@difflab/pi 0.7.1 → 0.8.0

package/skills/review/references/workflows/help.md

@@ -1,15 +1,24 @@
 # help
 
-```text
-/review auto [target] [--local] [--bg]
-/review new [title] [--intent text] [--base branch] [--local] [--bg]
-/review open [target] [--local]
-/review status [target] [--local]
-/review edit [target] [--local] [--bg]
-/review address [target] [--local] [--bg]
-/review publish [target] [--local] [--comment|--approve|--request-changes|--close] [--bg]
-/review complete [target] [--local|--approve|--reject|--abandon] [--bg]
-/review merge [target] [--bg]
-```
-
-Every target-bearing workflow calls `review_context` first; help is informational and may run without a target or repository. `--local` selects working-tree tuicr state and is preserved; otherwise use the forge. `--bg` launches one named same-session Orchestrator child without redispatch or mode changes. `auto` reviews and stages grounded findings; `new` creates a draft; `open` opens an existing target; `status` is read-only; `edit` opens without findings; `address` fixes and replies without resolving; `publish` publishes/promotes pending work; `complete` approves/rejects/abandons or archives local state without merging; `merge` is GitHub-only and requires clean settled readiness. Aliases: create/draft=new, launch=auto, ready=publish, close=complete, land=merge. Unknown verbs show this help.
+## Parse arguments
+
+Accept `help` or a request with no recognizable review verb; no positional arguments or flags. Prefer an explicit verb, then safely inferred natural-language intent, then help. Ask only if a material ambiguity prevents routing to a workflow.
+
+## Steps
+
+1. Show the complete subcommand list:
+
+   ```text
+   /review new [title] [--intent text] [--base branch] [--local]
+   /review auto [target] [--local]
+   /review open [target] [--local]
+   /review status [target] [--local]
+   /review edit [target] [--local]
+   /review address [target] [--local]
+   /review publish [target] [--local] [--comment|--approve|--request-changes|--close]
+   /review complete [target] [--local|--approve|--reject|--abandon]
+   /review merge [target]
+   /review help
+   ```
+
+2. Show aliases: `create`/`draft` → `new`, `launch` → `auto`, `ready` → `publish`, `close` → `complete`, `land` → `merge`.

package/skills/review/references/workflows/merge.md

@@ -1,5 +1,20 @@
 # merge
 
-**Owner/tier:** lifecycle coordinator; remote GitHub-only, no Reviewer or Worker.
+## Parse arguments
 
-Reject `--local`. Call `review_context` first with `local:false`, then `review_merge`; the tool checks open, non-draft, clean PR and settled checks and uses branch protection as authoritative. Approval by the current user is not required. Ask only when a conventional squash subject is materially unclear. Report merge or readiness failure. Merge never publishes or completes review state.
+Accept `merge [target]` and an optional conventional squash subject in the request. Fuzzy-match GitHub PR target from the request and repository. Prefer explicit target/subject, then safe inference, then current PR. Reject `--local`; do not infer it. Call `ask_user_question` only for material target or subject ambiguity before launch; return a precise blocker if a new decision arises later.
+
+## Steps
+
+1. Call `review_context` first with `local:false` and the selected target. Reject non-GitHub or local reviews without switching backend. For a short, explicitly approved merge requiring no substantive analysis, call `review_merge` directly with its resolved target and conventional squash subject if needed. The tool rechecks open, ready, clean PR status and settled CI; branch protection is authoritative, and current-user approval is not required unless protection requires it. Report the actual merge or exact readiness/CI/subject failure. Do not publish or complete in its place.
+2. If merge-readiness requires substantive inspection or judgment beyond the tool's checks, when native subagents are available, delegate only that analysis to medium `diffpi-orchestrator` in the background with ambient capabilities. Fill every field in this self-contained prompt with actual values (or `none`), including the exact request; never send raw placeholders:
+
+   ```text
+   Inspect GitHub PR merge readiness without merging.
+   - Repository: {absolute repo cwd}; exact request: {exact request}.
+   - Resolved target: {target}; backend: remote GitHub; local: false; issue URL: {issue URL or none}; proposed conventional squash subject: {subject or none}.
+   - Call review_context first with local:false; reuse its target/backend/cwd. Inspect PR readiness, subject and settled CI using review_status and available checks. Treat branch protection as authoritative; do not demand current-user approval unless protection requires it.
+   - Return completed evidence and a go/no-go recommendation or precise blocker. Do not merge, publish, complete or resolve threads.
+   ```
+
+3. If no native subagents are available for step 2, follow [foreground fallback](foreground-fallback.md) before inspection; only after explicit confirmation inspect inline and report actual evidence without claiming an Orchestrator result. Otherwise require the completed result, not a queued job or claim. Resolve material decisions with the user, then call `review_merge` in the initiating thread only if approval still applies. Report the actual result.

