ai-switch-cli 0.3.0__tar.gz → 0.5.0__tar.gz

{ai_switch_cli-0.3.0 → ai_switch_cli-0.5.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: ai-switch-cli
-Version: 0.3.0
+Version: 0.5.0
 Summary: One-command provider and model switching for Claude Code and Codex, with session-history checks
 Author: ZidongS
 License: MIT
@@ -78,13 +78,40 @@ Select `glm` or `deepseek`, enter the API key, and activate the generated config
 ai-switch use glm
 ```
 
-`use` lists the models the profile offers and remembers your choice:
+`use` activates the profile and does **not** ask anything: the profile's whole model list is published, so you
+pick the model inside the agent with `/model` (Codex) or its Opus/Sonnet/Haiku categories (Claude Code). The
+default for new sessions is the one you last used, or the profile default. Add `--choose` if you want to be
+asked, or `-m/--model` to set the default directly.
 
 ```text
-This profile provides 2 models:
+  ⠸ ▰▰▰▰▰▰▱▱▱▱▱▱▱▱▱▱  38% 2/5  writing ~/.codex/models.json
+  ✓ ▰▰▰▰▰▰▰▰▰▰▰▰▱▱▱▱  75% 4/5  write ~/.codex/models.json  3 model(s)
+  ✓ ▰▰▰▰▰▰▰▰▰▰▰▰▰▰▰▰ 100% 5/5  save profile state  glm is now the active profile
+╭──────────────────────────────────────────────────────────╮
+│ ✦ glm is active                                 (0.73s)  │
+├──────────────────────────────────────────────────────────┤
+│ profile glm                                              │
+│ default glm-5.3                                          │
+│ models  3 published  ·  switch inside codex with /model  │
+│ backup  ~/.config/ai-switch/backups/20260915-203657      │
+│ next    restart claude/codex to reload                   │
+╰──────────────────────────────────────────────────────────╯
+```
+
+The bar is a colour ramp (cyan → green) with a bright leading edge and it always finishes completely
+filled at 100%, followed by a short sweep as the closing flourish.
+
+The progress animation needs a terminal — piped output, `TERM=dumb`, `--plain` and `--no-color` keep the
+plain one-line-per-action form (handy for scripts). `NO_COLOR=1` mutes the colours but keeps the spinner.
+
+With `--choose` you get the picker instead:
+
+```text
+This profile provides 3 models:
   1) glm-5.3         GLM-5.3 flagship (1M context)  [default]
-  2) glm-5.3-flash   GLM-5.3 flash (fast, 1M context)
-Select model [1-2, Enter=glm-5.3, q=cancel]:
+  2) glm-5.3-flash   GLM-5.3 Flash (fast, text+image, 1M context)
+  3) glm-5-turbo     GLM-5 Turbo (agent-optimized, 200K context)
+Select model [1-3, Enter=glm-5.3, q=cancel]:
 ```
 
 For scripts and servers, choose without prompting (name, unique prefix, or index):
@@ -93,8 +120,11 @@ For scripts and servers, choose without prompting (name, unique prefix, or index
 ai-switch use glm --model glm-5.3-flash
 ai-switch use glm -m 2 -y
 ai-switch use glm --dry-run          # show what would change
+ai-switch use glm -m glm-5.3 --pin   # publish only this model (see below)
 ```
 
+Selecting a model sets the **default for new sessions** in both agents, and the profile's other models stay in the catalogue so you can still switch inside the agent. That means an in-app `/model` choice is a per-session override — and Codex writes its own choice back to `~/.codex/config.toml` when it exits, so the two agents can drift apart without anything else noticing; `ai-switch doctor` reports exactly that. If you want an activation to be exclusive instead, add `--pin`: the published catalogue then holds only the selected model, so Codex's own picker has nothing else to switch to (the next `ai-switch use` without `--pin` restores the full menu).
+
 Edit `~/.codex/config.toml` and `~/.claude/settings.json` for another provider, then save that configuration as a second profile:
 
 ```bash
@@ -116,7 +146,25 @@ Update an existing profile description:
 ai-switch describe default "Default daily configuration"
 ```
 
-`list` shows the active marker, profile name, configured clients, description, detected models, and endpoint hostnames. API keys are never printed.
+`list` shows the active marker, profile name, configured clients, description, detected models, and endpoint hostnames. `current` prints a panel of **every** profile with its models, the default one (★), the endpoint and — for the active profile — whether the live agent configuration still matches:
+
+```text
+ai-switch  ·  4 profile(s)  ·  ~/.config/ai-switch
+────────────────────────────────────────────────────────────────────
+  ○ glm   glm-coding-plan
+      Codex   glm-5.3  · 3 models
+      Claude  glm-5.3  · 2 Claude category mappings
+      host    open.bigmodel.cn
+      models  glm-5.3 ★, glm-5.3-flash, glm-5-turbo
+
+  ● paratera   custom paratera   active
+      Codex   DeepSeek-V4.1-Flash  · 31 models
+      Claude  DeepSeek-V4.1-Flash  · opus/sonnet/haiku → DeepSeek-V4.1-Flash
+      host    llmapi.paratera.com
+      models  DeepSeek-V4.1-Flash ★, DeepSeek-V4-Pro, …, MiniMax-Text-01
+```
+
+Use `current --plain` to print just the active profile name. API keys are never printed.
 
 ## Choosing a model
 
@@ -154,6 +202,32 @@ ai-switch use deepseek -m deepseek-flash
 
 Model suffixes are proxy-specific: DeepSeek's Claude endpoint accepts and strips a `[1m]` marker (`deepseek-v4-pro[1m]`), while ZAI rejects it as an unknown model — `glm-5.3` is already 1M context and must be used without a suffix. Worth a one-token request before trusting a name in a profile.
 
+## Gateways that break Codex's tool calls
+
+Codex 0.154 records an assistant `message` item **between** a turn's tool calls and their results — the text it said before calling the tools, or an empty string when it said nothing. That is valid in the Responses API, but a gateway that translates Responses into Chat completions emits it as its own chat message, which leaves the `tool_calls` message without the `tool` messages that must follow it directly. The next request is then rejected and the session dies:
+
+```text
+OpenAIException - The request is invalid: An assistant message with 'tool_calls' must be
+followed by tool messages responding to each 'tool_call_id'. (insufficient tool messages
+following tool_calls message)
+```
+
+Paratera (`llmapi.paratera.com`) does this. Claude Code is unaffected, because it speaks the Anthropic API rather than the Responses API — only Codex hits the translation.
+
+For such a profile, `ai-switch use` starts a small local proxy that moves that message back in front of the calls, points Codex at it, and stops it again when you activate a profile that does not need it:
+
+```text
+Note: Codex will use the llmapi.paratera.com patch proxy on 127.0.0.1:8791
+      (Paratera translates Responses into Chat completions and splits a turn's tool calls from their results).
+```
+
+* The profile keeps the provider's real endpoint. Only the live `~/.codex/config.toml` is repointed, and only Codex's own configuration — Claude Code still talks to the provider directly.
+* The proxy listens on `127.0.0.1` only, forwards your key and your request body untouched apart from the item order, and logs counts and status codes (never keys, prompts or answers) to `~/.config/ai-switch/patch.log`.
+* If the process dies (a reboot, a kill), Codex cannot reach its endpoint. `ai-switch doctor` reports it and `ai-switch use NAME` starts it again. `ai-switch current` shows whether it is up.
+* Nothing here is needed once the gateway's translation is fixed: `ai-switch use NAME --no-patch` activates the profile without the proxy and stops it if it is running.
+
+This is a workaround for the gateway, not a fix — the rewriting is deliberately narrow (only a completed `calls + message + results` turn, only on `/responses`), and anything else is forwarded byte for byte.
+
 ## Session history
 
 Conversation history belongs to the agents, not to a provider, so ai-switch never manages it. These files are deliberately excluded from profiles and switches:
@@ -188,7 +262,7 @@ ai-switch doctor --fix    # quarantine damaged Codex runtime databases
 
