ccse — one-line model switch across coding agents

ccse (cc-switch-enhanced) rewrites the model-name string in 23 coding-agent configs at once. It touches only model names — never baseurl / apikey / other fields. Every write is snapshotted first and is atomic, so it is safer than hand-editing JSON/TOML/YAML/env files.

Prefer ccse over manual Edit when the task is changing model names in agent configs. Hand-editing risks: wrong field, lost route prefix, broken JSON, no backup. ccse knows each adapter's exact slot path and keeps prefixes/suffixes correct.

Golden rule: preview before writing

Always dry-run first. Never apply blind.

ccse diff --model "glm-5.2"        # preview the exact diffs (writes nothing)
ccse --model "glm-5.2" --dry       # same, top-level form
ccse show                          # see current state first (★ = primary slot)

Core commands

ccse show                          # current model per agent/slot (★ = the one --model sets)
ccse list                          # known adapters + whether file exists
ccse --model "glm-5.2"             # switch every agent's PRIMARY slot (★) to glm-5.2
ccse --model "glm-5.2" --dry       # preview only
ccse --model "X" --only claude,codex,grok      # restrict to these adapters
ccse --model "X" --exclude pi,copaw            # skip these
ccse undo                          # revert the most recent apply
ccse undo 20260812-132700          # revert a specific snapshot
ccse snapshots                     # list saved snapshots
ccse history                       # apply history

ccse --model NAME is shorthand for ccse apply --model NAME. It hits only the primary slot (the ★ row in show) of each adapter, not every slot. To change multiple slots per agent (e.g. Claude's haiku/sonnet/opus separately), use a profile.

Prefix & suffix conventions — read before you pick a NAME

ccse show prints names like newapi/gpt-5.6-sol, krill/gpt-5.6-terra, glm-5.2[1M].

• Route prefix (newapi/, krill/, dmx/, …): tells the agent which upstream to use.

- A bare NAME (glm-5.2) keeps each adapter's existing prefix automatically. This is what you almost always want. - A NAME containing / (newapi/glm-5.2) is written verbatim, overriding the prefix. - --no-keep-prefix writes the name verbatim and drops the prefix — use only on purpose.

• Context suffix ([1M]): encodes the context-window size. Bare names preserve the suffix

the adapter already had; pass it explicitly (glm-5.2[1M]) to set it. Don't strip it unless asked — it carries the 1M-context intent.

When the user says "change model to X" with no prefix, pass the bare name and let ccse keep each adapter's prefix. Only inject a prefix if the user explicitly names a different upstream.

Profiles — multi-slot recipes

For anything beyond "one model for every primary slot", use profiles in ~/.ccse/profiles.toml:

[bailian-coding]
"claude.model"   = "qwen3.6-plus"
"claude.sonnet"  = "qwen3.8-max"     # different model per tier
"claude.opus"    = "qwen3.6-plus"
"codex.model"    = "qwen3.6-plus"
"grok.model"     = "qwen3.6-plus"

• Format: "adapter.slot" = "model-name". adapter selects the config, slot selects the

field inside it (see ccse show for exact slot names). Slots omitted from the profile are not touched.

• ccse profiles — list saved profiles
• ccse diff <profile> — preview
• ccse apply <profile> — write into all listed agents
• ccse genprofile --name snapshot — capture the current state as a reusable profile

This is how to express "Claude sonnet → qwen3.8-max, everything else → glm-5.2" cleanly.

Safety model

• Every apply / --model (non-dry) snapshots to ~/.ccse/snapshots/<stamp>/ before writing.

--no-backup skips this — do not use it unless the user insists.

• Restore re-backs the current file to *.ccse.pre-restore.<stamp> first, so undo never loses

data a second time.

• Atomic writes (temp file + os.replace); a parse failure aborts without touching the original.

If a user says "I changed a model and something broke", reach for ccse undo (or ccse snapshots → ccse undo <stamp>) before manual repair.

Supported adapters (23)

Two adapters look alike — don't confuse them:

• kilocode is the VS Code extension (~/.kilocode/cli/config.json, providers[] list).
• kilo is the standalone kilo CLI (~/.config/kilo/kilo.json). Its model is

referenced as newapi/<name> strings in many fields, and the CLI validates a model registry at provider.<provider>.models.<bare-name> at startup — the entry must carry modalities.output or kilo refuses to start. ccse handles both: it rewrites every newapi/<name> string (top-level model, subagentmodel, smallmodel, experimental.sweprunermodel, each agent.<name>.model, subagentvariantoverrides) AND (re)creates the registry entry for the new bare name, copying schema from the previous model when present.

ccse list shows which adapter files are present on this machine; ccse show prints the live value of every slot. Run ccse show before deciding what to switch.

Workflows

"Switch everything to one model"

ccse show                              # confirm what's primary now
ccse --model "glm-5.2" --dry           # preview
ccse --model "glm-5.2"                 # apply

"Switch only a few agents"

ccse --model "glm-5.2" --only claude,grok,codex --dry
ccse --model "glm-5.2" --only claude,grok,codex

"Different model per slot" (e.g. Claude sonnet differs from the rest)

# write ~/.ccse/profiles.toml, then:
ccse diff myprofile
ccse apply myprofile

"Revert a bad switch"

ccse snapshots        # find the stamp
ccse undo             # or: ccse undo <stamp>

Gotchas

• Preview first: ccse diff --model NAME writes nothing; apply snapshots before atomic replacement.
• Bare model names preserve each adapter's route prefix; names containing / override it.
• kilo and kilocode are different adapters. Kilo also needs its provider model registry entry updated.
• prime and pi resolve active models through their catalogs; switching only the visible settings field can leave them broken.
• OpenAI-compatible base URLs are normalized to /v1; an already suffixed URL is left unchanged.

Install & download

Requires ccse on PATH. Install from the release artifacts (Python ≥3.11) or from the repo:

# from the CNB release (0.2.4, latest)
uv tool install "cc-switch-enhanced @ https://cnb.cool/tonykw12/cc-switch-enhanced/-/releases/download/0.2.4/cc_switch_enhanced-0.2.4-py3-none-any.whl"
# or pip:
pip install "https://cnb.cool/tonykw12/cc-switch-enhanced/-/releases/download/0.2.4/cc_switch_enhanced-0.2.4-py3-none-any.whl"

# from the repo
uv tool install .    # or: pipx install .

Direct download links:

• wheel: https://cnb.cool/tonykw12/cc-switch-enhanced/-/releases/download/0.2.4/ccswitchenhanced-0.2.4-py3-none-any.whl
• sdist: https://cnb.cool/tonykw12/cc-switch-enhanced/-/releases/download/0.2.4/ccswitchenhanced-0.2.4.tar.gz

If command -v ccse fails after install, the CLI isn't on PATH — tell the user rather than hand-editing. Verify with ccse list (prints the 23 known adapters and which config files exist).

• env-rc adapters (kimi, copaw) edit a single export VAR=... line in ~/.zshrc; names with

special chars (glm-5.2[1M]) are single-quote-escaped safely. Other rc content is untouched.

• JSON-path selectors supported in profiles: providers[id=newapi].model,

providers[newapi].apiModelId (shorthand matching id/provider/name/key/envKey), providers.openai-codex-cli.settings.model (dotted/dashed keys verbatim).