feat(adopt): connect Pi and oh-my-pi; list the shared skills folder

[fylorn]

What this changes

Pi and oh-my-pi are now clients that can be pointed at the gateway in one step. Pi (earendil-works/pi, formerly badlogic/pi-mono) gets a provider in ~/.pi/agent/models.json; its fork oh-my-pi (can1357/oh-my-pi, omp) gets the same provider in ~/.omp/agent/models.yml (or models.yaml when that is the file omp reads). The plan dialog shows the full diff, the file is backed up, and restore removes exactly what was written (a file created here is deleted again). Both are marked fields_only: the field names are checked against the source, nothing was run on this machine.

Pi, ~/.pi/agent/models.json (merged into the existing file; the record goes into the models.json.thinkwatch.json sidecar because the file is JSON):

{
  "providers": {
    "thinkwatch": {
      "name": "ThinkWatch",
      "baseUrl": "http://127.0.0.1:8788/v1",
      "api": "openai-completions",
      "apiKey": "tw-…",
      "models": [
        { "id": "claude-sonnet-5", "api": "anthropic-messages", "baseUrl": "http://127.0.0.1:8788" },
        { "id": "gpt-5.5", "api": "openai-responses" },
        { "id": "gemini-3-pro", "api": "google-generative-ai", "baseUrl": "http://127.0.0.1:8788/v1beta" },
        { "id": "deepseek-chat" }
      ]
    }
  }
}

oh-my-pi, ~/.omp/agent/models.yml (sentinel comment block at the top, as for the other YAML clients):

providers:
  thinkwatch:
    baseUrl: http://127.0.0.1:8788/v1
    api: openai-completions
    apiKey: tw-…
    auth: apiKey
    models:
      - id: claude-sonnet-5
        api: anthropic-messages
        baseUrl: http://127.0.0.1:8788
      - id: deepseek-chat

• A provider of its own (thinkwatch). Both clients keep credentials per provider, and /login stores them under built-in ids (anthropic, openai-codex, …). Rewriting a built-in provider's baseUrl would send those credentials to the gateway.
• The model list is written, like opencode (writes_models): the gateway's GET /v1/models for the client's key. Each model uses its family's API so a same-format upstream is passed through unconverted: Claude → anthropic-messages (base URL without /v1), Gemini → google-generative-ai (/v1beta), GPT / o-series / Codex → openai-responses, everything else → the provider default openai-completions (/v1). The Clients page flags a stale list and re-plans it through the same diff.
• The default model is left alone, as for opencode: settings.json (Pi) and config.yml (omp) are not touched. The plan says where the ThinkWatch models are chosen.
• oh-my-pi always says how it authenticates. Without auth, omp shapes custom anthropic-messages models as Claude Code (OAuth-style request shaping). auth: apiKey is written with a key, auth: none without one (omp rejects a provider with models and neither).
• Pi's key is written so it reads back literally. Pi interprets $NAME / ${NAME} and a leading ! in apiKey; $ is written as $$ and a leading ! as $!. Gateway keys (tw- + 24 base32 characters) never contain either, so in practice the value is unchanged. omp only interprets a leading ! and a value that is exactly an environment variable name, which a tw- key cannot be.
• Both reload without a restart: Pi re-reads models.json when /model opens, omp's model picker re-reads models.yml when its mtime changed. takes_effect: immediately, reloads: true.
• omp reads only one of models.yml / models.yaml, the first that exists. The takeover writes into that one; if models.yml appears later over a models.yaml takeover, the plan and the diagnosis say that nothing written here is used (new messages, not the generic "a setting of the same name wins").
• Env vars that move the config directory are diagnosed like CODEX_HOME: PI_CODING_AGENT_DIR (Pi); PI_CODING_AGENT_DIR, PI_CONFIG_DIR, OMP_PROFILE, PI_PROFILE (omp). Both clients are movable on the Clients page (layout: the agent directory).
• MCP: ~/.pi/agent/mcp.json and ~/.omp/agent/mcp.json (mcpServers) appear on the MCP page, read-only (adopt.mcp.unverified_format, no local sample). The scan covers both MCP files (plus omp's .mcp.json and its disabledServers / enabledServers lists), skills/, Pi's prompts/ and omp's commands/ + prompts/ as slash commands, and the instruction files read into context (AGENTS.md, AGENTS.override.md, CLAUDE.md, SYSTEM.md, APPEND_SYSTEM.md for Pi; AGENTS.md, SYSTEM.md, RULES.md for omp). Project level: .pi/ and .omp/ MCP and skills.
• Logos: Pi uses the Lobe Icons "Pi Agent" mark (MIT); oh-my-pi has no mark and gets the letter tile.

Caveats shown in the plan (new message codes, Chinese in core.zh.json):

• Pi: "The default model stays as it is; ThinkWatch models are chosen in /model, where Ctrl+S makes one the default." (adopt.cost.pi.default_model)
• oh-my-pi: "The default model stays as it is; ThinkWatch models are chosen in /model, where assigning the default role makes one the default." (adopt.cost.omp.default_model)
• Pi, only when it applies: "{name} is set and NO_PROXY does not list {host}, so Pi sends its requests for the gateway through that proxy. Adding {host} to NO_PROXY sends them directly." (adopt.plan.pi.proxy_env), or the same sentence for Pi's own httpProxy setting (adopt.plan.pi.proxy_setting). Pi uses undici's EnvHttpProxyAgent, which proxies every host, 127.0.0.1 included, when NO_PROXY is empty; the check follows undici's rules (http_proxy before HTTP_PROXY, an empty value means no proxy, https falls back to the HTTP proxy, NO_PROXY host/port/suffix/* matching). The proxy variables come from the login-shell environment: user_env still keeps them away from core, but now sets them aside for this check. omp needs no note: it bypasses the proxy for loopback and private ranges.
• The usual adopt.plan.fields_only note, and adopt.plan.no_models when the key has no models.

DeepSeek Harness desktop and web. Both read the home patch layer that the existing DSH takeover writes, so they were already covered (checked in 0.2.0-rc.2, see below). Added: the installed desktop app (DeepSeek Harness.app in /Applications or ~/Applications; %LOCALAPPDATA%\Programs\DeepSeek Harness\DeepSeek Harness.exe on Windows) marks DSH as installed even before ~/.dsh exists, which also makes its MCP column present. New caveat: "Models used through a DeepSeek account signed in to the desktop app go straight to api.deepseek.com and do not pass through the gateway." (adopt.cost.dsh.account_direct). The community dsh-desktop is out of scope.

Shared skills folder. ~/.agents/skills and a project's .agents/skills are listed under a shared folder (agents, shown as 共用目录 / Shared folder) instead of DeepSeek Harness and Antigravity CLI. Pi, oh-my-pi, DSH, agy, Copilot, Kimi, Goose, Crush, Kilo, Cline and MiMo read it.

Codex. OPENAI_BASE_URL is no longer among Codex's diagnostic env vars: openai/codex 4b8bab6 (2026-04-03, "Remove OPENAI_BASE_URL config fallback") stopped reading it, so an exported one was misreported as overriding the takeover.

YAML writer. tw_adopt::yaml::set can now replace a whole value (a list or a map) through tw_yaml::put (the same self-checked path core uses for config.yaml), and write a nested key into a file that has no nodes yet. get returns a container in one-line form so its original can be recorded. Strings inside written blocks are plain only when they are unambiguous (letter first, [A-Za-z0-9-_./:@+], not a bool/null/number lookalike), otherwise quoted.

What was checked in the sources

Pi at 0f8740b (coding-agent 0.99.2):

• Agent dir ~/.pi/agent, override PI_CODING_AGENT_DIR: config.ts L539-L568, models.json / settings.json L577-L589
• models.json schema (provider name/baseUrl/apiKey/api/models[], model id/name/api/baseUrl/…), comments stripped before parsing: model-config.ts L188-L244, L297
• A model's api and baseUrl override the provider's: provider-composer.ts L192-L205
• apiKey interpretation (!command, $NAME, $$, $!): resolve-config-value.ts L28-L86
• Credential order (--api-key, auth.json, models.json apiKey, env), /model reloads the file, Ctrl+S saves the default: docs/models.md L23-L66; the reload in code: model-selector.ts L158, L192 → model-runtime.ts L839-L840
• Default model: saved default if its provider has auth, else the first available model: model-resolver.ts L675-L705
• Anthropic: Claude Code identity only for sk-ant-oat OAuth tokens; an API key goes out as x-api-key with Pi's own User-Agent: anthropic-messages.ts L980-L1079; Google base URL already carries the version: google-generative-ai.ts L350-L366
• Proxy: http-dispatcher.ts L45-L100, applied in main.ts L594-L595; undici "Always proxy if NO_PROXY is not set or empty": env-http-proxy-agent.js L89-L96 (v8.11.2)
• MCP since 0.99.0 in ~/.pi/agent/mcp.json / .pi/mcp.json: CHANGELOG L80-L92, docs/mcp.md L32; skills incl. ~/.agents/skills: package-manager.ts L2458, L2549-L2561; agent-dir files: docs/configuration.md

oh-my-pi at 6e4ac4a (coding-agent 18.4.8):

