@aiiware/aii 0.17.27 → 0.18.1

package/README.md

@@ -9,7 +9,7 @@ An autonomous AI assistant that lives in your terminal. Ask it anything — it r
 </p>
 
 <p align="center">
-  <strong>npm:</strong> <code>@aiiware/aii</code> · <strong>Version:</strong> 0.17.4 · <strong>Node.js:</strong> >= 18
+  <strong>npm:</strong> <code>@aiiware/aii</code> · <strong>Version:</strong> 0.18.0 · <strong>Node.js:</strong> >= 20
 </p>
 
 ---
@@ -160,6 +160,7 @@ During interactive sessions, type `/` to see all commands. Highlights:
 | `/theme [name]` | View or switch color theme |
 | `/auto` | Toggle auto-approve mode |
 | `/plan` | Toggle plan mode (read-only) |
+| `/goal [objective]` | Set a session goal the agent works toward autonomously (see [Session Goals](#session-goals)) |
 | `/diff` | See all file changes this session |
 | `/undo` | Revert the last file change |
 | `/stats` | Session stats — tokens, cost, duration |
@@ -271,6 +272,33 @@ Pause takes effect between tool calls — the agent finishes what it's doing, th
 
 The status bar updates in real time with iteration count, token usage, estimated cost, and tool calls — you always know what the loop is doing and what it's costing.
 
+### Session Goals (`/goal`)
+
+Give the interactive session a **persistent objective it keeps working toward until evidence says it's done** — no more typing "keep going" after every turn:
+
+```text
+/goal Make the checkout suite pass, verified by `npm test -w checkout`,
+      without changing public API behavior. --budget 2.00
+```
+
+While a goal is **active**, the agent sees the objective every turn; when a turn ends and you're not typing, Aii **auto-continues** — checking the evidence first, then either keeps working, declares `GOAL COMPLETE — evidence: …`, or declares `GOAL BLOCKED — <what it needs>`. Completion is evidence-based and shown verbatim; the lifecycle stays yours:
+
+| Command | What It Does |
+|---------|-------------|
+| `/goal <objective>` | Set the goal (`--budget <usd>`, `--auto-turns <n>` — `0` = unlimited; `--max-time 2h`, `--max-tokens 5M`; `--verify "cmd"`; `--auto` = no approval prompts) |
+| `/goal` | Status view — objective, status, verify, armed limits, time/tokens used, contextual next commands |
+| `/goal draft <plain description>` | The model proposes a strong goal contract (outcome · verification · constraints); you edit + submit it |
+| `/goal edit <text and/or flags>` | Revise in place (`/goal edit --budget 5` raises a limit without retyping) |
+| `/goal pause` / `resume` / `clear` | Suspend / re-arm / remove |
+
+**Verified completion** — arm `--verify "npm test"` and a `GOAL COMPLETE` claim only counts when that command exits 0: failures keep the goal active and feed the output back to the agent, so "done" means *provably* done. **Cloud-safe limits** — `--max-time` and `--max-tokens` work everywhere (including cloud mode, where `--budget` is unavailable by design).
+
+**Unattended runs**: in Normal mode, approval prompts pause the goal (it never escalates its own authority) — add `--auto` at set time (or type `/auto`) to auto-approve mutating tools for the session and let it run hands-off.
+
+Guardrails by default: auto-continuation caps at **10 turns** per arming (raise with `--auto-turns`, or `--auto-turns 0` for the run-until-done mode — loudly warned without a budget), budgets stop between turns with an honest summary (never claimed as "done"), a continuation that stops making tool calls pauses visibly, typing always wins the turn, Esc pauses the goal, and the `◎ Goal` status-bar segment keeps the state in view the whole time. Goals are session-scoped and survive restarts via `aii agent --continue` (same project only).
+
+`/goal` vs `/loop`: a loop grinds a fixed prompt with fresh context each iteration; a goal keeps one **conversation** converging on an outcome — the evidence trail stays in context. Starting a loop auto-pauses an active goal.
+
 ### Telegram Integration
 
 Run the full agent from Telegram: