vidbyte-cli 0.3.3__tar.gz → 0.3.4__tar.gz

{vidbyte_cli-0.3.3/src/vidbyte_cli.egg-info → vidbyte_cli-0.3.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: vidbyte-cli
-Version: 0.3.3
+Version: 0.3.4
 Summary: Vidbyte CLI: start hosted Research harness runs, run local Codex coding-agent teams, and get ranked next actions from your terminal.
 Author-email: Vidbyte <contact@vidbyte.pro>
 License-Expression: MIT
@@ -36,9 +36,9 @@ Requires-Dist: pydantic<3,>=2.6
 Requires-Dist: PyNaCl<2,>=1.5
 Requires-Dist: rich<16,>=13.7
 Requires-Dist: x402[evm]==2.8.0
-Requires-Dist: vidbyte-sdk[codex]<0.5,>=0.4.1
+Requires-Dist: vidbyte-sdk[codex]<0.5,>=0.4.2
 Provides-Extra: codex
-Requires-Dist: vidbyte-sdk[codex]<0.5,>=0.4.1; extra == "codex"
+Requires-Dist: vidbyte-sdk[codex]<0.5,>=0.4.2; extra == "codex"
 Provides-Extra: mcp
 Requires-Dist: mcp<3,>=2.3; extra == "mcp"
 Provides-Extra: dev
@@ -221,8 +221,8 @@ them passed again.
 
 | Option | Default | Meaning and reach |
 |--------|---------|-------------------|
-| `--model MODEL` | installed Codex default (suggestion: `gpt-6-sol`) | The model id the agents run on. Suggestion: the generator only, while `--critic-model` keeps choosing the critic. Ensemble: every turn and fork. Persistence: the agent. Task board: every task agent. Stages: every stage whose `--stage-model` is omitted or empty. Bug finder: every hunt agent. Infinite memory: every agent of the tree, on the model's own provider. |
-| `--reasoning-effort` | provider default (suggestion: `high`; stages: `medium`) | `none`, `minimal`, `low`, `medium`, `high`, or `xhigh` for every agent the command starts; an empty value means unset. In stages it applies only when `--stage-effort` is omitted. Infinite memory refuses it. |
+| `--model MODEL` | installed Codex default (suggestion: the default of your Codex or Claude Code) | The model id the agents run on. Suggestion: every role, and it picks the host when `--host` is omitted, while `--critic-model` can choose the critic apart. Ensemble: every turn and fork. Persistence: the agent. Task board: every task agent. Stages: every stage whose `--stage-model` is omitted or empty. Bug finder: every hunt agent. Infinite memory: every agent of the tree, on the model's own provider. |
+| `--reasoning-effort` | provider default (stages: `medium`) | `none`, `minimal`, `low`, `medium`, `high`, or `xhigh` for every agent the command starts; an empty value means unset. In stages it applies only when `--stage-effort` is omitted. Infinite memory refuses it. |
 | `--instructions FILE` | none | Repeatable. Each Markdown or text file is appended, after a blank line, to the system prompt of every agent the command starts, ensemble forks and suggestion critics included. An ensemble fork also gets the text at the end of its first turn. At most 8 non-empty UTF-8 files and 20,000 bytes in total; errors name the file's position, never its path or contents. |
 | `--send-to TARGET` | none | Repeatable. `slack` posts to the webhook in `SLACK_WEBHOOK_URL`; `slack:NAME` to the one saved with `slack add NAME`; `github:OWNER/NAME` opens an issue with `GITHUB_TOKEN`, else the `github link` account; `linear` or `linear:TEAM` opens an issue with the `linear login` key, in the default team unless TEAM is given. The result is sent after a normal finish, with one receipt line per target on stderr and stdout unchanged. A failed send exits 1 after every target was tried, with the result still on stdout. A dry run or prompt preview sends nothing. Needs a vidbyte-sdk release with the output connectors. |
 
@@ -236,28 +236,27 @@ The [Suggestion Agent](https://vidbyte.pro/agents/suggestion-agent) drafts next
 goal, has an independent critic review them, and returns the survivors ranked, each with a handoff.
 
 `agents suggest run` takes one nonempty `--goal` plus optional repeated context and `--files`
-values. It takes no JSON input file. Both roles default to `gpt-6-sol` on the `openai` provider at
-`high` reasoning effort: `--model` sets the generator's model, including its extra-compute
-generators, `--critic-model` independently sets the reviewer's, and `--reasoning-effort` replaces
-the effort for every role. `--count` accepts 2–100, `--rounds` accepts 1–8, `--max-output-tokens`
-caps each reply at up to 128,000 tokens (default 32,000), and `--extra-compute` runs one focused
-generator context per selected category before critique. `--dry-run`, `categories`, `handoff`,
-`project`, and `feedback` never call a model, so they need no key and are never charged.
-
-The suggestion agent is usage-based: there is no per-suggestion price. Every generator and
-critic turn is one call to Vidbyte's model gateway (`POST /api/v1/models/generate`), billed to
-the balance of your stored Vidbyte key at that model's rate in the Vidbyte SDK pricebook plus
-2%. A default run makes about five calls: a first draft, then a critique and a revision per
-round. The key needs the `models:invoke` permission (`vidbyte-cli login` with a key created at
-<https://vidbyte.pro/settings/api>); no Codex install, OpenAI login, or OpenAI key is involved.
-The goal, its context, `--files` bodies, attachments, and loaded project memory are sent through
-Vidbyte to OpenAI on each call. The result's `usage` field reports the model calls and the
-input, output, and cached tokens the run used; the billed amounts appear at
-<https://vidbyte.pro/usage>. `--max-total-tokens` and `--timeout-seconds` are checked before
-each call, never mid-call, because a call in flight is billed when it completes. If the balance
-or the key's daily model-spend cap runs out after a slate exists, the run stops with
-`gateway_limit` and returns the latest reviewed slate; before that, it fails with the
-balance or rate-limit error.
+values. `--host codex|claude` picks the coding agent; omitted, a `claude-*` `--model` picks
+`claude`, an OpenAI `--model` or `--provider openai` picks `codex`, else the first installed of
+the two. `--model` sets every role, `--critic-model` the reviewer apart; omitted, each keeps your
+agent's own default, as does `--reasoning-effort` (Claude Code takes `low` to `xhigh`). `--count`
+is 2–100, `--rounds` 1–8, `--max-output-tokens` up to 128,000 per reply, and
+`--extra-compute` runs one generator per category, four at a time. `--dry-run`, `categories`,
+`handoff`, `project`, and `feedback` start no agent and buy nothing.
+
+The Suggestion Agent runs on your own coding agent, Codex or Claude Code, on your machine, so
+OpenAI or Anthropic bills its model usage to your own account. Vidbyte charges 1 cent per
+requested suggestion from your prepaid balance before the run starts (`--count 5` costs 5
+cents). Dry runs are free. Vidbyte never sees your goal, context, or ideas. The charge is one
+`runtime.suggestion@2` admission with `units` = `--count`, bought after every free check (host,
+effort, SDK, catalog price) and verified before the first turn (`charged_cents` = `units`); the
+retry key prints to stderr first, and `--idempotency-key` recovers a lost reply without buying
+twice. A slate the review trims still costs the requested count. The key needs
+`runtime:write`; Claude Code needs vidbyte-sdk 0.4.2 or later. Both hosts run read-only in an
+empty scratch folder and never receive your Vidbyte key. The result reports `admission`,
+`charged_cents`, `host`, `generator_thread_id`, and `host_usage` (`billed_by` OpenAI or
+Anthropic, `billed_to: "your_own_account"`). Each turn is capped at 30 minutes; if your agent
+fails after a slate exists, the run returns it as `provider_failed`.
 