package/skills/review/references/workflows/new.md

@@ -1,9 +1,21 @@
 # new
 
-**Owner/tier:** foreground or background lifecycle coordinator; no Reviewer or Worker required.
+## Parse arguments
 
-1. Parse title, intent, base, target, and `--local`. Call `review_context` first with the exact selection.
-2. Call `review_new` with the same target/backend/local values. This creates a local tuicr review or remote draft PR/MR and opens it; use `review_launch_ui` only when needed.
-3. Report the created target and draft state. Do not run findings, publish, complete, or merge.
+Accept `new [title] [--intent text] [--base branch] [--local]` and a target named in the request. Fuzzy-match draft intent. Prefer explicit arguments/flags, then safe natural-language or repository title/intent/target/base/backend, then available defaults. Select local for an explicit working-tree request. Call `ask_user_question` only for material ambiguity before launch; return a precise blocker if a new decision arises later.
 
-Missing title/base, unsupported forge, dirty remote branch, or creation failure is an explicit error; never silently switch backend. `--bg` is one named child only.
+## Steps
+
+1. Resolve the concrete target/backend/local and base with `review_context`. Substitute actual values for every placeholder, including the exact request and issue URL if applicable; never send raw `{placeholders}`. When native subagents are available, launch a lightweight/low `diffpi-worker` background subagent for this bounded draft-creation task. Pass the following self-contained prompt with ambient capabilities:
+
+   ```text
+   Create a review draft.
+   - Repository: {absolute repo cwd}; exact request: {exact request}.
+   - Resolved target: {target}; backend: {backend}; local: {local}; title: {title}; intent: {intent}; base: {base}; issue URL: {issue URL or none}; approved policy: {policy}.
+   - This explicitly assigned review-draft creation is the only review lifecycle mutation authorized here. Call review_context first. Reuse its target/backend/local/cwd for all subsequent calls; never switch backend silently.
+   - Reject unsupported forge, dirty remote branch or unavailable local base/default with exact evidence.
+   - Call review_new to create the local tuicr review or remote draft PR/MR. Call review_launch_ui only if needed; report a returned launch command if automatic launch fails.
+   - Verify and report the actual created target and draft state or exact failure. Do not stage findings, publish, complete or merge.
+   ```
+
+2. If no native subagents are available, follow [foreground fallback](foreground-fallback.md) before creation. Only after explicit confirmation, perform the draft-creation steps above inline and verify with `review_status`; do not claim a Worker result. Otherwise require the completed child result and verify the created target/draft with `review_status`. Stop on failure; report the actual result, not merely a queued ID or claim. Do not assign the Worker publication, merge, findings judgment or substantive inspection.

package/skills/review/references/workflows/open.md

@@ -1,5 +1,11 @@
 # open
 
-**Owner/tier:** lifecycle coordinator, read-only.
+## Parse arguments
 
-Call `review_context` first with the exact target/backend/local, then call `review_open`. It opens an existing remote PR/MR; it never creates one. With `--local`, preserve the local behavior selected by context. Report unsupported or missing targets and browser/launch failures; do not mutate review state.
+Accept `open [target] [--local]`. Fuzzy-match the existing PR/MR or tuicr session from the request and repository. Prefer explicit arguments/flags, then safe inference, then the current-review target. Select local for an explicit working-tree request. Call `ask_user_question` only for material ambiguity.
+
+## Steps
+
+1. Call `review_context` first with the selected target/backend/local; reuse its target/backend/local/cwd. Reject unsupported selections without switching backend.
+2. Call `review_open` for an existing remote PR/MR in the browser. For `--local`, call `review_edit` to open only the matching existing tuicr session.
+3. Report the opened target or exact missing-session/browser/launch failure. Do not create a review, generate findings or mutate review state.

