luxin-cli 0.2.8 → 0.2.10

package/CHANGELOG.md

@@ -6,6 +6,113 @@ provenance; this file is the human- and agent-readable release map.
 
 ## Unreleased
 
+## 0.2.10 - 2026-11-02
+
+Saving files, reading prompts from files, guides that do not pick a model you
+cannot run, and errors that say what to do.
+
+- New: `--output PATH` (with `--overwrite`) on `create` and `edit` downloads
+  the produced asset to your disk and reports it in `data.downloads[]`. The
+  extension always matches the bytes: ask for `hero.png`, get JPEG, and the
+  file is written as `hero.jpg` with `extension_corrected: true`. An existing
+  file is refused before any spend, and a download that fails after the media
+  was created keeps the paid result and returns an `assets get` recovery
+  command. CLI downloads send a `luxin-cli/0.2.10` User-Agent.
+- `--prompt-file PATH` works with `create`, `edit`, and their `--guide` modes;
+  the guide's commands replay `--prompt-file` (and `--output`/`--overwrite`)
+  instead of inlining the text. (#2371)
+
+- `video create --guide` picks a text-to-video model for a prompt-only brief.
+  When a model that needs a starting image is chosen, `next_command` carries
+  `--reference-image IMAGE_PATH` and `next_command_missing_inputs` names it.
+  (#2357)
+- The guide reads more shape phrasing: "a wide landscape hero banner", "a wide
+  cinematic composition", "hero image", "website banner", and "youtube
+  thumbnail" are 16:9; "instagram story", "phone wallpaper", and "tiktok video"
+  are 9:16. A create without `--aspect-ratio` warns that it planned the 1:1
+  default and names the ratio the prompt asked for. (#2358)
+- New models: GPT Image 2.5 Flare and Sunburst (`openai.gpt-image-2.5-flare`,
+  `openai.gpt-image-2.5-sunburst`, plus 2026-09-08 snapshots; create and
+  edit), with `quality` up to `xhigh` and `max` and transparent backgrounds.
+  Create defaults to `quality: medium` when the budget allows, as GPT Image 2
+  does, so the quote is a real number. Nano Banana 2.1
+  (`fal.google-nano-banana-2-1`, create and edit) at 1K/2K/4K.
+- Fixed: `create --guide --model nano-banana` (any model alias) failed with
+  "Cannot access 'MODEL_ALIAS_RULES' before initialization".
+- The `imagen4`, `imagen4-fast`, and `imagen4-ultra` aliases are gone: Fal
+  retired those endpoints. `--model imagen4` now selects nothing instead of a
+  model that cannot run. The `grok-imagine-image-2` aliases now work in the
+  published CLI.
+
+- `luxin skill` prints the guidance markdown by default; `luxin skill --json`
+  keeps the envelope.
+- Under `2>&1` a live create/edit no longer adds the stderr `in_flight` line,
+  so the merged stream is one JSON document. Separate stderr still gets it.
+  (#2376)
+- A non-retryable failure (`BAD_REQUEST`, `MODERATION_BLOCKED`,
+  `IMAGE_TOO_SMALL`, `PROVIDER_VALIDATION_ERROR`) no longer carries a
+  `recover_command` that re-runs the same arguments; `recovery.action` says
+  what to change instead. The idempotency key is still returned.
+- When `api.luxin.sh` does not resolve, commands return `NETWORK_UNAVAILABLE`
+  with `retryable: false` and the fix in `error.recovery.network_fixes`: the
+  Codex sandbox flag and the hosts a generic sandbox must allow. Temporary
+  DNS, refused, reset, and timeout errors return a retryable `NETWORK_ERROR`
+  that keeps the idempotency key and names the sandbox only as a possible
+  cause. `create --guide` puts the recovery on its `service_unreachable`
+  blocker. Code change: a reset or timeout used to be
+  `HOSTED_API_REQUEST_FAILED`; it is now `NETWORK_ERROR`, still retryable.
+  (#2393)
+- The skill refresh nudge prints `npx skills update` with no name filter, so
+  installs under an alias name such as `image-skill` update too.
+
+- Claude Code and Codex plugin. Claude Code:
+  `claude plugin marketplace add danielgwilson/luxin`, then
+  `claude plugin install luxin@luxin`. Codex:
+  `codex plugin marketplace add danielgwilson/luxin`, then
+  `codex plugin add luxin@luxin`. The plugin ships the `luxin` skill only. Its
+  version is `<cli version>+<content hash>`, so updates arrive when a shipped
+  file changes.
+
+## 0.2.9 - 2026-08-17
+
+Discovery, batch work, and recovery — the release for agents doing real
+multi-image jobs.
+
+- New command: `luxin models overview` — a ~2KB orientation payload with
+  category groups and counts, the cheapest executable model per category,
+  alias hints, and runnable next commands. Read this instead of the full list.
+- New command: `luxin models find` — terse, cost-sorted decision rows with
+  `--max-credits`, `--limit` (default 10), `--sort credits|usd`, a free-text
+  query, and the existing filters. Rows carry aspect-ratio support and honest
+  availability, include cataloged-but-not-yet-wired models (clearly marked,
+  ranked last), and report `total_matched` with a narrowing hint. The query
+  that used to return 4,000+ lines now answers in ~130. (#2114)
+- `models list` gains the same `--limit` / `--sort` / `--max-credits` flags
+  and is byte-identical when you don't use them. Against an older hosted API,
+  find/overview/narrowing degrade honestly client-side and say so in
+  `warnings`.
+- New: `luxin create --batch items.json` (or `--batch -`) — sequential
+  multi-image work as one command: shared defaults plus per-item prompts,
+  whole-document validation before any spend. `--dry-run` prices the whole
+  job with the same debit math the live run charges, checks your credits AND
+  the daily job cap, and answers "your credits cover K of N; the batch stops
+  at item K+1" with the top-up sized to the actual need. At any wall the
+  completed items are delivered and the batch stops cleanly with one digest
+  and one recovery block; re-running with the same `--batch-id` skips
+  completed, unchanged items. `seed_policy: fixed | increment | random` where
+  the model supports a seed. (#2304)
+- Recovery handle in the envelope: live create/edit responses now carry
+  `data.idempotency_key` on success and `error.recovery.idempotency_key` plus
+  a re-runnable `error.recovery.recover_command` on failure, so an
+  interrupted spend is recoverable from the envelope alone. stdout is exactly
+  one JSON document per invocation — now stated plainly in the contract; the
+  in-flight recovery line stays on stderr. (#2240)
+- Hosted-side, already live and noted for completeness: create/edit responses
+  echo `effective_parameters` (each knob tagged explicit / default_policy /
+  provider_default) so you can see the configuration that actually ran, and
+  `resolution` without `quality` on xai.grok-imagine-image-2.0 now sends the
+  medium tier you were already being quoted. (#2310)
+
 ## 0.2.8 - 2026-08-12
 