 Context is limited to 5,000,000 characters per item and in aggregate, with truncation or omission
 recorded in the result manifest. Each surviving idea carries one primary category, evidence
@@ -509,7 +508,7 @@ plan included, can create an API key at <https://vidbyte.pro/settings/api>.
 | --- | --- | --- |
 | `vidbyte_setup_status` | Reports which keys and hosts are missing | Free, offline |
 | `vidbyte_plan(request)` | Asks Jev one yes/no question per agent and stores a 30-minute plan | One metered Jev call billed to your Vidbyte balance |
-| `vidbyte_run(plan_id, request_id)` | Runs the plan's recommended (or another passing) agent; your client asks you first | The agent's own price: ensemble, persistence, and task board charge a two-cent admission plus OpenAI usage; suggestion bills each model call to your Vidbyte balance at its model's rate plus 2%; research is usage-based up to $5.00 |
+| `vidbyte_run(plan_id, request_id)` | Runs the plan's recommended (or another passing) agent; your client asks you first | The agent's own price: ensemble, persistence, and task board charge a two-cent admission plus OpenAI usage; suggestion runs on your own Codex or Claude Code and charges 1 cent per requested suggestion (5 cents at the default count); research is usage-based up to $5.00 |
 | `vidbyte_run_status(job_id)` | Reports a job's status and pages its result | Free, offline |
 | `vidbyte_cancel_run(job_id)` | Stops a running job | Free; an admission or completed model calls are not refunded |
 

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/README.md

@@ -170,8 +170,8 @@ them passed again.
 
 | Option | Default | Meaning and reach |
 |--------|---------|-------------------|
-| `--model MODEL` | installed Codex default (suggestion: `gpt-6-sol`) | The model id the agents run on. Suggestion: the generator only, while `--critic-model` keeps choosing the critic. Ensemble: every turn and fork. Persistence: the agent. Task board: every task agent. Stages: every stage whose `--stage-model` is omitted or empty. Bug finder: every hunt agent. Infinite memory: every agent of the tree, on the model's own provider. |
-| `--reasoning-effort` | provider default (suggestion: `high`; stages: `medium`) | `none`, `minimal`, `low`, `medium`, `high`, or `xhigh` for every agent the command starts; an empty value means unset. In stages it applies only when `--stage-effort` is omitted. Infinite memory refuses it. |
+| `--model MODEL` | installed Codex default (suggestion: the default of your Codex or Claude Code) | The model id the agents run on. Suggestion: every role, and it picks the host when `--host` is omitted, while `--critic-model` can choose the critic apart. Ensemble: every turn and fork. Persistence: the agent. Task board: every task agent. Stages: every stage whose `--stage-model` is omitted or empty. Bug finder: every hunt agent. Infinite memory: every agent of the tree, on the model's own provider. |
+| `--reasoning-effort` | provider default (stages: `medium`) | `none`, `minimal`, `low`, `medium`, `high`, or `xhigh` for every agent the command starts; an empty value means unset. In stages it applies only when `--stage-effort` is omitted. Infinite memory refuses it. |
 | `--instructions FILE` | none | Repeatable. Each Markdown or text file is appended, after a blank line, to the system prompt of every agent the command starts, ensemble forks and suggestion critics included. An ensemble fork also gets the text at the end of its first turn. At most 8 non-empty UTF-8 files and 20,000 bytes in total; errors name the file's position, never its path or contents. |
 | `--send-to TARGET` | none | Repeatable. `slack` posts to the webhook in `SLACK_WEBHOOK_URL`; `slack:NAME` to the one saved with `slack add NAME`; `github:OWNER/NAME` opens an issue with `GITHUB_TOKEN`, else the `github link` account; `linear` or `linear:TEAM` opens an issue with the `linear login` key, in the default team unless TEAM is given. The result is sent after a normal finish, with one receipt line per target on stderr and stdout unchanged. A failed send exits 1 after every target was tried, with the result still on stdout. A dry run or prompt preview sends nothing. Needs a vidbyte-sdk release with the output connectors. |
 
@@ -185,28 +185,27 @@ The [Suggestion Agent](https://vidbyte.pro/agents/suggestion-agent) drafts next
 goal, has an independent critic review them, and returns the survivors ranked, each with a handoff.
 