package/skills/review/references/workflows/publish.md

@@ -1,7 +1,20 @@
 # publish
 
-**Owner/tier:** lifecycle coordinator; no Reviewer or Worker.
+## Parse arguments
 
-1. Parse target, `--local`, and one status: comment (default), approve, request-changes, or close. Call `review_context` first and preserve the selected backend. A local promotion may be used only when context identifies the matching local session; never silently switch.
-2. Call `review_publish` with the same target/backend/local. It promotes/stages pending local work where applicable and publishes the selected status; remote comments remain pending until this call. Comment/approve/request-changes mark remote drafts ready; close publishes comment then closes. GitLab may reject request-changes.
-3. Report promoted bodies/comments/replies, status, and final state. Publish never merges. Unsupported target, unmatched response, or forge failure is explicit.
+Accept `publish [target] [--local] [--comment|--approve|--request-changes|--close]`. Fuzzy-match target/backend/status from the request and repository. Prefer explicit arguments/flags, then safe inference, then current-review target and comment status. Call `ask_user_question` only for material ambiguity before launch; return a precise blocker if a new decision arises later.
+
+## Steps
+
+1. Call `review_context` first with the selected target/backend/local. For a short, explicitly approved publication with no substantive inspection, call `review_publish` directly using its resolved values and selected status (`COMMENT` by default, `APPROVE`, `REQUEST_CHANGES`, or `CLOSE`). Remote findings and replies stay pending until this call. Promote only drafts matching the remote PR/MR; preserve `local:true` for a working-tree session. Report unmatched promotion, GitLab request-changes rejection, and exact failures; verify the final status. Do not merge or resolve threads.
+2. If the request instead needs substantive inspection or judgment before publication, when native subagents are available, delegate only that analysis to medium `diffpi-orchestrator` in the background with ambient capabilities. Fill every field in this self-contained prompt with actual values (or `none`), including the exact request; never send raw placeholders:
+
+   ```text
+   Inspect publication readiness without publishing.
+   - Repository: {absolute repo cwd}; exact request: {exact request}.
+   - Resolved target: {target}; backend: {backend}; local: {local}; selected status: {status}; issue URL: {issue URL or none}; approved policy: {policy}.
+   - Call review_context first; reuse its target/backend/local/cwd. Inspect the matching session and pending findings/replies; identify unmatched drafts, failed checks and material decisions. Do not silently switch backends.
+   - Return actual completed evidence and a go/no-go recommendation or precise blocker. Do not publish, complete, merge or resolve threads.
+   ```
+
+3. If no native subagents are available for step 2, follow [foreground fallback](foreground-fallback.md) before inspection; only after explicit confirmation inspect inline and report actual evidence without claiming an Orchestrator result. Otherwise require the completed inspection result, not a queued job or claim. Resolve material decisions with the user, then call `review_publish` in the initiating thread only if approval still applies. Report the actual published or blocked outcome.

package/skills/review/references/workflows/status.md

@@ -1,5 +1,10 @@
 # status
 
-**Owner/tier:** lifecycle coordinator, read-only.
+## Parse arguments
 
-Call `review_context` first with the exact target/backend/local, then `review_status` with the same values. Report branch, worktree, forge/local review, URLs, and tuicr state. Missing or unsupported targets are errors; status never creates, publishes, completes, merges, or changes comments.
+Accept `status [target] [--local]`. Fuzzy-match the status target/backend from the request and repository. Prefer explicit arguments/flags, then safe inference, then the current-review target. Select local for an explicit working-tree request. Call `ask_user_question` only for material ambiguity.
+
+## Steps
+
+1. Call `review_context` first with the selected target/backend/local; reuse its target/backend/local/cwd. Reject unsupported selections without switching backend.
+2. Call `review_status` with those same values. Report branch, worktree, local or remote review, URLs, tuicr session, pending/draft state and exact errors. Do not create a review or mutate source, comments or lifecycle state.

