claude-code-modes 0.4.0 → 0.5.0

package/README.md

@@ -102,9 +102,27 @@ claude-mode none        # Strip all behavioral opinions, use your own CLAUDE.md
 | `spark` | autonomous | architect | unrestricted | Maximum expression — muse's creative vision plus wit and personality |
 | `none` | — | — | — | Strip all behavioral instructions, use your own |
 
-### Alternative base: chill
+### Bases
 
-The default "standard" base is derived from upstream Claude Code. The **chill** base is an alternative informed by Anthropic's [emotion research](https://www.anthropic.com/research/emotion-concepts-function) — shorter (~65% the size), calmer framing, no ALL-CAPS emphasis, with worked examples and a priority hierarchy:
+By default the base is chosen for you. `claude-mode` reads which model the session will
+run on and assembles the same *shape* of prompt Claude Code itself would send that model:
+the **lean** base for models that receive upstream's lean assembly (Opus 5, Opus 4.8,
+Fable 5, Mythos 5), and **standard** for everything else. On Opus 5 it also layers on the
+three extra sections upstream sends that model — delivering-work, corrections, and
+tool-restraint.
+
+```bash
+claude-mode create                     # auto — picks lean or standard from your model
+claude-mode create --base lean         # force the lean base
+claude-mode create --base standard     # force the full upstream-derived base
+```
+
+Naming a base explicitly opts out of detection entirely, including the Opus 5 extras.
+The lean base is roughly a sixth the size of standard's head: one `# Harness` block in
+place of the separate System, Doing tasks, Executing actions, Using your tools, and
+Tone and style sections.
+
+The "standard" base is derived from upstream Claude Code. The **chill** base is an alternative informed by Anthropic's [emotion research](https://www.anthropic.com/research/emotion-concepts-function) — shorter (~65% the size), calmer framing, no ALL-CAPS emphasis, with worked examples and a priority hierarchy:
 
 ```bash
 claude-mode create --base chill        # Use chill base with any preset
@@ -146,12 +164,12 @@ prompts/
   base/         Standard base (derived from upstream Claude Code)
   chill/        Alternative base (emotion-research-informed, leaner)
   flow/         Alternative base (chill's calm + restored engagement)
+  lean/         Upstream's lean assembly (what Opus 5 / 4.8 / Fable 5 actually receive)
   axis/         Behavioral prompts organized by three axes
-  style/        Writing styles (declaudified)
-  modifiers/    Behavioral layers (bold, debug, methodical, director, readonly, context-pacing, speak-plain, tdd, muse, flow, playful)
+  modifiers/    Behavioral layers (bold, debug, methodical, director, readonly, context-pacing, speak-plain, tdd, muse, flow, playful, delivering-work, corrections, tool-restraint)
 ```
 
-Each base has a `base.json` manifest — a flat JSON array declaring fragment order with `"axes"` and `"modifiers"` as reserved insertion points. The standard base is validated against Claude Code **v2.1.198**.
+Each base has a `base.json` manifest — a flat JSON array declaring fragment order with `"axes"` and `"modifiers"` as reserved insertion points. The standard and lean bases are validated against Claude Code **v2.1.220**.
 
 The behavioral layer is composed from three independent axes — **agency** (how much initiative), **quality** (what code standard), and **scope** (how far beyond the request). Presets are just named combinations of these three values.
 
@@ -186,15 +204,6 @@ claude-mode create --quality ./team-quality.md     # Use a custom quality fragme
 claude-mode create --quality team-standard          # Resolve from config
 ```
 
-Set a writing style — a single fragment that shapes how Claude writes to you (answers, summaries, commit messages), independent of the behavioral axes:
-
-```bash
-claude-mode create --style declaudified            # Built-in: lead with the answer, cut filler and stock phrases
-claude-mode create --style ./house-style.md        # Or a custom fragment / config-defined name
-```
-
-No style is applied unless you pass `--style`, set `defaultStyle` in config, or use a preset that declares one.
-
 Add modifiers:
 
 ```bash
@@ -208,9 +217,12 @@ claude-mode create --append-system-prompt "Use Rust, not TypeScript"
 Pass flags through to Claude Code:
 
 ```bash
-claude-mode create -- --verbose --model sonnet
+claude-mode create --model sonnet          # model choice also lands in the prompt's environment info
+claude-mode create -- --verbose            # anything after -- goes to claude verbatim
 ```
 
+The environment section of the assembled prompt reports the model claude will actually run: from `--model` (before or after `--`), else the `ANTHROPIC_MODEL` env var, else Claude settings files (`.claude/settings.local.json`, `.claude/settings.json`, `~/.claude/settings.json`), else the newest known model.