 `agents suggest run` takes one nonempty `--goal` plus optional repeated context and `--files`
-values. It takes no JSON input file. Both roles default to `gpt-6-sol` on the `openai` provider at
-`high` reasoning effort: `--model` sets the generator's model, including its extra-compute
-generators, `--critic-model` independently sets the reviewer's, and `--reasoning-effort` replaces
-the effort for every role. `--count` accepts 2–100, `--rounds` accepts 1–8, `--max-output-tokens`
-caps each reply at up to 128,000 tokens (default 32,000), and `--extra-compute` runs one focused
-generator context per selected category before critique. `--dry-run`, `categories`, `handoff`,
-`project`, and `feedback` never call a model, so they need no key and are never charged.
-
-The suggestion agent is usage-based: there is no per-suggestion price. Every generator and
-critic turn is one call to Vidbyte's model gateway (`POST /api/v1/models/generate`), billed to
-the balance of your stored Vidbyte key at that model's rate in the Vidbyte SDK pricebook plus
-2%. A default run makes about five calls: a first draft, then a critique and a revision per
-round. The key needs the `models:invoke` permission (`vidbyte-cli login` with a key created at
-<https://vidbyte.pro/settings/api>); no Codex install, OpenAI login, or OpenAI key is involved.
-The goal, its context, `--files` bodies, attachments, and loaded project memory are sent through
-Vidbyte to OpenAI on each call. The result's `usage` field reports the model calls and the
-input, output, and cached tokens the run used; the billed amounts appear at
-<https://vidbyte.pro/usage>. `--max-total-tokens` and `--timeout-seconds` are checked before
-each call, never mid-call, because a call in flight is billed when it completes. If the balance
-or the key's daily model-spend cap runs out after a slate exists, the run stops with
-`gateway_limit` and returns the latest reviewed slate; before that, it fails with the
-balance or rate-limit error.
+values. `--host codex|claude` picks the coding agent; omitted, a `claude-*` `--model` picks
+`claude`, an OpenAI `--model` or `--provider openai` picks `codex`, else the first installed of
+the two. `--model` sets every role, `--critic-model` the reviewer apart; omitted, each keeps your
+agent's own default, as does `--reasoning-effort` (Claude Code takes `low` to `xhigh`). `--count`
+is 2–100, `--rounds` 1–8, `--max-output-tokens` up to 128,000 per reply, and
+`--extra-compute` runs one generator per category, four at a time. `--dry-run`, `categories`,
+`handoff`, `project`, and `feedback` start no agent and buy nothing.
+
+The Suggestion Agent runs on your own coding agent, Codex or Claude Code, on your machine, so
+OpenAI or Anthropic bills its model usage to your own account. Vidbyte charges 1 cent per
+requested suggestion from your prepaid balance before the run starts (`--count 5` costs 5
+cents). Dry runs are free. Vidbyte never sees your goal, context, or ideas. The charge is one
+`runtime.suggestion@2` admission with `units` = `--count`, bought after every free check (host,
+effort, SDK, catalog price) and verified before the first turn (`charged_cents` = `units`); the
+retry key prints to stderr first, and `--idempotency-key` recovers a lost reply without buying
+twice. A slate the review trims still costs the requested count. The key needs
+`runtime:write`; Claude Code needs vidbyte-sdk 0.4.2 or later. Both hosts run read-only in an
+empty scratch folder and never receive your Vidbyte key. The result reports `admission`,
+`charged_cents`, `host`, `generator_thread_id`, and `host_usage` (`billed_by` OpenAI or
+Anthropic, `billed_to: "your_own_account"`). Each turn is capped at 30 minutes; if your agent
+fails after a slate exists, the run returns it as `provider_failed`.
 
 Context is limited to 5,000,000 characters per item and in aggregate, with truncation or omission
 recorded in the result manifest. Each surviving idea carries one primary category, evidence
@@ -458,7 +457,7 @@ plan included, can create an API key at <https://vidbyte.pro/settings/api>.
 | --- | --- | --- |
 | `vidbyte_setup_status` | Reports which keys and hosts are missing | Free, offline |
 | `vidbyte_plan(request)` | Asks Jev one yes/no question per agent and stores a 30-minute plan | One metered Jev call billed to your Vidbyte balance |
-| `vidbyte_run(plan_id, request_id)` | Runs the plan's recommended (or another passing) agent; your client asks you first | The agent's own price: ensemble, persistence, and task board charge a two-cent admission plus OpenAI usage; suggestion bills each model call to your Vidbyte balance at its model's rate plus 2%; research is usage-based up to $5.00 |
+| `vidbyte_run(plan_id, request_id)` | Runs the plan's recommended (or another passing) agent; your client asks you first | The agent's own price: ensemble, persistence, and task board charge a two-cent admission plus OpenAI usage; suggestion runs on your own Codex or Claude Code and charges 1 cent per requested suggestion (5 cents at the default count); research is usage-based up to $5.00 |
 | `vidbyte_run_status(job_id)` | Reports a job's status and pages its result | Free, offline |
 | `vidbyte_cancel_run(job_id)` | Stops a running job | Free; an admission or completed model calls are not refunded |
 

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "vidbyte-cli"
-version = "0.3.3"
+version = "0.3.4"
 description = "Vidbyte CLI: start hosted Research harness runs, run local Codex coding-agent teams, and get ranked next actions from your terminal."
 readme = "README.md"
 requires-python = ">=3.11"
@@ -57,22 +57,24 @@ dependencies = [
     # Styles the interactive login prompts; everything it renders goes to stderr.
     "rich>=13.7,<16",
     "x402[evm]==2.8.0",
-    # 0.4.1 is the floor: on 0.4.0, `agents infinite-memory` and JevAgent's swarm fail with a 400
-    # on the second call to a Claude thinking model, because a signed thinking block is replayed
-    # after its system prompt or tools changed (SDK #1204, #1207). 0.4.0 brought the native
+    # 0.4.2 is the floor: it adds ClaudeCodeHarnessAgent and the other ClaudeCode* names that
+    # `agents suggest --host claude` runs every turn on (SDK #1210). 0.4.1 fixed the 400 that
+    # `agents infinite-memory` and JevAgent's swarm hit on the second call to a Claude thinking
+    # model, when a signed thinking block was replayed after its system prompt or tools
+    # changed (SDK #1204, #1207). 0.4.0 brought the native
     # opencode, Grok, and Muse thread functions that `resolve run` calls and the InfiniteEvent and
     # InfiniteHandoff names that `agents infinite-memory` binds; 0.3.0 the output connectors that
     # `--send-to` delivers through; and 0.2.1 openai-codex 0.162, which releases a cancelled
     # turn's reader. The SDK is alpha and may change APIs between minor versions, so stay within
     # 0.4.x until a newer minor is verified against this CLI.
-    "vidbyte-sdk[codex]>=0.4.1,<0.5",
+    "vidbyte-sdk[codex]>=0.4.2,<0.5",
 ]
 
