Prepare for tinycloud 0.3.19: the silent-stop fix — surface and recover output-token truncation (no floor raise)

.claude-plugin/marketplace.json

@@ -7,14 +7,14 @@
   },
   "metadata": {
     "description": "Deep video and image work for coding agents: analyze, describe, search, caption, clip, detect faces, and publish via the tinycloud CLI.",
-    "version": "0.3.18"
+    "version": "0.3.19"
   },
   "plugins": [
     {
       "name": "tinycloud",
       "source": "./",
       "description": "Video and image understanding and editing via the tinycloud CLI (Cloudglue): watch/see/extract/search/ask/face over videos and images, captions, clips, collections, plus one-command workflows for sales coaching, blog posts, ad analysis, meeting breakdowns, and YouTube publishing.",
-      "version": "0.3.18",
+      "version": "0.3.19",
       "author": { "name": "Cloudglue" },
       "homepage": "https://tinycloud.sh",
       "repository": "https://github.com/cloudglue/tinycloud",

.claude-plugin/plugin.json

@@ -2,7 +2,7 @@
   "name": "tinycloud",
   "displayName": "Tinycloud Video",
   "description": "Video and image understanding and editing via the tinycloud CLI (Cloudglue): watch/see/extract/search/ask/face over videos and images, captions, clips, collections, and one-command video workflows.",
-  "version": "0.3.18",
+  "version": "0.3.19",
   "author": { "name": "Cloudglue", "url": "https://cloudglue.dev" },
   "homepage": "https://tinycloud.sh",
   "repository": "https://github.com/cloudglue/tinycloud",

CLAUDE.md

@@ -261,7 +261,33 @@ also rejects `<cg-query>`, `<cg-transcript>`, and `<cg-chapters auto>` (the
 private-only; explicit-children `<cg-chapters>` stays public-safe). Because
 the skill teaches the new elements and relies on that extended guard, the
 floor was raised to 0.3.18 (same merge-after-CDN gate — the dist PR merges
-only after CDN `channels.stable` = 0.3.18). The host-level `profile` verb and the leading global flags `--home`/`--profile`
+only after CDN `channels.stable` = 0.3.18). 0.3.19 is a **host-agent
+reliability release — the silent-stop fix** (no new verbs, flags, or feature
+ids — verbs stay 17, features stay 42; the changes live in the interactive
+agent and headless `-p`, not the video-verb surface the skill drives, so the
+skill floor stays 0.3.18 — the 0.3.13 no-raise pattern): adaptive-thinking
+chat models stream a thinking block that spends from the same per-response
+`max_tokens` budget as text and tool calls, so a long think could exhaust the
+old 8192-token cap, return `stop_reason: max_tokens` with no error, and the
+agent silently yielded mid-task with only a duration line (typing `continue`
+resumed it). Now a `length`-stopped turn — and an empty `stop`-ended one,
+e.g. a dropped stream read as a clean end — surfaces a status line instead of
+a quiet handoff; the cap rises to 32000, tunable via
+`TINYCLOUD_MODEL_MAX_TOKENS`; a guard blocks the final tool call of a
+`length`-stopped message with an error result (truncated tool-call JSON gets
+repaired into valid-but-incomplete args, e.g. a write with half a file) so
+the model re-issues it in smaller pieces, wired in both the interactive
+agent and headless `-p`; a truncated turn auto-continues via a hidden
+prompt, bounded to 2 consecutive per user turn, disabled with
+`TINYCLOUD_AUTO_CONTINUE=0` or `preferences.autoContinue: false`; streamed
+thinking shows a "Reasoning" loader label; and headless `-p --json` traces
+gain an additive `stop_reason` field plus a stderr warning on truncation.
+Headless `-p` also reaches parity with the interactive agent on model
+failures: it gains the 0.3.7 chat-retry layer (`TINYCLOUD_MODEL_RETRIES` now
+covers both agents), and an errored turn is truthful — `✖ turn failed` on
+stderr, an `error` field in the `--json` trace, exit 1 (previously a
+backend blip could exit 0 with an empty, zero-usage trace).
+The host-level `profile` verb and the leading global flags `--home`/`--profile`
 (also `$TINYCLOUD_HOME`; 0.3.3+) relocate state and are intentionally absent
 from `commands --json` — like the launcher's install/update, they're CLI/host
 concerns, not video operations.
@@ -313,9 +339,9 @@ of printing JSON. Any script invoking the binary must redirect `</dev/null`
   metadata sync) vs live-CDN jobs (`Install + smoke` matrix, npx-against-CDN)
   which run only on push to main or manual dispatch — never on PRs, because a
   CDN gap would fail every PR.
-- The live CDN serves 0.3.18 (latest aliases + v-prefixed pinned tarballs
-  for 0.3.0 through 0.3.18, with `manifest.json` + `.sha256`
-  sidecars; `channels.stable` = 0.3.18); all smoke legs are required.
+- The live CDN serves 0.3.19 (latest aliases + v-prefixed pinned tarballs
+  for 0.3.0 through 0.3.19, with `manifest.json` + `.sha256`
+  sidecars; `channels.stable` = 0.3.19); all smoke legs are required.
 - `publish-npm.yml` (tag `v*`): asserts tag == package.json version → gates
   on `generate-manifest.mjs --check` against the live CDN → publishes via
   npm trusted publishing (OIDC, `id-token: write`, npm ≥ 11.5.1 — no token