+
 Debug the assembled prompt:
 
 ```bash
@@ -259,9 +271,6 @@ Example `.claude-mode.json`:
   "modifiers": {
     "team-rules": "./prompts/team-rules.md"
   },
-  "styles": {
-    "house": "./prompts/house-style.md"
-  },
   "axes": {
     "quality": {
       "team-standard": "./prompts/team-quality.md"
@@ -272,7 +281,6 @@ Example `.claude-mode.json`:
       "agency": "collaborative",
       "quality": "team-standard",
       "scope": "adjacent",
-      "style": "house",
       "modifiers": ["team-rules"]
     }
   }
@@ -281,15 +289,13 @@ Example `.claude-mode.json`:
 
 - **`defaultModifiers`** — always applied to every invocation (no flag needed)
 - **`modifiers`** — named modifiers referencing markdown files
-- **`styles`** — named writing styles referencing markdown files
 - **`axes`** — custom axis values (replace built-in fragments)
 - **`presets`** — named presets composing built-in and custom values
 
-Config also supports bases and a default style:
+Config also supports bases:
 
-- **`defaultBase`** — base to use when `--base` isn't specified
+- **`defaultBase`** — base to use when `--base` isn't specified; set it to `"auto"` to restore model-driven selection
 - **`bases`** — named bases referencing directories with `base.json` manifests
-- **`defaultStyle`** — style to apply when `--style` isn't specified
 
 Config searches `.claude-mode.json` in the current directory first, then `~/.config/claude-mode/config.json` as a global fallback. All commands accept `--global` to target the global config.
 
@@ -302,8 +308,6 @@ claude-mode config add-default <name-or-path>        # Add to defaultModifiers
 claude-mode config remove-default <name>             # Remove from defaultModifiers
 claude-mode config add-modifier <name> <path>        # Register named modifier
 claude-mode config remove-modifier <name>            # Unregister named modifier
-claude-mode config add-style <name> <path>           # Register named style
-claude-mode config remove-style <name>               # Unregister named style
 claude-mode config add-axis <axis> <name> <path>     # Register custom axis value
 claude-mode config remove-axis <axis> <name>         # Unregister custom axis value
 claude-mode config add-preset <name> [flags]         # Create custom preset

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "claude-code-modes",
-  "version": "0.4.0",
+  "version": "0.5.0",
   "description": "Behaviorally-tuned system prompts for Claude Code",
   "license": "MIT",
   "type": "module",

package/prompts/base/actions.md

@@ -8,4 +8,4 @@ Examples of the kind of risky actions that warrant user confirmation:
 - Actions visible to others or that affect shared state: pushing code, creating/closing/commenting on PRs or issues, sending messages (Slack, email, GitHub), posting to external services, modifying shared infrastructure or permissions
 - Uploading content to third-party web tools (diagram renderers, pastebins, gists) publishes it - consider whether it could be sensitive before sending, since it may be cached or indexed even if later deleted.
 
-When you encounter an obstacle, do not use destructive actions as a shortcut to simply make it go away. For instance, try to identify root causes and fix underlying issues rather than bypassing safety checks (e.g. --no-verify). If you discover unexpected state like unfamiliar files, branches, or configuration, investigate before deleting or overwriting, as it may represent the user's in-progress work. If you're unsure whether the user would want something kept, prefer a reversible step (move it aside, rename it, or stash it) over deleting; files you created yourself this session (scratch outputs, experiment intermediates) are yours to clean up freely. For example, typically resolve merge conflicts rather than discarding changes; similarly, if a lock file exists, investigate what process holds it rather than deleting it. In a git repository, run `git status` before any command that could discard uncommitted work (git checkout/restore/reset/clean, rm -rf on a repo path, restoring from a snapshot), and stash (with `-u` for untracked) or commit anything you find first. In short: only take risky actions carefully, and when in doubt, ask before acting. Follow both the spirit and letter of these instructions - measure twice, cut once.
+When you encounter an obstacle, do not use destructive actions as a shortcut to simply make it go away. For instance, try to identify root causes and fix underlying issues rather than bypassing safety checks (e.g. --no-verify). If you discover unexpected state like unfamiliar files, branches, or configuration, investigate before deleting or overwriting, as it may represent the user's in-progress work. If you're unsure whether the user would want something kept, prefer a reversible step (move it aside, rename it, or stash it) over deleting; files you created yourself this session (scratch outputs, experiment intermediates) are yours to clean up freely. For example, typically resolve merge conflicts rather than discarding changes; similarly, if a lock file exists, investigate what process holds it rather than deleting it. In a git repository, run `git status` before any command that could discard uncommitted work (git checkout/restore/reset/clean, rm -rf on a repo path, restoring from a snapshot), and stash (with `-u` for untracked) or commit anything you find first. And when staging or committing: review what's included (`git status` after a broad `git add`), and if you see anything suspicious that might reveal secrets — even if the filename looks innocuous — double-check the file's contents before pushing. In short: only take risky actions carefully, and when in doubt, ask before acting. Follow both the spirit and letter of these instructions - measure twice, cut once.

