@actionway/cli-dev 0.0.0-dev.32112453761.a2fa31914131

package/assets/skill/actionway/SKILL.md

@@ -0,0 +1,56 @@
+---
+name: actionway
+description: Use the local Actionway CLI to discover, inspect, price, and call Actionway capabilities; continue asynchronous jobs or confirmed calls; read Actionway account usage, wallet, and transactions; authenticate; or update the CLI. Trigger when a local Agent can complete a task through `actionway`, when a `job_ref` or `call_ref` must be continued, or when the user asks about Actionway tools or account data.
+metadata:
+  version: "0.0.0-dev"
+---
+
+# Actionway
+
+Use the installed `actionway` CLI and its server-backed Capability Registry as the source of truth. Do not memorize a complete command catalog, Schema, provider list, or price table in this Skill.
+
+## Prepare
+
+1. Run Actionway commands with network access and permission to read `~/.actionway`. On Windows, use the generated `actionway.cmd` entrypoint when PowerShell policy blocks `actionway.ps1`. If setup or update fails, run `actionway.cmd doctor` on Windows or `actionway doctor` elsewhere from the host terminal and preserve its failed check IDs; do not weaken registry, proxy, CA, PATH, or PowerShell policy without user approval. If a sandboxed command reports `E_NOT_AUTHENTICATED` for an already logged-in user, rerun the same command with host permission before starting login again.
+2. If the user supplied a pending `job_ref`, skip updating and immediately run `actionway wait --job-ref=<uuid>`.
+3. Otherwise, once per Agent session, run `actionway update --check`. Continue with the installed CLI if the check is unavailable. If an update is available, run `actionway update` before starting a new call, then refresh this Skill from the returned `skill_source` for the next Agent session.
+4. Compare this Skill's `metadata.version` with successful command output `cli_version`. If they differ, replace the installed Actionway Skill with the directory at `skill_source`, tell the user once that the refreshed instructions apply from the next session, and continue the current task. Follow any output `notes`, especially minimum-version instructions.
+5. Run `actionway whoami`. On a genuine `E_NOT_AUTHENTICATED`, run `actionway init`, ask the user only to approve access in the browser, and install the staged Skill from `skill.source`. Never request, print, copy, or expose credentials.
+
+## Choose the command path
+
+- For an unknown, ambiguous, or long-tail capability, run `actionway tools search "<natural-language intent>"`. Use its filters only when the task supplies real constraints. Do not dump or preload the whole catalog.
+- If Search returns `clarification_required`, ask its bounded question. If it returns candidates, choose only a supported `tool_ref` and run `actionway tools inspect <tool_ref>` before the generic call. Inspect is the authority for the current input Schema, variants, effects, guidance, and Actionway USD price.
+- `actionway search` is the Web Search business capability. Capability discovery is always `actionway tools search`.
+- Use a high-frequency typed command only when `actionway --help` exposes an exact fit. Read `actionway <command> --help` and, when available, `--print-schema`; typed commands are parameter/file conveniences over the same Tool Call, confirmation, billing, and Job path.
+- For account questions, use only the requested read command: `actionway account usage`, `actionway account wallet`, or `actionway account transactions`.
+
+## Inspect and call
+
+1. Build arguments that match the inspected Schema exactly. Use the inspected canonical variant ID, never a guessed provider/model alias. For small objects use `--input-json`; for large or generated objects use `--input-file` or stdin.
+2. Run one command at a time. A generic call is `actionway tools call <tool_ref> [--variant <id>] --input-json '<object>'`. Reuse one stable `--idempotency-key` when retrying the same request.
+3. Parse stdout as structured JSON. Preserve a failure's `code`, `message`, `hint`, and safe details. Do not call internal binaries, endpoints, or unsupported flags.
+4. Treat server-projected pricing as authoritative. Show the current USD quote and effects when the user asks about price or when confirmation is required. Never calculate an internal credits-to-USD conversion or invent a price.
+5. Handle call status exactly:
+   - `confirmation_required`: do not execute or resume automatically. Present the server plan's capability/variant, effects, Usage label, and USD quote, then ask the user. On rejection, stop. On changed inputs, make a new call. On approval, run only `actionway tools call --resume <call_ref>`.
+   - `accepted`: immediately wait on the returned `job.jobRef`.
+   - `completed`: consume the returned result.
+   - `confirmation: strong`: never treat chat text alone as approval; follow a server-verifiable approval instruction or report that execution cannot continue.
+6. On `E_INSUFFICIENT_BALANCE`, stop the call. When the failure includes `top_up_url`, the user-facing response must render that exact field value as a Markdown link by substituting it into `[Top up Actionway balance](<top_up_url>)`; never leave it only in terminal JSON, replace it with a terminal command, or synthesize a URL for another environment. Ask the user to retry the original task after the top-up is complete.
+
+## Handle asynchronous jobs
+
+1. Preserve the full UUID `job_ref` returned by submit.
+2. If a typed command returns pending, or a Tool Call returns `accepted`, immediately run `actionway wait --job-ref=<uuid>` in the same Agent turn.
+3. Let `actionway wait` poll internally. Do not create a background process or a shell `sleep`, `for`, or `while` polling loop.
+4. If stderr emits a `long_wait` event after ten minutes, tell the user once that processing continues and they may come back later. Continue waiting unless the user asks to stop.
+5. If the wait is interrupted or times out, run `actionway wait` again with the same `job_ref`. Use `actionway poll` only for a one-shot diagnostic snapshot.
+6. Consume the terminal payload returned by `actionway wait` as the result. Do not wait for a server callback or for a prior Agent session to resume.
+
+Do not resubmit the original capability call or typed command merely because a job is still pending. Keep using its original `job_ref`.
+
+## Deliver artifacts
+
+- Download terminal `assets` into an `outputs/` directory under the current task using stable, descriptive filenames.
+- Present the downloaded local files to the user. Do not return only remote URLs when a downloadable artifact is available.
+- If downloading fails, preserve the terminal payload and `job_ref`; never resubmit generation to recover a file transfer.