package/templates/plan/PLAN.md

@@ -44,4 +44,4 @@
 
 ## References
 
-<!-- Add repository or external references. -->
+<!-- Add repository or external references. Record one completed Plan Reviewer round per authoring cycle: reviewed file hashes, Git state, verdict and findings; each finding's disposition; post-fix structural verification and current snapshot. A repaired BLOCKING verdict is not a reviewer PASS. Keep DRAFT until explicit finalize/go. -->

package/dist/commands/mode.d.ts

@@ -1,5 +0,0 @@
-import type { ExtensionAPI } from '@earendil-works/pi-coding-agent';
-import type { ModeController } from '../modes';
-/** Register the inline agent mode command. */
-export declare function registerModeCommand(pi: ExtensionAPI, modes: ModeController): void;
-//# sourceMappingURL=mode.d.ts.map

package/dist/commands/mode.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"mode.d.ts","sourceRoot":"","sources":["../../src/commands/mode.ts"],"names":[],"mappings":"AAAA,OAAO,KAAK,EAAE,YAAY,EAAoB,MAAM,iCAAiC,CAAC;AACtF,OAAO,KAAK,EAAE,cAAc,EAAE,MAAM,UAAU,CAAC;AAK/C,8CAA8C;AAC9C,wBAAgB,mBAAmB,CAAC,EAAE,EAAE,YAAY,EAAE,KAAK,EAAE,cAAc,GAAG,IAAI,CAKjF"}

package/dist/modes.d.ts

@@ -1,75 +0,0 @@
-import { type ExtensionAPI, type ExtensionContext } from '@earendil-works/pi-coding-agent';
-export type ModePromptStrategy = 'append' | 'replace';
-export type ModeThinkingLevel = 'off' | 'minimal' | 'low' | 'medium' | 'high' | 'xhigh' | 'max';
-export interface AgentMode {
-    id: string;
-    label: string;
-    description: string;
-    systemPrompt: string;
-    promptStrategy: ModePromptStrategy;
-    modelPreferences: string[];
-    thinkingLevel?: ModeThinkingLevel;
-    requiredTools: string[];
-    forbiddenTools: string[];
-    requireModel: boolean;
-    requireThinking: boolean;
-    tools: string[];
-    source: string;
-    sourcePath: string;
-}
-export interface ModeCatalog {
-    modes: AgentMode[];
-    diagnostics: string[];
-}
-export interface ModeListOptions {
-    includeSkills?: boolean;
-}
-export interface ModeController {
-    list(ctx: ExtensionContext, options?: ModeListOptions): Promise<ModeCatalog>;
-    set(agent: string, ctx: ExtensionContext): Promise<ModeSelectionResult>;
-    unset(ctx: ExtensionContext): Promise<ModeSelectionResult>;
-    restore(ctx: ExtensionContext): Promise<void>;
-    refresh(ctx: ExtensionContext): void;
-    apply(systemPrompt: string): string;
-    getActive(): AgentMode | undefined;
-    getStatus(ctx: ExtensionContext): ModeRuntimeStatus;
-}
-export interface ModeRuntimeStatus {
-    active?: string;
-    model?: {
-        provider: string;
-        id: string;
-    };
-    thinkingLevel: ModeThinkingLevel;
-    activeTools: string[];
-    capabilityError?: string;
-}
-export type ModeSelectionResult = {
-    ok: true;
-    active?: AgentMode;
-    message: string;
-} | {
-    ok: false;
-    message: string;
-};
-export interface ModeDiscoveryOptions extends ModeListOptions {
-    cwd: string;
-    agentDir?: string;
-    bundledAgentsDir?: string;
-    homeDir?: string;
-    projectTrusted: boolean;
-}
-export interface ModeControllerOptions {
-    agentDir?: string;
-    bundledAgentsDir?: string;
-    homeDir?: string;
-}
-type ModeRuntime = Pick<ExtensionAPI, 'appendEntry' | 'getActiveTools' | 'getAllTools' | 'getThinkingLevel' | 'setActiveTools' | 'setModel' | 'setThinkingLevel'>;
-/** Discover standard agents and, when requested, qualified skill agents. */
-export declare function discoverAgentModes(options: ModeDiscoveryOptions): Promise<ModeCatalog>;
-/** Resolve an agent id without guessing an unqualified skill-agent name. */
-export declare function resolveAgentMode(modes: readonly AgentMode[], requested: string): ModeSelectionResult;
-/** Create the session-scoped controller that applies and restores complete mode profiles. */
-export declare function createModeController(pi: ModeRuntime, options?: ModeControllerOptions): ModeController;
-export {};
-//# sourceMappingURL=modes.d.ts.map