 ## Safety and storage
 
-Before activation, the current Codex and Claude files are backed up under `~/.config/ai-switch/backups/<timestamp>/` (files that a switch removes are kept under `removed/`). Profiles are stored under `~/.config/ai-switch/profiles/`; directories use mode 700 and files use mode 600. The active profile and the last used model are recorded in `~/.config/ai-switch/state.json`. Restart `claude` or `codex` after switching so the process reloads its configuration.
+Before activation, the current Codex and Claude files are backed up under `~/.config/ai-switch/backups/<timestamp>/` (files that a switch removes are kept under `removed/`). Profiles are stored under `~/.config/ai-switch/profiles/`; directories use mode 700 and files use mode 600. The active profile, the last used model and the running gateway patch proxy (pid, port, endpoint) are recorded in `~/.config/ai-switch/state.json`; that proxy's log is `~/.config/ai-switch/patch.log`. Restart `claude` or `codex` after switching so the process reloads its configuration.
 
 Set `AI_SWITCH_HOME` to use a different profile directory, `CODEX_HOME`/`CLAUDE_CONFIG_DIR` for relocated agent configurations, and `AI_SWITCH_API_KEY` to fill the API key prompt non-interactively.
 
@@ -197,9 +271,10 @@ Set `AI_SWITCH_HOME` to use a different profile directory, `CODEX_HOME`/`CLAUDE_
 ```text
 ai-switch init NAME [-d DESCRIPTION]  Save current files as a new profile
 ai-switch list                        List profiles and configuration summaries
-ai-switch use NAME [-m MODEL] [-y] [-n]  Back up and activate a profile (and a model)
+ai-switch use NAME [-m MODEL] [--choose] [--pin] [--plain] [--no-patch]
+                                      Back up and activate a profile (animated, -m sets the default)
 ai-switch models [NAME] [--json]      Show the models a profile offers
-ai-switch current                     Print the active profile
+ai-switch current [-p]                Show every profile, its models and the live state (-p: name only)
 ai-switch describe NAME TEXT          Set a profile description
 ai-switch upgrade NAME                Derive models.json for a profile from an older release
 ai-switch add                         Create a profile interactively (GLM, DeepSeek, or custom)

{ai_switch_cli-0.3.0 → ai_switch_cli-0.5.0}/README.md

@@ -64,13 +64,40 @@ Select `glm` or `deepseek`, enter the API key, and activate the generated config
 ai-switch use glm
 ```
 
-`use` lists the models the profile offers and remembers your choice:
+`use` activates the profile and does **not** ask anything: the profile's whole model list is published, so you
+pick the model inside the agent with `/model` (Codex) or its Opus/Sonnet/Haiku categories (Claude Code). The
+default for new sessions is the one you last used, or the profile default. Add `--choose` if you want to be
+asked, or `-m/--model` to set the default directly.
 
 ```text
-This profile provides 2 models:
+  ⠸ ▰▰▰▰▰▰▱▱▱▱▱▱▱▱▱▱  38% 2/5  writing ~/.codex/models.json
+  ✓ ▰▰▰▰▰▰▰▰▰▰▰▰▱▱▱▱  75% 4/5  write ~/.codex/models.json  3 model(s)
+  ✓ ▰▰▰▰▰▰▰▰▰▰▰▰▰▰▰▰ 100% 5/5  save profile state  glm is now the active profile
+╭──────────────────────────────────────────────────────────╮
+│ ✦ glm is active                                 (0.73s)  │
+├──────────────────────────────────────────────────────────┤
+│ profile glm                                              │
+│ default glm-5.3                                          │
+│ models  3 published  ·  switch inside codex with /model  │
+│ backup  ~/.config/ai-switch/backups/20260915-203657      │
+│ next    restart claude/codex to reload                   │
+╰──────────────────────────────────────────────────────────╯
+```
+
+The bar is a colour ramp (cyan → green) with a bright leading edge and it always finishes completely
+filled at 100%, followed by a short sweep as the closing flourish.
+
+The progress animation needs a terminal — piped output, `TERM=dumb`, `--plain` and `--no-color` keep the
+plain one-line-per-action form (handy for scripts). `NO_COLOR=1` mutes the colours but keeps the spinner.
+
+With `--choose` you get the picker instead:
+
+```text
+This profile provides 3 models:
   1) glm-5.3         GLM-5.3 flagship (1M context)  [default]
-  2) glm-5.3-flash   GLM-5.3 flash (fast, 1M context)
-Select model [1-2, Enter=glm-5.3, q=cancel]:
+  2) glm-5.3-flash   GLM-5.3 Flash (fast, text+image, 1M context)
+  3) glm-5-turbo     GLM-5 Turbo (agent-optimized, 200K context)
+Select model [1-3, Enter=glm-5.3, q=cancel]:
 ```
 
