dispatched once more, then fails the job instead of hanging). The agent's engine is served on its container transport automatically; an engine passed as `durable: engine` takes `durableSteps: transport` (or an engine `handleStep`); `durableSteps: false` keeps the phases as local steps. Version `1` stays registered, so runs in flight during the upgrade finish on their old path. `ChannelWorkflowEngine` gains optional `handleStep`, `ChannelWorkflowCtx` optional `step` / `sleepUntil` / `sleep`; `ChannelStepServer`, `ChannelStepLog`, `ChannelCoalesceOptions` and `ChannelInboundPart` are exported.
+
+## 0.84.0
+
+### Minor Changes
+
+- [#440](https://github.com/DavideCarvalho/adonis-agora-agent/pull/440) [`c113c76`](https://github.com/DavideCarvalho/adonis-agora-agent/commit/c113c76a890b076caab8bc143c7d2a10d570b7be) Thanks [@DavideCarvalho](https://github.com/DavideCarvalho)! - A run's time to first token, failure code and model are stored, and the governance dashboard and the Telescope tab chart them:
+  
+  - **Stored on the run.** `agent_run` gains nullable `ttft_ms`, `error_code` and `model_id`, added by `createAgentTables`' schema repair (run at boot; with `autoCreateTables: false`, re-run the `create_agent_tables` migration). `RecordRunEndInput` takes `ttftMs`, `errorCode` and `modelId`. A completed run records its time to first token and last model; a failed one records the same stable code its stream `error` frame carries (`quota_exceeded`, `model_no_output`, `run_failed`…), next to the `error` text as before. A run found dead, or failed by OpenCode, is `run_failed`. `RunSummaryRow` carries the three as optional fields.
+  - **Tool timing.** A tool call's `execution_ms` is now filled, measured inside its `tool:<id>` checkpoint: how long the tool ran, not how long a replay took to read it back. A replayed call writes none.
+  - **Read-model.** `RunReliability` gains optional `latency: { ttft, duration }` (`{ samples, p50Ms, p95Ms }`, nearest-rank, over completed turns a person started; a delegation's child run is left out) and `failuresByCode` (`{ code, count }[]`, most frequent first; a failure from before codes is `unknown`). `PerToolStatRow` gains optional `p50ExecutionMs` / `p95ExecutionMs`. `summarizeLatency` / `summarizeRunMetrics` are exported; `GenuiLatency` is now an alias of the shared `Latency`.
+  - **Dashboard.** Reliability shows time to first token and turn duration (p50 / p95) and a failures-by-code donut and table; Tools shows p50 / p95 per tool.
+  - **Telescope.** A new "Run latency" section: TTFT and turn-duration distributions (p50 / p95 / p99) from `run.finished`, and failures by code from `run.failed` (`agent.latency`, `agent.failuresByCode`). Active runs now subtract `run.failed` and `run.cancelled` too, and the tool-call counters skip a call re-published on replay.
+
 ## 0.83.0
 
 ### Minor Changes