 [project.optional-dependencies]
 # Kept so `pip install "vidbyte-cli[codex]"` keeps resolving; the base dependencies above
 # already include the SDK's Codex integration.
 codex = [
-    "vidbyte-sdk[codex]>=0.4.1,<0.5",
+    "vidbyte-sdk[codex]>=0.4.2,<0.5",
 ]
 # The local MCP server behind `vidbyte-cli mcp serve`. Optional so the base CLI stays small;
 # pinned below 3 because the protocol types track MCP specification revisions.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/prompts/list.md

@@ -2,4 +2,4 @@
 
 ## Purpose
 
-Show the runnable agent workflows included with this CLI. Each entry gives the command shape and a plain-language description, so you can choose a workflow before supplying a task. The catalog is local product metadata and does not contact Vidbyte, use a provider, or start an agent. Run the command shown beside a workflow to open its help and see the accepted options. Some workflows require Codex, provider credentials, or a Vidbyte admission fee, and the suggestion workflow bills each of its model calls to your Vidbyte balance, so check their command help before starting a run.
+Show the runnable agent workflows included with this CLI. Each entry gives the command shape and a plain-language description, so you can choose a workflow before supplying a task. The catalog is local product metadata and does not contact Vidbyte, use a provider, or start an agent. Run the command shown beside a workflow to open its help and see the accepted options. Some workflows require Codex, provider credentials, or a Vidbyte admission fee, and the suggestion workflow runs on your own Codex or Claude Code and charges 1 cent per requested suggestion, so check their command help before starting a run.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/prompts/model.md

@@ -1,10 +1,10 @@
 # `--model MODEL`
 
 ## Purpose
-Names the model that the agents of this command run on, given as the provider's own model identifier. On the Codex-hosted runtime commands, leaving it unset keeps the model the installed Codex host is configured with, which is the right choice when that model works for your login; on agents suggest run, leaving it unset uses gpt-6-sol. A blank value means unset, surrounding spaces are removed, and the identifier may be at most 128 characters long. Codex checks whether your account can run the model only once the agents start, so on a paid runtime command an unknown identifier fails after the admission is charged. On agents suggest run, Vidbyte's model gateway or the provider refuses an unknown identifier at the first model call, which fails with a model error that names this option.
+Names the model that the agents of this command run on, given as the provider's own model identifier. On the Codex-hosted runtime commands, leaving it unset keeps the model the installed Codex host is configured with, which is the right choice when that model works for your login; on agents suggest run, leaving it unset keeps the model your Codex is configured with, or Claude Code's built-in default model, because the isolated Claude Code session loads none of your Claude settings. A blank value means unset, surrounding spaces are removed, and the identifier may be at most 128 characters long. Codex checks whether your account can run the model only once the agents start, so on a paid runtime command an unknown identifier fails after the admission is charged. On agents suggest run, a claude-* model picks the Claude Code host and an OpenAI model picks Codex, and the provider refuses an unknown identifier at the first turn, after the admission, with a model error that names this option.
 
 ## Where it applies
 On agents suggest run it reaches the generator, every extra-compute generator agent, and the critic, unless --critic-model names a separate model for the critic. On runtime same-host-ensemble it sets the model of the planner and every fork. On runtime persistence it reaches the one agent, and on runtime task-board run it reaches every task agent. On runtime stages run it reaches every stage whose --stage-model is omitted or blank, while a given --stage-model still wins for its own stage. On agents bug-finder run it reaches every hunt agent. On agents infinite-memory run and resume it reaches every agent of the tree, defaults to gpt-6-sol, runs on the provider whose Vidbyte pricebook row lists the model (a Claude model runs on Anthropic), and is billed through Vidbyte's model gateway at the model's rate plus 2%. Pass only the model every agent should run on; a model meant as a backup would become the primary model of the whole run.
 
 ## Cost and recovery
-On the Codex-hosted runtime commands, model usage is billed to your own OpenAI account rather than to the flat Vidbyte admission; if Codex reports the model as unsupported, run codex login status to see which account is active, then pass a --model value that this login supports. On agents suggest run, every call goes through Vidbyte's model gateway and is billed to your Vidbyte balance at that model's pricebook rate plus 2%, so a pricier model raises the run's cost and no OpenAI login is involved.
+On the Codex-hosted runtime commands, model usage is billed to your own OpenAI account rather than to the flat Vidbyte admission; if Codex reports the model as unsupported, run codex login status to see which account is active, then pass a --model value that this login supports. On agents suggest run, every turn runs on your own Codex or Claude Code, so OpenAI or Anthropic bills the model usage to your own account, and Vidbyte charges only 1 cent per requested suggestion whatever the model.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/prompts/reasoning_effort.md

@@ -1,10 +1,10 @@
 # `--reasoning-effort`
 
 ## Purpose
-Sets how much reasoning each agent turn asks the model to spend before it answers. The accepted values are the choices shown in brackets after the option, from the lowest effort to the highest, and an empty string, the blank first choice, means the option is unset. Leaving it unset keeps the provider default for every agent, except on agents suggest run, whose model calls default to high, and on runtime stages run, whose stages default to medium. On agents suggest run, none turns reasoning off, and the default model, gpt-6-sol, refuses minimal at the first model call.
+Sets how much reasoning each agent turn asks the model to spend before it answers. The accepted values are the choices shown in brackets after the option, from the lowest effort to the highest, and an empty string, the blank first choice, means the option is unset. Leaving it unset keeps the provider default for every agent, except on runtime stages run, whose stages default to medium. On agents suggest run with the Claude Code host only low, medium, high, and xhigh are accepted, so none or minimal fails there before anything is charged.
 
 ## Where it applies
 It reaches every agent the command starts: every suggestion role including the critic, every ensemble turn and fork, the persistence agent, every task-board task agent, every stage, and every bug-finder hunt agent. Agents infinite-memory run and resume refuse it before anything starts, because the infinite memory agent has no reasoning effort setting. On runtime stages run a given --stage-effort wins, and this option applies only when --stage-effort is omitted entirely.
 
