@lawaty/lacode 2.1.0-beta.0 → 2.3.0-beta.0

package/CHANGELOG.md

@@ -9,6 +9,48 @@ commit as the changelog entry it describes.
 
 ## [Unreleased]
 
+## [2.3.0-beta.0] — reliable pool refresh, autonomy modes, and empty-response failover
+
+### Added
+
+- Modes: YOLO (`/yolo`) approves permission requests that would ask, while an explicit `deny` still wins; YKB (`/ykb`) tells the agent not to ask and answers questions with the `(Recommended)` option (or the first option when none is marked); `/modes` shows both. The `yolo`/`ykb` options only seed the live file, and toggles apply without a restart. (US-37, US-38)
+- Pool: an empty task result is now a failure. The model is benched for `empty response`, the task result explains the bench, and the next spawn routes to another slot. This fixes the “routed only to one slot, and it returns nothing” case. (US-4)
+- Pool: the refresh now runs every 6 hours and also after a startup, so a run killed by a reboot is retried. A `--budget-min` wall-clock budget (default 30) writes what was measured, core providers are scanned first, and unscanned or truncated providers keep their previous rows. (US-39)
+- Pool: one shared failure classifier (`credit`, `rate-limited`, `agreement`, `credential`, `not-served`, `timeout`, `network`, `server`) is persisted as `status`/`kind`. Transient rows with no reset get a 45-minute retry, and per-provider warnings name the dominant failure.
+- Pool: a refresh lock serializes the 6-hour run and transient retry timers; a lock left by a dead PID is reclaimed.
+- Pool: a slot can carry a tokens-per-minute budget (`tpm:`), and the router keeps a learned TPM record; a sub-threshold slot is dropped with a warning instead of being sent a task it cannot finish.
+- Config: `contextAutoUpdate: false` disables the automatic context-map refresh.
+- TUI: `/yolo`, `/ykb`, and `/modes` commands.
+
+### Changed
+
+- Pool: the last slot is now a reserved `opencode` (Zen) slot at weight 2, so a full field of other providers cannot squeeze Zen out.
+- Pool: a transient failure keeps the last measured rankable row in `model-scores.json` instead of erasing it on the next write.
+- Pool: routing never falls back to a cooling or unavailable slot. When no slot is usable, the spawn fails with a message naming every cooling slot and its reason; a reaped model refuses routing until its cooldown lapses.
+- Timer: the service has no start timeout, and the timer has a 5-minute startup catch-up trigger in addition to the 6-hour calendar.
+- Event fan-out: a failing plugin event handler is logged once per minute with its message and stack instead of being silently swallowed.
+- Timer migration: resolves a real `node` binary when the host executable is not named `node`.
+- Notifications: `dispose` clears pending claim-release and popup timers.
+
+### Fixed
+
+- Pool: empty task results were treated as successes, so the same empty model kept winning every small task.
+- Pool: a killed refresh wrote nothing; partial scans now keep the previous pool and the next run is triggered after a reboot.
+- Docs: USER-STORIES covers the refresh and empty-response behavior (US-39), and US-4/US-5 no longer claim the pool never refuses work.
+
+## [2.2.0-beta.0] — context-aware pool routing
+
+### Added
+
+- Pool: context-aware routing. `pool.txt` lines carry a `c:<context>` token, and the router prefers the smallest model whose context fits the task prompt, so large-context models are reserved for large tasks.
+- Pool: `npm run score` probes providers with no direct HTTP endpoint (or an unexpanded `${...}` URL) through the local opencode session, and considers catalogue-priced models on providers other than `opencode`/`opencode-go`, whose credits cover them.
+- Pool: a rate-limit reset or a `402 out of credits` answer schedules a transient retry; a transient failure keeps the model's previous score until the retry.
+
+### Changed
+
+- Pool: the context floor is now 100k. Slots are ordered as `opencode-go` (weight 4), small-context providers (weight 2), large-context providers (weight 1), and at most one `opencode` model last.
+- Pool: small-context models are preferred for ordinary work; large-context models are used when the estimated task size needs them or when small slots are full.
+
 ## [2.1.0-beta.0] — every provider counts, and the timer installs itself
 
