Tags: nowa/ri

ri v0.81.4

Two production-reported fixes for long-running agents behind a
concurrency-limited gateway (longbridge#7 and #8):

- The event log returned by agent_loop_* no longer retains per-chunk
  progress events, which grew it quadratically with streamed output
  (hundreds of MB of RSS per long round, discarded unread).
- Streaming errors carry the gateway's announced retry-after so retries
  land on a free slot instead of backing off blindly, and a dropped
  stream consumer aborts the request so the upstream concurrency slot is
  released immediately.

Pi baseline unchanged (v0.81.1 @ dd6bea41).
Workspace tests: 1552 passed / 0 failed.

ri v0.81.3

Behavior-parity audit (constants, retry/error handling, streaming edges,
plus a live payload differential against pi), cross-implementation
interop testing against pi-written files, and the three feature gaps the
audit deferred: Copilot entitlement filtering, Bedrock/Vertex interactive
logins, and SDK-contract maxRetries.

Pi baseline unchanged (v0.81.1 @ dd6bea41).
Workspace tests: 1547 passed / 0 failed.

ri v0.81.2

Patch release: prompt-text parity with pi — typebox-exact validation
error text, pi abort checkpoints, Responses empty-text replay,
summarization truthiness and budget fallbacks, skill/template name
edges. Pi baseline unchanged (v0.81.1 @ dd6bea41).
Workspace tests: 1471 passed / 0 failed.

ri v0.81.1

Patch release: recover truncated tool-call arguments with
streaming-JSON parsing on the OpenAI Completions and Responses paths.
Pi baseline unchanged (v0.81.1 @ dd6bea41).
Workspace tests: 1470 passed / 0 failed.

ri v0.81.0 — initial release

Rust port of pi's LLM provider and agent runtime.
Pi baseline: v0.81.1 @ dd6bea41 (audited case-by-case, 1179 cases).
Workspace tests: 1469 passed / 0 failed.

Snapshot before the supplementary migration that produced v0.81.0

Last state prior to the large follow-up port: remaining six-stage plan
items, interactive OAuth, ri-owned catalog generator, SQLite lazy-read
architecture, PR #4 merge, the case-by-case pi test-parity audit
(1179 cases), and gap-closure waves 1+2.