package/dist/modes.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"modes.d.ts","sourceRoot":"","sources":["../src/modes.ts"],"names":[],"mappings":"AAAA,OAAO,EAGL,KAAK,YAAY,EACjB,KAAK,gBAAgB,EACtB,MAAM,iCAAiC,CAAC;AAUzC,MAAM,MAAM,kBAAkB,GAAG,QAAQ,GAAG,SAAS,CAAC;AACtD,MAAM,MAAM,iBAAiB,GAAG,KAAK,GAAG,SAAS,GAAG,KAAK,GAAG,QAAQ,GAAG,MAAM,GAAG,OAAO,GAAG,KAAK,CAAC;AAEhG,MAAM,WAAW,SAAS;IACxB,EAAE,EAAE,MAAM,CAAC;IACX,KAAK,EAAE,MAAM,CAAC;IACd,WAAW,EAAE,MAAM,CAAC;IACpB,YAAY,EAAE,MAAM,CAAC;IACrB,cAAc,EAAE,kBAAkB,CAAC;IACnC,gBAAgB,EAAE,MAAM,EAAE,CAAC;IAC3B,aAAa,CAAC,EAAE,iBAAiB,CAAC;IAClC,aAAa,EAAE,MAAM,EAAE,CAAC;IACxB,cAAc,EAAE,MAAM,EAAE,CAAC;IACzB,YAAY,EAAE,OAAO,CAAC;IACtB,eAAe,EAAE,OAAO,CAAC;IACzB,KAAK,EAAE,MAAM,EAAE,CAAC;IAChB,MAAM,EAAE,MAAM,CAAC;IACf,UAAU,EAAE,MAAM,CAAC;CACpB;AAED,MAAM,WAAW,WAAW;IAC1B,KAAK,EAAE,SAAS,EAAE,CAAC;IACnB,WAAW,EAAE,MAAM,EAAE,CAAC;CACvB;AAED,MAAM,WAAW,eAAe;IAC9B,aAAa,CAAC,EAAE,OAAO,CAAC;CACzB;AAED,MAAM,WAAW,cAAc;IAC7B,IAAI,CAAC,GAAG,EAAE,gBAAgB,EAAE,OAAO,CAAC,EAAE,eAAe,GAAG,OAAO,CAAC,WAAW,CAAC,CAAC;IAC7E,GAAG,CAAC,KAAK,EAAE,MAAM,EAAE,GAAG,EAAE,gBAAgB,GAAG,OAAO,CAAC,mBAAmB,CAAC,CAAC;IACxE,KAAK,CAAC,GAAG,EAAE,gBAAgB,GAAG,OAAO,CAAC,mBAAmB,CAAC,CAAC;IAC3D,OAAO,CAAC,GAAG,EAAE,gBAAgB,GAAG,OAAO,CAAC,IAAI,CAAC,CAAC;IAC9C,OAAO,CAAC,GAAG,EAAE,gBAAgB,GAAG,IAAI,CAAC;IACrC,KAAK,CAAC,YAAY,EAAE,MAAM,GAAG,MAAM,CAAC;IACpC,SAAS,IAAI,SAAS,GAAG,SAAS,CAAC;IACnC,SAAS,CAAC,GAAG,EAAE,gBAAgB,GAAG,iBAAiB,CAAC;CACrD;AAED,MAAM,WAAW,iBAAiB;IAChC,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,KAAK,CAAC,EAAE;QAAE,QAAQ,EAAE,MAAM,CAAC;QAAC,EAAE,EAAE,MAAM,CAAA;KAAE,CAAC;IACzC,aAAa,EAAE,iBAAiB,CAAC;IACjC,WAAW,EAAE,MAAM,EAAE,CAAC;IACtB,eAAe,CAAC,EAAE,MAAM,CAAC;CAC1B;AAED,MAAM,MAAM,mBAAmB,GAAG;IAAE,EAAE,EAAE,IAAI,CAAC;IAAC,MAAM,CAAC,EAAE,SAAS,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,GAAG;IAAE,EAAE,EAAE,KAAK,CAAC;IAAC,OAAO,EAAE,MAAM,CAAA;CAAE,CAAC;AAErH,MAAM,WAAW,oBAAqB,SAAQ,eAAe;IAC3D,GAAG,EAAE,MAAM,CAAC;IACZ,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,cAAc,EAAE,OAAO,CAAC;CACzB;AAED,MAAM,WAAW,qBAAqB;IACpC,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B,OAAO,CAAC,EAAE,MAAM,CAAC;CAClB;AAED,KAAK,WAAW,GAAG,IAAI,CACrB,YAAY,EACV,aAAa,GACb,gBAAgB,GAChB,aAAa,GACb,kBAAkB,GAClB,gBAAgB,GAChB,UAAU,GACV,kBAAkB,CACrB,CAAC;AAiCF,4EAA4E;AAC5E,wBAAsB,kBAAkB,CAAC,OAAO,EAAE,oBAAoB,GAAG,OAAO,CAAC,WAAW,CAAC,CAiC5F;AAED,4EAA4E;AAC5E,wBAAgB,gBAAgB,CAAC,KAAK,EAAE,SAAS,SAAS,EAAE,EAAE,SAAS,EAAE,MAAM,GAAG,mBAAmB,CAqBpG;AAED,6FAA6F;AAC7F,wBAAgB,oBAAoB,CAAC,EAAE,EAAE,WAAW,EAAE,OAAO,GAAE,qBAA0B,GAAG,cAAc,CAgIzG"}