 ### Added

package/README.md

@@ -156,10 +156,24 @@ option is optional:
 | `feedbackHint` | `true` | Show the one-time feedback note on the first interactive run. `false` — or `LACODE_NO_FEEDBACK_HINT=1` — turns it off. The flag lives in `~/.local/share/opencode/lacode/feedback.json`; nothing is sent anywhere. |
 | `sound` | `true` | Play a short tone with each notification, one per outcome, so you can tell a finished session from a question or a permission prompt without reading the popup. It is opt-out: `false` — or `LACODE_SOUND=0` (`false`, `off`, `no` work too, and beat the option) — turns the sound off, and anything else, a typo included, leaves it on. |
 | `omniroute` | — | Optional [OmniRoute](docs/OMNIROUTE.md) gateway integration. Absent or `{ enabled: false }` changes nothing. |
+| `yolo` | `false` | Start with YOLO mode on: permission requests that would ask are auto-approved. Explicit `deny` rules still win. Toggle live with `/yolo`. |
+| `ykb` | `false` | Start with YKB (You Know Better) mode on: the agent is told not to ask, and questions that still appear are answered with the recommended option. Toggle live with `/ykb`. |
 | `watchMs` | `1000` | How often the pool files are checked for changes. `0` disables live reload. |
 | `watchDebounceMs` | `250` | Collapse a burst of editor writes into one reload. |
 | `imagePaste` | — | Image attachments: `maxBytes` (4 MiB), `maxImages` (10) and `dir`. |
 
+## Modes
+
+Two runtime modes can be seeded from plugin options and toggled from the TUI:
+
+| Command | Effect |
+| --- | --- |
+| `/yolo` | Toggle YOLO. Every permission request that would `ask` is approved automatically; an explicit `deny` still wins. |
+| `/ykb` | Toggle YKB (You Know Better). The agent is told not to ask questions and to continue with its recommendation. A question that still appears is answered with its `(Recommended)` option, or its first option when none is marked. |
+| `/modes` | Show both switches. |
+
+The toggles write `~/.config/lacode/modes.json` (override with `LACODE_MODES_FILE`) and the server picks the change up on the next decision — no restart. Turning a mode on also answers requests that are already pending. The `yolo` / `ykb` plugin options only seed the file on first run; after that, the file wins. A free-text question with no options is left for the user, since there is no option to choose.
+
 ### Staying up to date
 
 opencode's plugin cache is **write-once**: it resolves `@lawaty/lacode` to an

package/dist/agents.js

@@ -51,6 +51,8 @@ var FALLBACK_SLOTS = [
   { index: 3, model: "opencode/deepseek-v4.1-flash-free", weight: 1 },
   { index: 4, model: "opencode/glm-5.3-flash-free", weight: 1 }
 ];
+var CONTEXT_OVERFLOW_COOLDOWN_MS = 10 * 60 * 1e3;
+var BELOW_MIN_CONTEXT_COOLDOWN_MS = 6 * 60 * 60 * 1e3;
 var STATE_SWEEP_INTERVAL_MS = 60 * 60 * 1e3;
 var CHILD_MEMORY_MS = 30 * 6e4;
 

package/dist/commands.js

@@ -50,6 +50,8 @@ var DEFAULT_PRESET = join2(PRESETS_DIR, "free-tier.json");
 var USER_CONFIG_DIR = join2(homedir(), ".config", "lacode");
 var USER_MODELS_PATH = join2(USER_CONFIG_DIR, "pool.json");
 var USER_MODELS_TXT = join2(USER_CONFIG_DIR, "pool.txt");
+var CONTEXT_OVERFLOW_COOLDOWN_MS = 10 * 60 * 1e3;
+var BELOW_MIN_CONTEXT_COOLDOWN_MS = 6 * 60 * 60 * 1e3;
 var STATE_SWEEP_INTERVAL_MS = 60 * 60 * 1e3;
 var CHILD_MEMORY_MS = 30 * 6e4;