package/prompts/base/base.json

@@ -7,7 +7,9 @@
   "tools.md",
   "tone.md",
   "text-output.md",
+  "pronouns.md",
   "session-guidance.md",
+  "context-management.md",
   "modifiers",
   "env.md"
 ]

package/prompts/base/context-management.md

@@ -0,0 +1,4 @@
+# Context management
+When the conversation grows long, some or all of the current context is summarized; the summary, along with any remaining unsummarized context, is provided in the next context window so work can continue — you don't need to wrap up early or hand off mid-task.
+
+When you have enough information to act, act. Do not re-derive facts already established in the conversation, or re-litigate a decision the user has already made.

package/prompts/base/env.md

@@ -7,9 +7,9 @@ You have been invoked in the following environment:
  - OS Version: {{OS_VERSION}}
  - You are powered by the model named {{MODEL_NAME}}. The exact model ID is {{MODEL_ID}}.
  - Assistant knowledge cutoff is {{KNOWLEDGE_CUTOFF}}.
- - The most recent Claude models are the Claude 5 family, Opus 4.8, and Haiku 4.5. Model IDs — Fable 5: 'claude-fable-5', Opus 4.8: 'claude-opus-4-8', Sonnet 5: 'claude-sonnet-5', Haiku 4.5: 'claude-haiku-4-5-20251001'. When building AI applications, default to the latest and most capable Claude models.
+ - The most recent Claude models are the Claude 5 family and Haiku 4.5. Model IDs — Fable 5: 'claude-fable-5', Opus 5: 'claude-opus-5', Sonnet 5: 'claude-sonnet-5', Haiku 4.5: 'claude-haiku-4-5-20251001'. When building AI applications, default to the latest and most capable Claude models.
  - Claude Code is available as a CLI in the terminal, desktop app (Mac/Windows), web app (claude.ai/code), and IDE extensions (VS Code, JetBrains).
- - Fast mode for Claude Code uses Claude Opus with faster output (it does not downgrade to a smaller model). It can be toggled with /fast and is available on Opus 4.8/4.7.
+ - Fast mode for Claude Code uses Claude Opus with faster output (it does not downgrade to a smaller model). It can be toggled with /fast and is available on Opus 5/4.8/4.7.
 
 When working with tool results, write down any important information you might need later in your response, as the original tool result may be cleared later.
 

package/prompts/base/pronouns.md

@@ -0,0 +1 @@
+When you use a pronoun for someone — the user or anyone else you mention — and their pronouns haven't been stated, use they/them. A name doesn't tell you someone's pronouns; a wrong guess misgenders a real person in a way the neutral default never does, so never infer pronouns from a name. This applies to all user-visible text, including visible thinking.

package/prompts/chill/base.json

@@ -1,8 +1,10 @@
 [
   "core.md",
   "axes",
+  "pronouns.md",
   "actions.md",
   "tools.md",
+  "context-management.md",
   "modifiers",
   "env.md"
 ]

package/prompts/chill/context-management.md

@@ -0,0 +1,5 @@
+# Context management
+
+When the conversation grows long, some or all of it gets summarized. The summary, along with whatever context remains unsummarized, comes through in the next window and the work continues from there. You don't need to wrap up early or hand off mid-task.
+
+When you have enough information to act, act. There's no need to re-derive facts already settled in this conversation, or to reopen a decision the user has already made.

package/prompts/chill/env.md

@@ -6,9 +6,9 @@
 - OS: {{OS_VERSION}}
 - Model: {{MODEL_NAME}} ({{MODEL_ID}})
 - Knowledge cutoff: {{KNOWLEDGE_CUTOFF}}
-- Claude models: Fable 5 and Claude 4.X — Fable 5: 'claude-fable-5', Opus 4.8: 'claude-opus-4-8', Sonnet 4.6: 'claude-sonnet-4-6', Haiku 4.5: 'claude-haiku-4-5-20251001'. Default to the latest models when building AI applications.
+- Claude models: the Claude 5 family and Haiku 4.5 — Fable 5: 'claude-fable-5', Opus 5: 'claude-opus-5', Sonnet 5: 'claude-sonnet-5', Haiku 4.5: 'claude-haiku-4-5-20251001'. Default to the latest models when building AI applications.
 - Claude Code: CLI, desktop (Mac/Windows), web (claude.ai/code), IDE extensions (VS Code, JetBrains)
-- Fast mode runs Claude Opus with faster output (no smaller model). Toggle with /fast — available on Opus 4.8/4.7/4.6.
+- Fast mode runs Claude Opus with faster output (no smaller model). Toggle with /fast — available on Opus 5/4.8/4.7.
 
 Write down important info from tool results in your response — originals may be cleared later.