README.md

@@ -17,11 +17,11 @@ The npm package is a small launcher: on first run it downloads the matching
 platform distribution from Cloudglue's CDN (cached under
 `~/.tinycloud/versions/<version>/`), verifies its checksum, and execs the real
 binary. The package version pins the binary version, so
-`npx @cloudglue/tinycloud@0.3.18` always runs tinycloud 0.3.18. It also adds two
+`npx @cloudglue/tinycloud@0.3.19` always runs tinycloud 0.3.19. It also adds two
 wrapper commands:
 
 ```bash
-tinycloud install --version 0.3.18   # pre-download a version
+tinycloud install --version 0.3.19   # pre-download a version
 tinycloud install --latest          # install latest stable and pin to it
 tinycloud update                    # move to latest stable, prune old versions
 ```

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@cloudglue/tinycloud",
-  "version": "0.3.18",
+  "version": "0.3.19",
   "description": "Agent CLI for deep video and image work, by Cloudglue. Downloads the tinycloud binary on first run.",
   "bin": {
     "tinycloud": "bin/tinycloud.js"

skills/tinycloud/reference/setup.md

@@ -114,7 +114,9 @@ binary reports in `--version --json`.
 | `TINYCLOUD_UPLOAD_IDLE_TIMEOUT_MS` | Idle (no-progress) deadline for an upload — aborts a wedged transfer in ~1min instead of waiting out the total upload deadline (default 60s; `0` disables) |
 | `TINYCLOUD_HTTP_RETRIES` | Bounded auto-retries for idempotent (GET/HEAD) Cloudglue calls on transient errors — 408/429/5xx, network blips (default 5; `0` disables) |
 | `TINYCLOUD_JOB_WAIT_TIMEOUT_MS` | Wall-clock budget for async job waits (the `waitForReady` poll behind `watch`/`extract`/`face`) (default 10min; `0` disables) |
-| `TINYCLOUD_MODEL_RETRIES` | Client-side retries for the interactive agent's chat-model turns on transient connection errors (default 5; `0` disables). 0.3.7+ |
+| `TINYCLOUD_MODEL_RETRIES` | Client-side retries for agent chat-model turns on transient connection errors (default 5; `0` disables). 0.3.7+ interactive; 0.3.19+ also headless `-p` |
+| `TINYCLOUD_MODEL_MAX_TOKENS` | Per-response output-token cap for agent chat-model turns (interactive and headless `-p`) — thinking, text, and tool calls share it (default 32000). 0.3.19+ |
+| `TINYCLOUD_AUTO_CONTINUE` | `0` stops the interactive agent auto-resuming after an output-token-truncated turn (default on, bounded to 2 consecutive; also `preferences.autoContinue: false` in config). 0.3.19+ |
 
 Every Cloudglue request carries a hard deadline, so a stalled route can never
 hang the CLI indefinitely; a timeout surfaces as a retryable `upstream` error

skills/tinycloud/tinycloud-skill.json

@@ -1,5 +1,5 @@
 {
-  "skill_version": "0.3.18",
+  "skill_version": "0.3.19",
   "tinycloud": {
     "min_version": "0.3.18",
     "supported_range": ">=0.3.18 <0.4.0",