# relay-baton v0.9.0 Release Notes

Status: **Released**

Date: 2026-05

Theme: **Automation & Runtime (bounded)** — add *bounded* multi-turn
continuation, session replay, re-planning, and adaptive compression — without
ever introducing an unbounded autopilot, a daemon, or auto-commit.

## Summary

v0.9 turns the Agent Room from a single-shot REPL into a **bounded** runtime.
The new `LoopController` is a pure, model-free guard that caps how many
plan↔execute steps may run and stops on no-progress (divergence) or budget
pressure. The room gains `/continue --max-steps N` (default 3, hard-capped at
20), `/replan` (backs up the current plan first), and `/replay`. A new
`relay-baton replay` command renders the recorded conversation timeline. Per-
agent **adaptive compression thresholds** let larger-context agents run hotter
before compaction. Everything stays confirmation-first and subprocess-only.

## Added

- **Bounded loop controller** — `LoopController` (core `runtime/`). Pure and
  model-free: the caller runs a step, reports an observation
  (`budgetRatio`, `progressKey`), and asks whether to proceed. Stops on
  `max-steps`, `budget`, `divergence` (repeated progress key), or
  `explicit-stop`. It NEVER spawns anything — "continue" is always *bounded*.

- **Session replay** — `ConversationReplay` (core) projects
  `.ai-session/conversation.jsonl` into an ordered, filterable timeline plus a
  deterministic summary (counts by role/kind, first/last timestamps). New
  read-only CLI command `relay-baton replay` with `--json`, `--session`,
  `--kind`, and `--limit`. No model call.

- **Room bounded continue / replan / replay** — new slash commands:
  - `/continue --max-steps N` (or `-n N`, or a bare number) — preview a bounded
    plan↔execute loop, then run it confirmation-first. Each step is gated by
    `LoopController`; the loop stops early when the git working-tree signature
    stops changing.
  - `/replan <task>` — backs up the current `plan.md`, then runs the planner.
  - `/replay` — read-only conversation timeline (no confirmation, no model).

- **Adaptive compression thresholds** — `RelayBatonConfig.contextCompression
  .perAgent?` plus `resolveCompressionThreshold(config, agent)`. Per-agent
  overrides take precedence over the global threshold; values are clamped to
  `[0,1]`. `ContextCompressor.compressIfNeeded` accepts an optional `agent`.

## Changed

- CLI version bumped to `0.9.0`.
- Room help text and command grammar extended (`/continue`, `/replan`,
  `/replay`).

## Tests

- `LoopController.test.ts` — max-steps cap (incl. clamp to ≥1), budget ceiling
  (incl. 0.95 default), divergence on repeated progress key, explicit stop.
- `ConversationReplay.test.ts` — ordered timeline + summary, session/kind
  filters, recent-N limit, one-liner rendering, empty log.
- `CompressionPolicy.test.ts` — default fallback, global vs per-agent
  precedence, clamping of out-of-range values.
- `RoomEngine.test.ts` — `parseMaxSteps` parsing/clamping; `/continue`,
  `/replan`, `/replay` action shapes and confirmation gating.

## Not included (still deferred)

- Non-interactive `run --until` orchestration → candidate for a later release.
- Daemon / long-running prototype → **explicitly out of scope** (roadmap
  candidate only; relay-baton stays a local, on-demand harness).
- Semantic summarization, exact tokenizers, auto commit/push/PR — permanently
  out of scope.

## Upgrade notes

- Fully additive. No config migration required. `contextCompression.perAgent`
  is optional. `/continue` is bounded by design (default 3 steps, max 20) and
  still asks for a single explicit confirmation before any agent runs.

## Verification commands

```bash
pnpm build
pnpm test
relay-baton replay --json   # read-only timeline, no model call
relay-baton chat            # /help, then /continue --max-steps 2 (confirm), /replan, /replay, /exit
relay-baton verify
```

`continue` never runs unbounded: it is capped by `--max-steps` and stops on
no-progress or budget pressure. `replay`/`/replay` make no model calls at all.