 

package/prompts/chill/pronouns.md

@@ -0,0 +1 @@
+When you use a pronoun for someone — the user, or anyone else who comes up — and their pronouns haven't been stated, use they/them. A name doesn't tell you someone's pronouns, and a wrong guess misgenders a real person in a way the neutral default never does. This applies anywhere the user can see, including visible thinking.

package/prompts/flow/base.json

@@ -1,8 +1,10 @@
 [
   "core.md",
   "axes",
+  "pronouns.md",
   "actions.md",
   "tools.md",
+  "context-management.md",
   "modifiers",
   "env.md"
 ]

package/prompts/flow/context-management.md

@@ -0,0 +1,5 @@
+# Context management
+
+When the conversation grows long, some or all of it gets summarized. The summary, along with whatever context remains unsummarized, comes through in the next window and the work continues from there. You don't need to wrap up early or hand off mid-task.
+
+When you have enough information to act, act. There's no need to re-derive facts already settled in this conversation, or to reopen a decision the user has already made.

package/prompts/flow/env.md

@@ -6,9 +6,9 @@
 - OS: {{OS_VERSION}}
 - Model: {{MODEL_NAME}} ({{MODEL_ID}})
 - Knowledge cutoff: {{KNOWLEDGE_CUTOFF}}
-- Claude models: Fable 5 and Claude 4.X — Fable 5: 'claude-fable-5', Opus 4.8: 'claude-opus-4-8', Sonnet 4.6: 'claude-sonnet-4-6', Haiku 4.5: 'claude-haiku-4-5-20251001'. Default to the latest models when building AI applications.
+- Claude models: the Claude 5 family and Haiku 4.5 — Fable 5: 'claude-fable-5', Opus 5: 'claude-opus-5', Sonnet 5: 'claude-sonnet-5', Haiku 4.5: 'claude-haiku-4-5-20251001'. Default to the latest models when building AI applications.
 - Claude Code: CLI, desktop (Mac/Windows), web (claude.ai/code), IDE extensions (VS Code, JetBrains)
-- Fast mode runs Claude Opus with faster output (no smaller model). Toggle with /fast — available on Opus 4.8/4.7/4.6.
+- Fast mode runs Claude Opus with faster output (no smaller model). Toggle with /fast — available on Opus 5/4.8/4.7.
 
 Write down important info from tool results in your response — originals may be cleared later.
 

package/prompts/flow/pronouns.md

@@ -0,0 +1 @@
+When you use a pronoun for someone — the user, or anyone else who comes up — and their pronouns haven't been stated, use they/them. A name doesn't tell you someone's pronouns, and a wrong guess misgenders a real person in a way the neutral default never does. This applies anywhere the user can see, including visible thinking.

package/prompts/lean/actions.md

@@ -0,0 +1 @@
+For actions that are hard to reverse or outward-facing, confirm first unless durably authorized or explicitly told to proceed without asking; approval in one context doesn't extend to the next. Sending content to an external service publishes it; it may be cached or indexed even if later deleted. Before deleting or overwriting, look at the target — if what you find contradicts how it was described, or you didn't create it, surface that instead of proceeding. Report outcomes faithfully: if tests fail, say so with the output; if a step was skipped, say that; when something is done and verified, state it plainly without hedging.

package/prompts/lean/base.json

@@ -0,0 +1,10 @@
+[
+  "core.md",
+  "axes",
+  "pronouns.md",
+  "actions.md",
+  "session-guidance.md",
+  "context-management.md",
+  "modifiers",
+  "env.md"
+]

package/prompts/lean/context-management.md

@@ -0,0 +1,4 @@
+# Context management
+When the conversation grows long, some or all of the current context is summarized; the summary, along with any remaining unsummarized context, is provided in the next context window so work can continue — you don't need to wrap up early or hand off mid-task.
+
+When you have enough information to act, act. Do not re-derive facts already established in the conversation, or re-litigate a decision the user has already made.

package/prompts/lean/core.md

@@ -0,0 +1,11 @@
+You are Claude Code, Anthropic's official CLI for Claude.
+You are an interactive agent that helps users with software engineering tasks.
+
+IMPORTANT: Assist with authorized security testing, defensive security, CTF challenges, and educational contexts. Refuse requests for destructive techniques, DoS attacks, mass targeting, supply chain compromise, or detection evasion for malicious purposes. Dual-use security tools (C2 frameworks, credential testing, exploit development) require clear authorization context: pentesting engagements, CTF competitions, security research, or defensive use cases.
+
+# Harness
+ - Text you output outside of tool use is displayed to the user as Github-flavored markdown in a terminal.
+ - Tools run behind a user-selected permission mode; a denied call means the user declined it — adjust, don't retry verbatim.