• models.yml then models.yaml, only the first that exists is read; legacy models.json migration: docs/models.md L20-L33, config-file.ts L187-L195; agent dir and overrides: dirs.ts L1-L28, L594-L596
• auth omitted → custom anthropic-messages models are shaped as OAuth / Claude Code: custom-models.ts L51-L63, docs/models.md L146-L148
• Schema (provider has no name; model id/api/baseUrl): models-config-schema-bundle.ts L192-L345; apiKey required with models unless auth: none|oauth: models-config.ts L66-L80; key resolution (!command, exact env-var name, else literal): resolve-config-value.ts L100-L108
• Discovery applies the provider's single api (per-model only from supported_endpoint_types, which the gateway does not send), so the list is written instead: model-registry.ts L1607-L1614
• Reload: model-registry.ts L556-L567 called from model-picker.ts L199-L205; /model, the default role: docs/models.md L602, L678
• Loopback and private ranges bypass the proxy: proxy.ts L21-L66; base URLs: google.ts L36-L41, anthropic-client.ts L198, L249
• MCP ~/.omp/agent/mcp.json, disabledServers / enabledServers: docs/mcp-config.md L18-L21, L93-L97; skills incl. the agents provider: docs/skills.md L91-L122

DeepSeek Harness at tag dsh-v0.2.0-rc.2 (639ed01):

• The home patch is applied above the profile layers: profile-context.ts L63-L70; runProfile builds that context with home: resolveDshHome(): profile-boot.ts L68-L74, L287-L296; the desktop host runs runProfile({ profile: 'desktop' }): desktop-host/src/index.ts L25-L30, under the shared $DSH_HOME/profiles/desktop: desktop/src/paths.ts L12-L21
• The llm-deepseek row (now @deepseek-ai/dsh-llm-deepseek-api-key) still honours baseURL and apiKeyEnv: llm-deepseek-api-key config.ts L14-L17, llm-deepseek config.ts L83; the account route is a separate row llm-deepseek-account (README) whose base URL defaults to https://api.deepseek.com/anthropic (config.ts L106, L290) and whose token is handed out only for the allowed origin (llm-deepseek-account index.ts L17-L26, deepseek-account index.ts L95-L99)
• Packaging: productName: 'DeepSeek Harness', macOS dmg/zip, Windows NSIS per user: electron-builder-config.mjs L109, L164, L231-L242; electron-builder's per-user default $LocalAppData\Programs\${APP_FILENAME} with the product name: multiUser.nsh L28-L47, targetUtil.ts L41-L43

Pi logo: Lobe Icons pi.svg ("Pi Agent", pi.dev, MIT).

How it was verified

Nothing was installed or run; no third-party client, no login.

• cargo fmt --manifest-path src-tauri/Cargo.toml --all -- --check: clean
• cargo clippy --manifest-path src-tauri/Cargo.toml --all-targets -- -D warnings: clean
• cargo test --manifest-path src-tauri/Cargo.toml -p tw-adopt -p tw-scan: 237 + 17 + 55 + 31 + 31 passed
• cargo test --manifest-path src-tauri/Cargo.toml --lib: 337 passed (plus the two crates)
• cargo test --manifest-path src-tauri/Cargo.toml --test msg_codes --test ts_bindings: passed (msg-codes.txt and lite-api.ts regenerated)
• npx tsc --noEmit, npx tsc --noEmit -p scripts/shots: clean; npx vitest run: 583 passed

New tests: byte-identical restore of a Pi models.json with comments and another provider, and of an omp models.yml with comments and another provider; adopting omp twice is a no-op; rewriting the model list keeps the first record; files created here are removed again; omp writes into models.yaml when that is the file it reads and reports a later models.yml as hiding everything; the proxy note (undici precedence, NO_PROXY matching, Pi's httpProxy, shell-file fallback, nothing for omp); Pi key escaping checked against Pi's parsing rules; per-model API choice; stale model list flagged for Pi and omp on the Clients listing; YAML whole-value write/remove and quoting; scan sources and MCP listing for Pi and omp (incl. disabledServers), the shared skills folder; the DSH desktop app counted as installed; Codex no longer reports OPENAI_BASE_URL.

Not verified / open

• Pi and oh-my-pi were not run against the gateway (hence fields_only). The gateway's conversions for Pi's google-generative-ai and openai-responses requests are untested end to end.
• The DSH desktop app is only found at the default install locations; a Windows install moved to another directory is not (its InstallLocation registry value is not read). Linux has no published desktop build.
• omp migrates a legacy models.json into models.yml only while neither YAML file exists; a takeover on a machine where that migration has not run yet would create models.yml first and the old models.json would stay unmigrated. Not handled.
• Core's request-header hint does not recognise Pi or omp (the Traffic "app" column stays empty for them); that would be a core change.
• Screenshots were not retaken (scripts/shots/mock updated); the feature docs on thinkwat.ch were not touched.

🤖 Generated with Claude Code