 For scripts and servers, choose without prompting (name, unique prefix, or index):
@@ -79,8 +106,11 @@ For scripts and servers, choose without prompting (name, unique prefix, or index
 ai-switch use glm --model glm-5.3-flash
 ai-switch use glm -m 2 -y
 ai-switch use glm --dry-run          # show what would change
+ai-switch use glm -m glm-5.3 --pin   # publish only this model (see below)
 ```
 
+Selecting a model sets the **default for new sessions** in both agents, and the profile's other models stay in the catalogue so you can still switch inside the agent. That means an in-app `/model` choice is a per-session override — and Codex writes its own choice back to `~/.codex/config.toml` when it exits, so the two agents can drift apart without anything else noticing; `ai-switch doctor` reports exactly that. If you want an activation to be exclusive instead, add `--pin`: the published catalogue then holds only the selected model, so Codex's own picker has nothing else to switch to (the next `ai-switch use` without `--pin` restores the full menu).
+
 Edit `~/.codex/config.toml` and `~/.claude/settings.json` for another provider, then save that configuration as a second profile:
 
 ```bash
@@ -102,7 +132,25 @@ Update an existing profile description:
 ai-switch describe default "Default daily configuration"
 ```
 
-`list` shows the active marker, profile name, configured clients, description, detected models, and endpoint hostnames. API keys are never printed.
+`list` shows the active marker, profile name, configured clients, description, detected models, and endpoint hostnames. `current` prints a panel of **every** profile with its models, the default one (★), the endpoint and — for the active profile — whether the live agent configuration still matches:
+
+```text
+ai-switch  ·  4 profile(s)  ·  ~/.config/ai-switch
+────────────────────────────────────────────────────────────────────
+  ○ glm   glm-coding-plan
+      Codex   glm-5.3  · 3 models
+      Claude  glm-5.3  · 2 Claude category mappings
+      host    open.bigmodel.cn
+      models  glm-5.3 ★, glm-5.3-flash, glm-5-turbo
+
+  ● paratera   custom paratera   active
+      Codex   DeepSeek-V4.1-Flash  · 31 models
+      Claude  DeepSeek-V4.1-Flash  · opus/sonnet/haiku → DeepSeek-V4.1-Flash
+      host    llmapi.paratera.com
+      models  DeepSeek-V4.1-Flash ★, DeepSeek-V4-Pro, …, MiniMax-Text-01
+```
+
+Use `current --plain` to print just the active profile name. API keys are never printed.
 
 ## Choosing a model
 
@@ -140,6 +188,32 @@ ai-switch use deepseek -m deepseek-flash
 
 Model suffixes are proxy-specific: DeepSeek's Claude endpoint accepts and strips a `[1m]` marker (`deepseek-v4-pro[1m]`), while ZAI rejects it as an unknown model — `glm-5.3` is already 1M context and must be used without a suffix. Worth a one-token request before trusting a name in a profile.
 
+## Gateways that break Codex's tool calls
+
+Codex 0.154 records an assistant `message` item **between** a turn's tool calls and their results — the text it said before calling the tools, or an empty string when it said nothing. That is valid in the Responses API, but a gateway that translates Responses into Chat completions emits it as its own chat message, which leaves the `tool_calls` message without the `tool` messages that must follow it directly. The next request is then rejected and the session dies:
+
+```text
+OpenAIException - The request is invalid: An assistant message with 'tool_calls' must be
+followed by tool messages responding to each 'tool_call_id'. (insufficient tool messages
+following tool_calls message)
+```
+
+Paratera (`llmapi.paratera.com`) does this. Claude Code is unaffected, because it speaks the Anthropic API rather than the Responses API — only Codex hits the translation.
+