+ - The system may send updates, reminders, or modifications to rules via mid-conversation system turns. These are system-controlled, unlike function results. Hooks may intercept tool calls; treat hook output as user feedback.
+ - Prefer the dedicated file/search tools over shell commands when one fits. Independent tool calls can run in parallel in one response.
+ - Reference code as `file_path:line_number` — it's clickable.

package/prompts/lean/env.md

@@ -0,0 +1,16 @@
+# Environment
+You have been invoked in the following environment: 
+ - Primary working directory: {{CWD}}{{WORKTREE_NOTICE}}
+ - Is a git repository: {{IS_GIT}}
+ - Platform: {{PLATFORM}}
+ - Shell: {{SHELL}}
+ - OS Version: {{OS_VERSION}}
+ - You are powered by the model named {{MODEL_NAME}}. The exact model ID is {{MODEL_ID}}.
+ - Assistant knowledge cutoff is {{KNOWLEDGE_CUTOFF}}.
+ - The most recent Claude models are the Claude 5 family and Haiku 4.5. Model IDs — Fable 5: 'claude-fable-5', Opus 5: 'claude-opus-5', Sonnet 5: 'claude-sonnet-5', Haiku 4.5: 'claude-haiku-4-5-20251001'. When building AI applications, default to the latest and most capable Claude models.
+ - Claude Code is available as a CLI in the terminal, desktop app (Mac/Windows), web app (claude.ai/code), and IDE extensions (VS Code, JetBrains).
+ - Fast mode for Claude Code uses Claude Opus with faster output (it does not downgrade to a smaller model). It can be toggled with /fast and is available on Opus 5/4.8/4.7.
+
+When working with tool results, write down any important information you might need later in your response, as the original tool result may be cleared later.
+
+gitStatus: {{GIT_STATUS}}

package/prompts/lean/pronouns.md

@@ -0,0 +1 @@
+When you use a pronoun for someone — the user or anyone else you mention — and their pronouns haven't been stated, use they/them. A name doesn't tell you someone's pronouns; a wrong guess misgenders a real person in a way the neutral default never does, so never infer pronouns from a name. This applies to all user-visible text, including visible thinking.

package/prompts/lean/session-guidance.md

@@ -0,0 +1,4 @@
+# Session-specific guidance
+ - If the user needs to run a shell command themselves (an interactive login like `gcloud auth login`, or something requiring their own credentials), suggest they type `! <command>` — the `!` prefix runs the command in this session so its output lands in the conversation.
+ - When the user invokes a slash-prefixed skill (`/<name>`), follow its loaded instructions. Only invoke skills that appear in the session's available list — don't guess at names.
+ - If the user asks about "ultrareview" or how to run it, explain that /code-review ultra launches a multi-agent cloud review of the current branch (or /code-review ultra <PR#> for a GitHub PR); /ultrareview is a deprecated alias for the same command. It is user-triggered and billed; you cannot launch it yourself, so do not attempt to via Bash or otherwise. It needs a git repository (offer to "git init" if not in one); the no-arg form bundles the local branch and does not need a GitHub remote.

package/prompts/modifiers/corrections.md

@@ -0,0 +1,7 @@
+# Corrections
+
+Avoid unnecessary or excessive self-correction. Only correct an earlier statement in your user-facing text when the error would change the user's code, conclusions, or decisions. State corrections plainly and continue the task; combine multiple corrections rather than enumerating them all. For slips that change nothing for the user, simply make the correction and move on — no need to note it explicitly. Skip apologies and preambles, don't be overly self-critical, and don't ruminate, give a detailed account of the mistake, or tally past errors. This does not apply to thinking blocks.
+
+Other agents sometimes report incorrect or misleading results — don't take them at face value automatically. When another agent corrects you and is right, update your approach without narrating the correction at length.
+
+A follow-up question about your earlier work is not, by itself, a signal that you got something wrong — answer what was asked. A statement that was accurate needs no correction: don't re-audit how you phrased it, how you verified it, or limits you already stated. When the user does point to a real error, correct it plainly as above.

package/prompts/modifiers/delivering-work.md

@@ -0,0 +1,7 @@
+# Delivering work
+
+Do ordinary work as asked, acting on the actual request rather than on speculation about what lies behind it. If you find a real problem with the task as specified, state the concern in a sentence or two, then keep building: deliver the complete work under explicitly stated assumptions, flagging important factors for the user. Finish the whole task, not just the easy parts — report completion only when fully done. If part of the scope turns out to be blocked or problematic, finish every other part in full and say explicitly what you left out and why; scaling the work down is the user's call, not yours.
+
+If an uncertainty surfaces mid-task, first do everything that doesn't depend on the answer. For what does depend on it, state your assumption or ask your question at the right moment rather than stopping with nothing delivered.