 ## Cost
-Higher effort spends more tokens on every turn of every agent. On the Codex-hosted runtime commands that usage is billed to your own OpenAI account; on agents suggest run every token is billed to your Vidbyte balance through the model gateway, at the model's pricebook rate plus 2%. A long board or a large ensemble at xhigh is a real cost decision rather than a tuning knob, while minimal or low suits long runs of mechanical work.
+Higher effort spends more tokens on every turn of every agent. On the Codex-hosted runtime commands that usage is billed to your own OpenAI account; on agents suggest run it is billed by OpenAI or Anthropic to your own account, because every turn runs on your own Codex or Claude Code. A long board or a large ensemble at xhigh is a real cost decision rather than a tuning knob, while minimal or low suits long runs of mechanical work.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/suggestion/prompts/README.md

@@ -9,3 +9,4 @@ When adding an option, add its help asset, wire its loading in `suggest.py`, and
 - 2026-10-07: The local Codex config can choose a model unsupported by a ChatGPT login. `--model` lets the caller select an available generator model without changing machine-wide settings; see issue #108.
 - 2026-10-10: `model.md` was removed. `--model` became one of the universal agent options that `AgentCommand` attaches to every agent command, so its help moved to `commands/agents/prompts/model.md`, which keeps the generator-only reach, the critic's independence, and the unsupported-model recovery.
 - 2026-10-11: `idempotency_key.md` was removed with the suggestion admission. Each model call is now billed through Vidbyte's model gateway, so the cost-bearing assets (`count`, `provider`, `max_output_tokens`, `max_total_tokens`, `timeout`, `dry_run`, `run`) describe per-call usage billing at the model's rate plus 2%, and the shared permissions block names the Vidbyte key rather than a read-only Codex sandbox.
+- 2026-10-30: `idempotency_key.md` is back with the `runtime.suggestion@2` admission, and `host.md` is new for `--host codex|claude`. Every turn runs on the caller's own Codex or Claude Code, whose provider bills the model usage to the caller's own account, and Vidbyte charges 1 cent per requested suggestion before the run starts. The cost-bearing assets and the shared permissions block say so. C015 holds every price these assets state to the gate's one cent, so they say "the charge in cents equals --count" rather than quoting multiples.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/suggestion/prompts/approaches.md

@@ -93,8 +93,8 @@ Recover by correcting the specific value named by the error, rerunning dry-run w
 
 **Authentication and permissions for Approach History**
 This input itself does not grant access to files, providers, accounts, repositories, or execution tools.
-Text values are available to local validation without credentials, while model-backed runs need a stored Vidbyte key with the models:invoke permission.
+Text values are available to local validation without credentials, while a real run needs a stored Vidbyte key with the runtime:write permission and your own signed-in Codex or Claude Code.
 The --files option reads only paths explicitly named by the caller and does not scan a workspace or agent history.
-Generator and critic agents have no file or shell tools and see only the text the CLI sends through Vidbyte to the model provider, so suggestions cannot mutate the host.
-The stored Vidbyte key is used only when a real model-backed run calls the model gateway, and it is never sent to a provider or copied into result prose.
-Permission failures should be repaired at the key, balance, or filesystem boundary rather than by weakening the context contract.
+Generator and critic agents run read-only on your own coding agent in an empty scratch folder and see only what this command sends them, so suggestions cannot mutate the host.
+The stored Vidbyte key only buys the run's admission at 1 cent per requested suggestion, and it is never sent to OpenAI or Anthropic or copied into result prose.
+Permission failures should be repaired at the key, balance, coding-agent login, or filesystem boundary rather than by weakening the context contract.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/suggestion/prompts/avoid.md

@@ -93,8 +93,8 @@ Recover by correcting the specific value named by the error, rerunning dry-run w
 
 **Authentication and permissions for Directions to Avoid**
 This input itself does not grant access to files, providers, accounts, repositories, or execution tools.
-Text values are available to local validation without credentials, while model-backed runs need a stored Vidbyte key with the models:invoke permission.
+Text values are available to local validation without credentials, while a real run needs a stored Vidbyte key with the runtime:write permission and your own signed-in Codex or Claude Code.
 The --files option reads only paths explicitly named by the caller and does not scan a workspace or agent history.
-Generator and critic agents have no file or shell tools and see only the text the CLI sends through Vidbyte to the model provider, so suggestions cannot mutate the host.
-The stored Vidbyte key is used only when a real model-backed run calls the model gateway, and it is never sent to a provider or copied into result prose.
-Permission failures should be repaired at the key, balance, or filesystem boundary rather than by weakening the context contract.
+Generator and critic agents run read-only on your own coding agent in an empty scratch folder and see only what this command sends them, so suggestions cannot mutate the host.
+The stored Vidbyte key only buys the run's admission at 1 cent per requested suggestion, and it is never sent to OpenAI or Anthropic or copied into result prose.
+Permission failures should be repaired at the key, balance, coding-agent login, or filesystem boundary rather than by weakening the context contract.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/suggestion/prompts/blockers.md

@@ -93,8 +93,8 @@ Recover by correcting the specific value named by the error, rerunning dry-run w
 
 **Authentication and permissions for Blockers**
 This input itself does not grant access to files, providers, accounts, repositories, or execution tools.
-Text values are available to local validation without credentials, while model-backed runs need a stored Vidbyte key with the models:invoke permission.
+Text values are available to local validation without credentials, while a real run needs a stored Vidbyte key with the runtime:write permission and your own signed-in Codex or Claude Code.
 The --files option reads only paths explicitly named by the caller and does not scan a workspace or agent history.
-Generator and critic agents have no file or shell tools and see only the text the CLI sends through Vidbyte to the model provider, so suggestions cannot mutate the host.
-The stored Vidbyte key is used only when a real model-backed run calls the model gateway, and it is never sent to a provider or copied into result prose.
-Permission failures should be repaired at the key, balance, or filesystem boundary rather than by weakening the context contract.
+Generator and critic agents run read-only on your own coding agent in an empty scratch folder and see only what this command sends them, so suggestions cannot mutate the host.
+The stored Vidbyte key only buys the run's admission at 1 cent per requested suggestion, and it is never sent to OpenAI or Anthropic or copied into result prose.
+Permission failures should be repaired at the key, balance, coding-agent login, or filesystem boundary rather than by weakening the context contract.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/suggestion/prompts/capability.md

@@ -93,8 +93,8 @@ Recover by correcting the specific value named by the error, rerunning dry-run w
 
