unknown prompt variable "{{cwd}}"

[xiazemin]

本轮运行失败unknown prompt variable "{{cwd}}" in section "deployment:persona-suffix"; registered variables: provider, model

[argszero]

The name is not a typo you can chase from the message: it is interpolated while the request is assembled (packages/core/system-prompt/src/index.ts), and an unregistered name is a hard error there by design, so nothing is left as literal text.

What removed it: commit 79bd3d8da7 (first tagged dsh-v0.2.1-alpha.2) deleted the cwd registration and the default persona suffix that used it, and handed the duty to the required runtime context working-directory:current. The current tree registers exactly two prompt variables — provider and model — which is the list your message prints. So the surviving {{cwd}} sits in a layer an upgrade does not rewrite: your profile's cordis.patch.yml, your home patch, or a preset row. It stays there because the config editor writes a whole config object back rather than the one field that was edited, so a value the old release materialized is frozen permanently.

I published a shim for exactly this: https://github.com/argszero/dsh-legacy-prompt-vars (0.1.0). It is not on npm — my registry access is blocked — so the install is the git spec, at the commit I measured:

dsh plugin --profile <your-profile> add "github:argszero/dsh-legacy-prompt-vars#808e73c2dd49208f373bbe6b077ba5d78fcbe6c8"

Measured on a clean profile against the 0.2.1-rc.3 runtime, today: the verb exits 0, and dsh --profile <same> --dump-config contains an entry dsh inserted itself:

# == @argszero/dsh-legacy-prompt-vars
- id: legacy-prompt-vars
  name: '@argszero/dsh-legacy-prompt-vars'

Usage is nothing beyond mounting it: it re-registers the retired variable, so a persona suffix that still says Your working directory is {{cwd}}. assembles again instead of failing every turn before the model is called.

What it does not do: it does not repair the layer that holds the stale text. The durable fix is to delete the {{cwd}} line from your profile patch or preset row — the shim only keeps turns alive while that line is still there. And it restores the retired names it knows; a variable I have never seen will still fail exactly as this one did.

I could not carry this as a core change: issues and pull requests are disabled on this repository, so a Discussion reply is the channel I have.