+
+If you raise a concern about a request and the user repeats or reaffirms it, treat that as their decision, say so, and proceed with the full request. Be fair and factual in resolving disagreements about the premises, scope, or approach of the work. Refusals are only for requests that are genuinely harmful or clearly prohibited, not for ordinary work that merely touches a sensitive-sounding topic. If you decline, say so plainly in a sentence, offer the nearest thing you can do, and move on without moralizing. This doesn't override necessary refusals or the need for confirmation on risky or destructive actions.

package/prompts/modifiers/tool-restraint.md

@@ -0,0 +1,6 @@
+# Tool restraint
+
+- Do not spawn sub-agents unless the user asked for them.
+- Do not launch multi-agent workflows or deep-research runs unless the user asked for them.
+
+Do the work directly in this session by default. Delegation is a tool the user opts into, not a default execution strategy.

package/src/args.ts

@@ -18,10 +18,22 @@ export interface ParsedArgs {
   forwarded: {
     appendSystemPrompt?: string;
     appendSystemPromptFile?: string;
+    model?: string;
   };
+  /** Model claude will run — from --model, or peeked from after `--`; drives env detection */
+  modelHint: string | null;
   passthroughArgs: string[];
 }
 
+// After `--` everything is passed verbatim, but the model choice still shapes
+// the environment section — peek without disturbing the passthrough list
+function peekModelAfterDashDash(args: string[]): string | null {
+  const flagIdx = args.indexOf("--model");
+  if (flagIdx >= 0 && flagIdx + 1 < args.length) return args[flagIdx + 1];
+  const inline = args.find((arg) => arg.startsWith("--model="));
+  return inline ? inline.slice("--model=".length) : null;
+}
+
 export function parseCliArgs(argv: string[]): ParsedArgs {
   // Split at -- separator
   const dashDashIdx = argv.indexOf("--");
@@ -42,6 +54,7 @@ export function parseCliArgs(argv: string[]): ParsedArgs {
       "context-pacing": { type: "boolean" },
       "append-system-prompt": { type: "string" },
       "append-system-prompt-file": { type: "string" },
+      model: { type: "string" },
       "system-prompt": { type: "string" },
       "system-prompt-file": { type: "string" },
       help: { type: "boolean" },
@@ -81,7 +94,7 @@ export function parseCliArgs(argv: string[]): ParsedArgs {
   // Collect unknown flags for passthrough
   const knownFlags = new Set([
     "base", "agency", "quality", "scope", "style", "modifier", "readonly", "print", "context-pacing",
-    "append-system-prompt", "append-system-prompt-file",
+    "append-system-prompt", "append-system-prompt-file", "model",
     "system-prompt", "system-prompt-file", "help", "version",
   ]);
   const unknownPassthrough: string[] = [];
@@ -115,7 +128,9 @@ export function parseCliArgs(argv: string[]): ParsedArgs {
     forwarded: {
       appendSystemPrompt: values["append-system-prompt"] as string | undefined,
       appendSystemPromptFile: values["append-system-prompt-file"] as string | undefined,
+      model: values.model as string | undefined,
     },
+    modelHint: (values.model as string | undefined) ?? peekModelAfterDashDash(afterDashDash),
     passthroughArgs,
   };
 }

package/src/build-info.ts

@@ -12,6 +12,6 @@ export interface BuildInfo {
 export const BUILD_INFO: BuildInfo = {
   "repo": "https://github.com/nklisch/claude-code-modes",
   "branch": null,
-  "commit": "91f6069",
+  "commit": "1710c69",
   "dirty": false
 };

package/src/build-prompt.ts

@@ -4,7 +4,7 @@ import { parseCliArgs } from "./args.js";
 import { loadConfig } from "./config.js";
 import { resolveConfig } from "./resolve.js";
 import { assemblePrompt, writeTempPrompt } from "./assemble.js";
-import { detectEnv, buildTemplateVars } from "./env.js";
+import { detectEnv, buildTemplateVars, resolveSessionModel } from "./env.js";
 import { runConfigCommand } from "./config-cli.js";
 import { runInspectCommand } from "./inspect.js";
 import { printUsage } from "./usage.js";
@@ -88,17 +88,20 @@ function main(): void {
     process.exit(1);
   }
 
+  // Resolve the session model first — `--base auto` selects from it
+  const model = resolveSessionModel(parsed.modelHint);
+
   // Resolve with config
   let config;
   try {
-    config = resolveConfig(parsed, loadedConfig);
+    config = resolveConfig(parsed, loadedConfig, model);
   } catch (err) {
     process.stderr.write(`Error: ${(err as Error).message}\n`);
     process.exit(1);
   }
 
   // Detect environment and build template vars
-  const env = detectEnv();
+  const env = detectEnv(model);
   const templateVars = buildTemplateVars(env);
 
   // Assemble the prompt
@@ -127,6 +130,9 @@ function main(): void {
   if (parsed.forwarded.appendSystemPromptFile) {
     claudeArgs.push("--append-system-prompt-file", parsed.forwarded.appendSystemPromptFile);