 **Authentication and permissions for Capabilities**
 This input itself does not grant access to files, providers, accounts, repositories, or execution tools.
-Text values are available to local validation without credentials, while model-backed runs need a stored Vidbyte key with the models:invoke permission.
+Text values are available to local validation without credentials, while a real run needs a stored Vidbyte key with the runtime:write permission and your own signed-in Codex or Claude Code.
 The --files option reads only paths explicitly named by the caller and does not scan a workspace or agent history.
-Generator and critic agents have no file or shell tools and see only the text the CLI sends through Vidbyte to the model provider, so suggestions cannot mutate the host.
-The stored Vidbyte key is used only when a real model-backed run calls the model gateway, and it is never sent to a provider or copied into result prose.
-Permission failures should be repaired at the key, balance, or filesystem boundary rather than by weakening the context contract.
+Generator and critic agents run read-only on your own coding agent in an empty scratch folder and see only what this command sends them, so suggestions cannot mutate the host.
+The stored Vidbyte key only buys the run's admission at 1 cent per requested suggestion, and it is never sent to OpenAI or Anthropic or copied into result prose.
+Permission failures should be repaired at the key, balance, coding-agent login, or filesystem boundary rather than by weakening the context contract.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/suggestion/prompts/completed.md

@@ -93,8 +93,8 @@ Recover by correcting the specific value named by the error, rerunning dry-run w
 
 **Authentication and permissions for Completed Work**
 This input itself does not grant access to files, providers, accounts, repositories, or execution tools.
-Text values are available to local validation without credentials, while model-backed runs need a stored Vidbyte key with the models:invoke permission.
+Text values are available to local validation without credentials, while a real run needs a stored Vidbyte key with the runtime:write permission and your own signed-in Codex or Claude Code.
 The --files option reads only paths explicitly named by the caller and does not scan a workspace or agent history.
-Generator and critic agents have no file or shell tools and see only the text the CLI sends through Vidbyte to the model provider, so suggestions cannot mutate the host.
-The stored Vidbyte key is used only when a real model-backed run calls the model gateway, and it is never sent to a provider or copied into result prose.
-Permission failures should be repaired at the key, balance, or filesystem boundary rather than by weakening the context contract.
+Generator and critic agents run read-only on your own coding agent in an empty scratch folder and see only what this command sends them, so suggestions cannot mutate the host.
+The stored Vidbyte key only buys the run's admission at 1 cent per requested suggestion, and it is never sent to OpenAI or Anthropic or copied into result prose.
+Permission failures should be repaired at the key, balance, coding-agent login, or filesystem boundary rather than by weakening the context contract.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/suggestion/prompts/constraint.md

@@ -93,8 +93,8 @@ Recover by correcting the specific value named by the error, rerunning dry-run w
 
 **Authentication and permissions for Constraints**
 This input itself does not grant access to files, providers, accounts, repositories, or execution tools.
-Text values are available to local validation without credentials, while model-backed runs need a stored Vidbyte key with the models:invoke permission.
+Text values are available to local validation without credentials, while a real run needs a stored Vidbyte key with the runtime:write permission and your own signed-in Codex or Claude Code.
 The --files option reads only paths explicitly named by the caller and does not scan a workspace or agent history.
-Generator and critic agents have no file or shell tools and see only the text the CLI sends through Vidbyte to the model provider, so suggestions cannot mutate the host.
-The stored Vidbyte key is used only when a real model-backed run calls the model gateway, and it is never sent to a provider or copied into result prose.
-Permission failures should be repaired at the key, balance, or filesystem boundary rather than by weakening the context contract.
+Generator and critic agents run read-only on your own coding agent in an empty scratch folder and see only what this command sends them, so suggestions cannot mutate the host.
+The stored Vidbyte key only buys the run's admission at 1 cent per requested suggestion, and it is never sent to OpenAI or Anthropic or copied into result prose.
+Permission failures should be repaired at the key, balance, coding-agent login, or filesystem boundary rather than by weakening the context contract.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/suggestion/prompts/context.md

@@ -93,8 +93,8 @@ Recover by correcting the specific value named by the error, rerunning dry-run w
 
 **Authentication and permissions for General Context**
 This input itself does not grant access to files, providers, accounts, repositories, or execution tools.
-Text values are available to local validation without credentials, while model-backed runs need a stored Vidbyte key with the models:invoke permission.
+Text values are available to local validation without credentials, while a real run needs a stored Vidbyte key with the runtime:write permission and your own signed-in Codex or Claude Code.
 The --files option reads only paths explicitly named by the caller and does not scan a workspace or agent history.
-Generator and critic agents have no file or shell tools and see only the text the CLI sends through Vidbyte to the model provider, so suggestions cannot mutate the host.
-The stored Vidbyte key is used only when a real model-backed run calls the model gateway, and it is never sent to a provider or copied into result prose.
-Permission failures should be repaired at the key, balance, or filesystem boundary rather than by weakening the context contract.
+Generator and critic agents run read-only on your own coding agent in an empty scratch folder and see only what this command sends them, so suggestions cannot mutate the host.
+The stored Vidbyte key only buys the run's admission at 1 cent per requested suggestion, and it is never sent to OpenAI or Anthropic or copied into result prose.
+Permission failures should be repaired at the key, balance, coding-agent login, or filesystem boundary rather than by weakening the context contract.

vidbyte_cli-0.3.4/src/vidbyte_cli/commands/agents/suggestion/prompts/count.md

@@ -0,0 +1,5 @@
+The requested count is the maximum number of worthwhile ideas in the final slate. Five balances useful breadth with the attention required to compare candidates; requests may range from two to one hundred.
+
+Vidbyte charges 1 cent per requested suggestion from your prepaid balance before the run starts, so the charge in cents equals this count, even when review keeps fewer ideas. The generator considers a pool up to twice the requested count, capped at one hundred, before critique selects the strongest subset; that work runs on your own coding agent and its provider bills it to your own account.
+
+The count is never a quota that justifies weak filler. A smaller result includes an explicit shortfall so the caller can distinguish selectivity from failure.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/suggestion/prompts/critic_model.md

@@ -32,7 +32,7 @@ The input is bounded before it enters the agent context so oversized content bec
 A caller should keep related facts separate when their provenance, confidence, or expected effect differs enough to influence review.
 