 - New command: `luxin skill` prints the full usage guidance from bundled

package/README.md

@@ -13,6 +13,7 @@ discoverable on skills.sh:
 
 ```bash
 npx skills add danielgwilson/luxin --full-depth --skill luxin -g -a codex -y
+npx skills add danielgwilson/luxin --full-depth --skill luxin -g -a claude-code -y
 ```
 
 Intent-named aliases are published from the same repo and point to the same
@@ -28,10 +29,40 @@ npx skills add danielgwilson/luxin --full-depth --skill image-to-3d -g -a codex
 npx skills add danielgwilson/luxin --full-depth --skill creative-media -g -a codex -y
 ```
 
-Or install straight from the hosted public contract for the always-latest build:
+Or install straight from the hosted public contract, which serves the build
+deployed to luxin.sh. `npx skills update luxin` refreshes either install:
 
 ```bash
 npx skills add https://luxin.sh --skill luxin -g -a codex -y
+npx skills add https://luxin.sh --skill luxin -g -a claude-code -y
+```
+
+Or install the Luxin plugin, which carries the same skill and updates through
+the plugin manager. In Claude Code:
+
+```bash
+claude plugin marketplace add danielgwilson/luxin
+claude plugin install luxin@luxin
+```
+
+Inside a session, `/plugin marketplace add danielgwilson/luxin` and then
+`/plugin install luxin@luxin` do the same. Update with
+`claude plugin marketplace update luxin && claude plugin update luxin@luxin`.
+
+In Codex:
+
+```bash
+codex plugin marketplace add danielgwilson/luxin
+codex plugin add luxin@luxin
+```
+
+Update with `codex plugin marketplace upgrade luxin`.
+
+Codex blocks network in its default sandbox, and Luxin calls
+`https://api.luxin.sh`. Start Codex with network on:
+
+```bash
+codex -s workspace-write -c sandbox_workspace_write.network_access=true
 ```
 
 If the Codex/global skill target is read-only or missing, keep the tracked slug
@@ -87,8 +118,8 @@ installed binary is `luxin`.
 
 ## Keeping The Skill Current
 
-The installed `SKILL.md` is a ~4KB discovery stub. Everything volatile — flags,
-workflows, recovery, funding — lives in the guidance the CLI itself carries, so
+The installed `SKILL.md` is a ~4KB discovery stub. Everything volatile (flags,
+workflows, recovery, funding) lives in the guidance the CLI itself carries, so
 it always matches the code you are running:
 
 ```bash
@@ -116,7 +147,7 @@ command and never rewrites your skill files, agent config, or hook files.
 ### Optional: refresh at session start (copy-paste, opt-in)
 
 Some runtimes can run a command when a session starts. If you want an automatic
-refresh, add it yourself — Luxin will not install it for you, and nothing below
+refresh, add it yourself. Luxin will not install it for you, and nothing below
 runs unless you put it there. For Claude Code, in `.claude/settings.json`:
 
 ```json