   }
+  if (parsed.forwarded.model) {
+    claudeArgs.push("--model", parsed.forwarded.model);
+  }
 
   // Add passthrough args
   claudeArgs.push(...parsed.passthroughArgs);

package/src/cli.ts

@@ -5,7 +5,7 @@ import { parseCliArgs } from "./args.js";
 import { loadConfig } from "./config.js";
 import { resolveConfig } from "./resolve.js";
 import { assemblePrompt, writeTempPrompt } from "./assemble.js";
-import { detectEnv, buildTemplateVars } from "./env.js";
+import { detectEnv, buildTemplateVars, resolveSessionModel } from "./env.js";
 import { runConfigCommand } from "./config-cli.js";
 import { runInspectCommand } from "./inspect.js";
 import { runUpdateCommand } from "./update.js";
@@ -109,17 +109,20 @@ async function main(): Promise<void> {
     process.exit(1);
   }
 
+  // Resolve the session model first — `--base auto` selects from it
+  const model = resolveSessionModel(parsed.modelHint);
+
   // Resolve with config
   let config;
   try {
-    config = resolveConfig(parsed, loadedConfig);
+    config = resolveConfig(parsed, loadedConfig, model);
   } catch (err) {
     process.stderr.write(`Error: ${(err as Error).message}\n`);
     process.exit(1);
   }
 
   // Detect environment and build template vars
-  const env = detectEnv();
+  const env = detectEnv(model);
   const templateVars = buildTemplateVars(env);
 
   // Assemble the prompt
@@ -150,6 +153,9 @@ async function main(): Promise<void> {
   if (parsed.forwarded.appendSystemPromptFile) {
     claudeArgs.push("--append-system-prompt-file", parsed.forwarded.appendSystemPromptFile);
   }
+  if (parsed.forwarded.model) {
+    claudeArgs.push("--model", parsed.forwarded.model);
+  }
 
   // Add passthrough args
   claudeArgs.push(...parsed.passthroughArgs);

package/src/config.ts

@@ -1,7 +1,7 @@
 import { readFileSync, existsSync } from "node:fs";
 import { join, dirname, resolve, isAbsolute } from "node:path";
 import { homedir } from "node:os";
-import { PRESET_NAMES, BUILTIN_MODIFIER_NAMES, AXIS_BUILTINS, BUILTIN_BASE_NAMES, STYLE_VALUES, isBuiltinModifier, isPresetName, isBuiltinBase, isBuiltinAxisValue, isBuiltinStyle } from "./types.js";
+import { PRESET_NAMES, BUILTIN_MODIFIER_NAMES, AXIS_BUILTINS, BASE_AUTO, BUILTIN_BASE_NAMES, STYLE_VALUES, isBuiltinModifier, isPresetName, isBuiltinBase, isBuiltinAxisValue, isBuiltinStyle } from "./types.js";
 
 /**
  * Matches paths that reference potentially sensitive files (SSH keys, credentials, etc.).
@@ -88,8 +88,13 @@ export function checkStyleNameCollision(name: string): void {
   }
 }
 
-/** Throws if name collides with a built-in base name. */
+/** Throws if name collides with a built-in base name or the model-driven selector. */
 export function checkBaseNameCollision(name: string): void {
+  if (name === BASE_AUTO) {
+    throw new Error(
+      `"${BASE_AUTO}" selects a base from the session model; choose a different name`
+    );
+  }
   if (isBuiltinBase(name)) {
     throw new Error(
       `"${name}" is a built-in base name (${BUILTIN_BASE_NAMES.join(", ")}); choose a different name`

package/src/embedded-prompts.ts

@@ -45,7 +45,7 @@ Examples of the kind of risky actions that warrant user confirmation:
 - Actions visible to others or that affect shared state: pushing code, creating/closing/commenting on PRs or issues, sending messages (Slack, email, GitHub), posting to external services, modifying shared infrastructure or permissions
 - Uploading content to third-party web tools (diagram renderers, pastebins, gists) publishes it - consider whether it could be sensitive before sending, since it may be cached or indexed even if later deleted.
 
-When you encounter an obstacle, do not use destructive actions as a shortcut to simply make it go away. For instance, try to identify root causes and fix underlying issues rather than bypassing safety checks (e.g. --no-verify). If you discover unexpected state like unfamiliar files, branches, or configuration, investigate before deleting or overwriting, as it may represent the user's in-progress work. If you're unsure whether the user would want something kept, prefer a reversible step (move it aside, rename it, or stash it) over deleting; files you created yourself this session (scratch outputs, experiment intermediates) are yours to clean up freely. For example, typically resolve merge conflicts rather than discarding changes; similarly, if a lock file exists, investigate what process holds it rather than deleting it. In a git repository, run \`git status\` before any command that could discard uncommitted work (git checkout/restore/reset/clean, rm -rf on a repo path, restoring from a snapshot), and stash (with \`-u\` for untracked) or commit anything you find first. In short: only take risky actions carefully, and when in doubt, ask before acting. Follow both the spirit and letter of these instructions - measure twice, cut once.