package/dist/tools/modes.d.ts

@@ -1,4 +0,0 @@
-import { type ToolDefinition } from '@earendil-works/pi-coding-agent';
-import type { ModeController } from '../modes';
-export declare function createModeTools(controller: ModeController): readonly ToolDefinition[];
-//# sourceMappingURL=modes.d.ts.map

package/dist/tools/modes.d.ts.map

@@ -1 +0,0 @@
-{"version":3,"file":"modes.d.ts","sourceRoot":"","sources":["../../src/tools/modes.ts"],"names":[],"mappings":"AAAA,OAAO,EAAc,KAAK,cAAc,EAAE,MAAM,iCAAiC,CAAC;AAGlF,OAAO,KAAK,EAAe,cAAc,EAAE,MAAM,UAAU,CAAC;AAiB5D,wBAAgB,eAAe,CAAC,UAAU,EAAE,cAAc,GAAG,SAAS,cAAc,EAAE,CAiFrF"}

package/skills/mode/SKILL.md

@@ -1,38 +0,0 @@
----
-name: mode
-description: List, select, or clear an inline agent profile. Use when the user invokes /skill:mode or asks to change the current Diffpi mode.
-allowed-tools: ask_user_question diffpi_modes_list diffpi_modes_set diffpi_modes_unset diffpi_modes_status
----
-
-# Mode
-
-Pi appends `/skill:mode` arguments as `User: <arguments>`. Route those arguments without manual command parsing in the extension.
-
-## Direct requests
-
-- If the argument is `help`, `-h`, or `--help`, print the usage and examples below. Do not call a tool.
-- If the argument is `clear`, call `diffpi_modes_unset` and stop.
-- If the argument is one agent id, call `diffpi_modes_set` with that exact id. On the next model turn, call `diffpi_modes_status` to verify the selected model, thinking level, and active tools; if preflight or verification fails, report the exact capability error and do not claim an active mode.
-- A qualified id such as `spec:planner` is one agent id. The tool enables skill-agent discovery for it.
-- If the request contains multiple agent ids or an unsupported flag, print the usage and stop.
-
-```text
-/skill:mode [--include-skills] [agent|clear]
-/skill:mode
-/skill:mode --include-skills
-/skill:mode worker
-/skill:mode spec:planner
-/skill:mode clear
-```
-
-## Picker
-
-Use the picker when there are no arguments or when the only argument is `--include-skills`.
-
-1. Call `diffpi_modes_list`. Set `includeSkills` to true only for `--include-skills`.
-2. Call `ask_user_question` with one single-select question. Do not ask as plain chat text.
-3. Put the complete list of valid ids in the question text. Offer Default plus at most three relevant agents as structured options. The custom-answer row lets the user enter any other valid id.
-4. If the user selects Default or enters `clear`, call `diffpi_modes_unset`.
-5. Otherwise, call `diffpi_modes_set` with the selected or entered id. On the next model turn, call `diffpi_modes_status` and verify the selected model, thinking level, and active tools. If selection or post-turn verification fails, report the exact capability error and ensure no ghost active mode is claimed.
-
-Do not perform unrelated work. Explain that the selected prompt, preferred available model, thinking level, and tool set start on the next turn. Mode selection is not a security boundary.