package/assets/skill/actionway/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Actionway"
+  short_description: "Discover and run Actionway capabilities"
+  default_prompt: "Use $actionway to discover the right capability, inspect its current schema and price, and complete this task with the local Actionway CLI."

package/assets/skill/actionway-dev/SKILL.md

@@ -0,0 +1,29 @@
+---
+name: actionway-dev
+description: Use the installed Actionway development channel to discover, inspect, price, and call capabilities against dev.actionway.ai; test confirmed calls and async jobs; or inspect the deployed dev CLI source SHA without touching the stable Actionway installation. Trigger for Actionway dev environment, staging verification, a newly merged dev deployment, or pre-production Codex testing.
+---
+
+# Actionway Dev
+
+Use only the `actionway-dev` executable. Never substitute the stable `actionway` command.
+
+## Prepare
+
+1. Run `actionway-dev update` once per Agent session. It follows public `@actionway/cli-dev@latest` from the latest successful `dev` deployment and never updates the stable `@actionway/cli` package.
+2. Run `actionway-dev status` when the user asks which source SHA is installed.
+3. Run `actionway-dev whoami`. On `E_NOT_AUTHENTICATED`, run `actionway-dev init` and ask the user only to approve browser login. Never request or expose credentials.
+
+## Discover and call
+
+1. For long-tail capabilities run `actionway-dev tools search "<intent>"`, then inspect exactly one supported result with `actionway-dev tools inspect <tool_ref>`.
+2. Match the inspected Schema exactly and call with `actionway-dev tools call <tool_ref> --input-json '<object>'`.
+3. On `confirmation_required`, present the server plan and quote. Resume only after approval with `actionway-dev tools call --resume <call_ref>`.
+4. On `accepted`, immediately run `actionway-dev wait --job-ref=<uuid>`. Reuse the same reference until terminal state; never resubmit a pending call.
+
+Use typed commands only when `actionway-dev --help` exposes an exact fit. Read command help and `--print-schema`, then run `--dry-run` before a real typed call.
+
+## Boundaries
+
+- Treat the development Gateway as the authority for Schema, price, confirmation, and results.
+- Development calls can consume development credits and real upstream vendor cost.
+- Do not run the stable CLI's updater, publish npm packages, or change production configuration.

package/assets/skill/actionway-dev/agents/openai.yaml

@@ -0,0 +1,4 @@
+interface:
+  display_name: "Actionway Dev"
+  short_description: "Use the isolated Actionway development channel"
+  default_prompt: "Use $actionway-dev to test a capability against the latest Actionway development deployment."