+For such a profile, `ai-switch use` starts a small local proxy that moves that message back in front of the calls, points Codex at it, and stops it again when you activate a profile that does not need it:
+
+```text
+Note: Codex will use the llmapi.paratera.com patch proxy on 127.0.0.1:8791
+      (Paratera translates Responses into Chat completions and splits a turn's tool calls from their results).
+```
+
+* The profile keeps the provider's real endpoint. Only the live `~/.codex/config.toml` is repointed, and only Codex's own configuration — Claude Code still talks to the provider directly.
+* The proxy listens on `127.0.0.1` only, forwards your key and your request body untouched apart from the item order, and logs counts and status codes (never keys, prompts or answers) to `~/.config/ai-switch/patch.log`.
+* If the process dies (a reboot, a kill), Codex cannot reach its endpoint. `ai-switch doctor` reports it and `ai-switch use NAME` starts it again. `ai-switch current` shows whether it is up.
+* Nothing here is needed once the gateway's translation is fixed: `ai-switch use NAME --no-patch` activates the profile without the proxy and stops it if it is running.
+
+This is a workaround for the gateway, not a fix — the rewriting is deliberately narrow (only a completed `calls + message + results` turn, only on `/responses`), and anything else is forwarded byte for byte.
+
 ## Session history
 
 Conversation history belongs to the agents, not to a provider, so ai-switch never manages it. These files are deliberately excluded from profiles and switches:
@@ -174,7 +248,7 @@ ai-switch doctor --fix    # quarantine damaged Codex runtime databases
 
 ## Safety and storage
 
-Before activation, the current Codex and Claude files are backed up under `~/.config/ai-switch/backups/<timestamp>/` (files that a switch removes are kept under `removed/`). Profiles are stored under `~/.config/ai-switch/profiles/`; directories use mode 700 and files use mode 600. The active profile and the last used model are recorded in `~/.config/ai-switch/state.json`. Restart `claude` or `codex` after switching so the process reloads its configuration.
+Before activation, the current Codex and Claude files are backed up under `~/.config/ai-switch/backups/<timestamp>/` (files that a switch removes are kept under `removed/`). Profiles are stored under `~/.config/ai-switch/profiles/`; directories use mode 700 and files use mode 600. The active profile, the last used model and the running gateway patch proxy (pid, port, endpoint) are recorded in `~/.config/ai-switch/state.json`; that proxy's log is `~/.config/ai-switch/patch.log`. Restart `claude` or `codex` after switching so the process reloads its configuration.
 
 Set `AI_SWITCH_HOME` to use a different profile directory, `CODEX_HOME`/`CLAUDE_CONFIG_DIR` for relocated agent configurations, and `AI_SWITCH_API_KEY` to fill the API key prompt non-interactively.
 
@@ -183,9 +257,10 @@ Set `AI_SWITCH_HOME` to use a different profile directory, `CODEX_HOME`/`CLAUDE_
 ```text
 ai-switch init NAME [-d DESCRIPTION]  Save current files as a new profile
 ai-switch list                        List profiles and configuration summaries
-ai-switch use NAME [-m MODEL] [-y] [-n]  Back up and activate a profile (and a model)
+ai-switch use NAME [-m MODEL] [--choose] [--pin] [--plain] [--no-patch]
+                                      Back up and activate a profile (animated, -m sets the default)
 ai-switch models [NAME] [--json]      Show the models a profile offers
-ai-switch current                     Print the active profile
+ai-switch current [-p]                Show every profile, its models and the live state (-p: name only)
 ai-switch describe NAME TEXT          Set a profile description
 ai-switch upgrade NAME                Derive models.json for a profile from an older release
 ai-switch add                         Create a profile interactively (GLM, DeepSeek, or custom)