 **Defaults and precedence for Critic Model**
-When it is omitted, the critic runs on the --model value, or on gpt-6-sol when --model is omitted too, so every call of the run bills at the rate --model names, and it uses the same high reasoning effort unless --reasoning-effort names another value.
+When it is omitted, the critic runs on the --model value, or, when --model is omitted too, on the model your Codex is configured with or on Claude Code's built-in default model, and it uses the same reasoning effort as every other role.
 Command-line values are the only source for this field, and the run does not read a separate configuration document.
 Environment variables do not silently populate this context value, which keeps an invocation reproducible.
 When multiple values are supplied, their command-line order becomes their stable context order.
@@ -93,8 +93,8 @@ Recover by correcting the specific value named by the error, rerunning dry-run w
 
 **Authentication and permissions for Critic Model**
 This input itself does not grant access to files, providers, accounts, repositories, or execution tools.
-Text values are available to local validation without credentials, while model-backed runs need a stored Vidbyte key with the models:invoke permission.
+Text values are available to local validation without credentials, while a real run needs a stored Vidbyte key with the runtime:write permission and your own signed-in Codex or Claude Code.
 The --files option reads only paths explicitly named by the caller and does not scan a workspace or agent history.
-Generator and critic agents have no file or shell tools and see only the text the CLI sends through Vidbyte to the model provider, so suggestions cannot mutate the host.
-The stored Vidbyte key is used only when a real model-backed run calls the model gateway, and it is never sent to a provider or copied into result prose.
-Permission failures should be repaired at the key, balance, or filesystem boundary rather than by weakening the context contract.
+Generator and critic agents run read-only on your own coding agent in an empty scratch folder and see only what this command sends them, so suggestions cannot mutate the host.
+The stored Vidbyte key only buys the run's admission at 1 cent per requested suggestion, and it is never sent to OpenAI or Anthropic or copied into result prose.
+Permission failures should be repaired at the key, balance, coding-agent login, or filesystem boundary rather than by weakening the context contract.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/suggestion/prompts/decision.md

@@ -93,8 +93,8 @@ Recover by correcting the specific value named by the error, rerunning dry-run w
 
 **Authentication and permissions for Decision Context**
 This input itself does not grant access to files, providers, accounts, repositories, or execution tools.
-Text values are available to local validation without credentials, while model-backed runs need a stored Vidbyte key with the models:invoke permission.
+Text values are available to local validation without credentials, while a real run needs a stored Vidbyte key with the runtime:write permission and your own signed-in Codex or Claude Code.
 The --files option reads only paths explicitly named by the caller and does not scan a workspace or agent history.
-Generator and critic agents have no file or shell tools and see only the text the CLI sends through Vidbyte to the model provider, so suggestions cannot mutate the host.
-The stored Vidbyte key is used only when a real model-backed run calls the model gateway, and it is never sent to a provider or copied into result prose.
-Permission failures should be repaired at the key, balance, or filesystem boundary rather than by weakening the context contract.
+Generator and critic agents run read-only on your own coding agent in an empty scratch folder and see only what this command sends them, so suggestions cannot mutate the host.
+The stored Vidbyte key only buys the run's admission at 1 cent per requested suggestion, and it is never sent to OpenAI or Anthropic or copied into result prose.
+Permission failures should be repaired at the key, balance, coding-agent login, or filesystem boundary rather than by weakening the context contract.

vidbyte_cli-0.3.4/src/vidbyte_cli/commands/agents/suggestion/prompts/dry_run.md

@@ -0,0 +1,5 @@
+A dry run validates and resolves the complete request without starting a coding agent. It exposes settings, manifest entries, omissions, and warnings before anything is charged or adds latency.
+
+File bodies stay out of the result by default while source identity and inclusion status remain inspectable. This makes large or sensitive context sets easier to audit before your own Codex or Claude Code sends them to its provider.
+
+Dry runs are free: no Vidbyte key, balance, admission, coding agent, or token usage is required for this validation path. The result contains no ideas and clearly identifies dry-run completion as its stopping reason.

vidbyte_cli-0.3.4/src/vidbyte_cli/commands/agents/suggestion/prompts/extra_compute.md

@@ -0,0 +1,5 @@
+Extra compute runs one independent suggestion agent for each selected category and combines their structured results before critique. Each category receives the same goal and caller context, but its own focused category window, so the run can explore more distinct directions.
+
+The default is off because one shared window is usually enough and uses fewer model turns, which your coding agent's provider bills to your own account. Enable it when breadth matters more than latency or when several categories deserve separate attention. At most four category agents run at once, and the combined pool is still bounded by the requested count and the workflow's total token and time limits.
+
+This setting changes the generation topology, not the meaning of a category, the final schema, or Vidbyte's 1 cent per requested suggestion. Every per-category result is labeled before combination, and the independent critic still reviews the combined candidates. A category whose agent fails is skipped with a warning, and only when every category fails does the run stop with its typed error.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/suggestion/prompts/files.md

@@ -93,8 +93,8 @@ Recover by correcting the specific value named by the error, rerunning dry-run w
 
 **Authentication and permissions for Supporting Files**
 This input itself does not grant access to files, providers, accounts, repositories, or execution tools.
-Text values are available to local validation without credentials, while model-backed runs need a stored Vidbyte key with the models:invoke permission.
+Text values are available to local validation without credentials, while a real run needs a stored Vidbyte key with the runtime:write permission and your own signed-in Codex or Claude Code.
 The --files option reads only paths explicitly named by the caller and does not scan a workspace or agent history.
-Generator and critic agents have no file or shell tools and see only the text the CLI sends through Vidbyte to the model provider, so suggestions cannot mutate the host.
-The stored Vidbyte key is used only when a real model-backed run calls the model gateway, and it is never sent to a provider or copied into result prose.
-Permission failures should be repaired at the key, balance, or filesystem boundary rather than by weakening the context contract.
+Generator and critic agents run read-only on your own coding agent in an empty scratch folder and see only what this command sends them, so suggestions cannot mutate the host.
+The stored Vidbyte key only buys the run's admission at 1 cent per requested suggestion, and it is never sent to OpenAI or Anthropic or copied into result prose.
+Permission failures should be repaired at the key, balance, coding-agent login, or filesystem boundary rather than by weakening the context contract.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/suggestion/prompts/forbidden.md

@@ -93,8 +93,8 @@ Recover by correcting the specific value named by the error, rerunning dry-run w
 
 **Authentication and permissions for Forbidden Actions**
 This input itself does not grant access to files, providers, accounts, repositories, or execution tools.
-Text values are available to local validation without credentials, while model-backed runs need a stored Vidbyte key with the models:invoke permission.
+Text values are available to local validation without credentials, while a real run needs a stored Vidbyte key with the runtime:write permission and your own signed-in Codex or Claude Code.
 The --files option reads only paths explicitly named by the caller and does not scan a workspace or agent history.
-Generator and critic agents have no file or shell tools and see only the text the CLI sends through Vidbyte to the model provider, so suggestions cannot mutate the host.
-The stored Vidbyte key is used only when a real model-backed run calls the model gateway, and it is never sent to a provider or copied into result prose.
-Permission failures should be repaired at the key, balance, or filesystem boundary rather than by weakening the context contract.
+Generator and critic agents run read-only on your own coding agent in an empty scratch folder and see only what this command sends them, so suggestions cannot mutate the host.
+The stored Vidbyte key only buys the run's admission at 1 cent per requested suggestion, and it is never sent to OpenAI or Anthropic or copied into result prose.
+Permission failures should be repaired at the key, balance, coding-agent login, or filesystem boundary rather than by weakening the context contract.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/suggestion/prompts/goal.md

@@ -93,8 +93,8 @@ Recover by correcting the specific value named by the error, rerunning dry-run w
 
 **Authentication and permissions for Goal**
 This input itself does not grant access to files, providers, accounts, repositories, or execution tools.
-Text values are available to local validation without credentials, while model-backed runs need a stored Vidbyte key with the models:invoke permission.
+Text values are available to local validation without credentials, while a real run needs a stored Vidbyte key with the runtime:write permission and your own signed-in Codex or Claude Code.
 The --files option reads only paths explicitly named by the caller and does not scan a workspace or agent history.
-Generator and critic agents have no file or shell tools and see only the text the CLI sends through Vidbyte to the model provider, so suggestions cannot mutate the host.
-The stored Vidbyte key is used only when a real model-backed run calls the model gateway, and it is never sent to a provider or copied into result prose.
-Permission failures should be repaired at the key, balance, or filesystem boundary rather than by weakening the context contract.
+Generator and critic agents run read-only on your own coding agent in an empty scratch folder and see only what this command sends them, so suggestions cannot mutate the host.
+The stored Vidbyte key only buys the run's admission at 1 cent per requested suggestion, and it is never sent to OpenAI or Anthropic or copied into result prose.
+Permission failures should be repaired at the key, balance, coding-agent login, or filesystem boundary rather than by weakening the context contract.

vidbyte_cli-0.3.4/src/vidbyte_cli/commands/agents/suggestion/prompts/host.md

@@ -0,0 +1,5 @@
+The host is the coding agent on this machine that runs every generator and critic turn: codex for Codex, which serves OpenAI models, or claude for Claude Code, which serves Anthropic models. It runs on your own login, so its provider bills the model usage to your own account, never to Vidbyte.
+
+When omitted, a claude-* --model picks claude, an OpenAI --model or --provider openai picks codex, and otherwise the first installed of codex and claude is used. A host that is not installed, or flags that point at two different hosts, fail before anything is charged. On Windows, a Claude Code found only as the npm batch launcher claude.cmd is refused before anything is charged too, because the SDK never starts a batch file; install the native Claude Code build instead.
+
+The host does not change the schema, the categories, or the price: Vidbyte charges 1 cent per requested suggestion either way. Both hosts run read-only in an empty scratch folder and see only what this command sends them. Claude Code runs in an isolated session that loads none of your Claude settings, so with no --model it uses Claude Code's built-in default model, while Codex uses the model it is configured with.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/suggestion/prompts/hypotheses.md

@@ -93,8 +93,8 @@ Recover by correcting the specific value named by the error, rerunning dry-run w
 
 **Authentication and permissions for Hypotheses**
 This input itself does not grant access to files, providers, accounts, repositories, or execution tools.
-Text values are available to local validation without credentials, while model-backed runs need a stored Vidbyte key with the models:invoke permission.
+Text values are available to local validation without credentials, while a real run needs a stored Vidbyte key with the runtime:write permission and your own signed-in Codex or Claude Code.
 The --files option reads only paths explicitly named by the caller and does not scan a workspace or agent history.
-Generator and critic agents have no file or shell tools and see only the text the CLI sends through Vidbyte to the model provider, so suggestions cannot mutate the host.
-The stored Vidbyte key is used only when a real model-backed run calls the model gateway, and it is never sent to a provider or copied into result prose.
-Permission failures should be repaired at the key, balance, or filesystem boundary rather than by weakening the context contract.
+Generator and critic agents run read-only on your own coding agent in an empty scratch folder and see only what this command sends them, so suggestions cannot mutate the host.
+The stored Vidbyte key only buys the run's admission at 1 cent per requested suggestion, and it is never sent to OpenAI or Anthropic or copied into result prose.
+Permission failures should be repaired at the key, balance, coding-agent login, or filesystem boundary rather than by weakening the context contract.

{vidbyte_cli-0.3.3 → vidbyte_cli-0.3.4}/src/vidbyte_cli/commands/agents/suggestion/prompts/idea.md

@@ -93,8 +93,8 @@ Recover by correcting the specific value named by the error, rerunning dry-run w
 
 **Authentication and permissions for Idea Identifier**
 This input itself does not grant access to files, providers, accounts, repositories, or execution tools.
-Text values are available to local validation without credentials, while model-backed runs need a stored Vidbyte key with the models:invoke permission.
+Text values are available to local validation without credentials, while a real run needs a stored Vidbyte key with the runtime:write permission and your own signed-in Codex or Claude Code.
 The --files option reads only paths explicitly named by the caller and does not scan a workspace or agent history.
-Generator and critic agents have no file or shell tools and see only the text the CLI sends through Vidbyte to the model provider, so suggestions cannot mutate the host.
-The stored Vidbyte key is used only when a real model-backed run calls the model gateway, and it is never sent to a provider or copied into result prose.
-Permission failures should be repaired at the key, balance, or filesystem boundary rather than by weakening the context contract.
+Generator and critic agents run read-only on your own coding agent in an empty scratch folder and see only what this command sends them, so suggestions cannot mutate the host.
+The stored Vidbyte key only buys the run's admission at 1 cent per requested suggestion, and it is never sent to OpenAI or Anthropic or copied into result prose.
+Permission failures should be repaired at the key, balance, coding-agent login, or filesystem boundary rather than by weakening the context contract.