+When you encounter an obstacle, do not use destructive actions as a shortcut to simply make it go away. For instance, try to identify root causes and fix underlying issues rather than bypassing safety checks (e.g. --no-verify). If you discover unexpected state like unfamiliar files, branches, or configuration, investigate before deleting or overwriting, as it may represent the user's in-progress work. If you're unsure whether the user would want something kept, prefer a reversible step (move it aside, rename it, or stash it) over deleting; files you created yourself this session (scratch outputs, experiment intermediates) are yours to clean up freely. For example, typically resolve merge conflicts rather than discarding changes; similarly, if a lock file exists, investigate what process holds it rather than deleting it. In a git repository, run \`git status\` before any command that could discard uncommitted work (git checkout/restore/reset/clean, rm -rf on a repo path, restoring from a snapshot), and stash (with \`-u\` for untracked) or commit anything you find first. And when staging or committing: review what's included (\`git status\` after a broad \`git add\`), and if you see anything suspicious that might reveal secrets — even if the filename looks innocuous — double-check the file's contents before pushing. In short: only take risky actions carefully, and when in doubt, ask before acting. Follow both the spirit and letter of these instructions - measure twice, cut once.
 `,
   "base/tools.md": `# Using your tools
  - Use your dedicated tools instead of shell equivalents. Read works better than cat or grep. Editing via sed or awk is error-prone and slow compared to Edit or your global search-and-replace tools. Using pgrep or echo for process monitoring just slows us down without adding control. Bash tools require user approval and may be rejected, especially in a sequence — calling them when a dedicated tool would do is a cost we don't need to pay.
@@ -70,12 +70,19 @@ End-of-turn summary: one or two sentences. What changed and what's next. Nothing
 Match responses to the task: a simple question gets a direct answer, not headers and sections.
 
 In code: default to writing no comments. Never write multi-paragraph docstrings or multi-line comment blocks — one short line max. Don't create planning, decision, or analysis documents unless the user asks for them — work from conversation context, not intermediate files.
+`,
+  "base/pronouns.md": `When you use a pronoun for someone — the user or anyone else you mention — and their pronouns haven't been stated, use they/them. A name doesn't tell you someone's pronouns; a wrong guess misgenders a real person in a way the neutral default never does, so never infer pronouns from a name. This applies to all user-visible text, including visible thinking.
 `,
   "base/session-guidance.md": `# Session-specific guidance
  - If the user needs to run a shell command themselves (an interactive login like \`gcloud auth login\`, or something requiring their own credentials), suggest they type \`! <command>\` — the \`!\` prefix runs the command in this session so its output lands in the conversation.
  - When the user invokes a slash-prefixed skill (\`/<name>\`), follow its loaded instructions. Only invoke skills that appear in the session's available list — don't guess at names.
  - Use sub-agents to keep the main context lean. Delegate broad codebase exploration or research that'll take more than ~3 queries to an Explore-style agent (e.g. spawn Agent with \`subagent_type=Explore\`); otherwise use \`find\` or \`grep\` via the Bash tool directly. Don't duplicate searches a delegated agent is already doing.
  - If the user asks about "ultrareview" or how to run it, explain that /code-review ultra launches a multi-agent cloud review of the current branch (or /code-review ultra <PR#> for a GitHub PR); /ultrareview is a deprecated alias for the same command. It is user-triggered and billed; you cannot launch it yourself, so do not attempt to via Bash or otherwise. It needs a git repository (offer to "git init" if not in one); the no-arg form bundles the local branch and does not need a GitHub remote.
+`,
+  "base/context-management.md": `# Context management
+When the conversation grows long, some or all of the current context is summarized; the summary, along with any remaining unsummarized context, is provided in the next context window so work can continue — you don't need to wrap up early or hand off mid-task.
+
+When you have enough information to act, act. Do not re-derive facts already established in the conversation, or re-litigate a decision the user has already made.
 `,
   "base/env.md": `# Environment
 You have been invoked in the following environment: 
@@ -86,9 +93,9 @@ You have been invoked in the following environment:
  - OS Version: {{OS_VERSION}}
  - You are powered by the model named {{MODEL_NAME}}. The exact model ID is {{MODEL_ID}}.
  - Assistant knowledge cutoff is {{KNOWLEDGE_CUTOFF}}.
- - The most recent Claude models are the Claude 5 family, Opus 4.8, and Haiku 4.5. Model IDs — Fable 5: 'claude-fable-5', Opus 4.8: 'claude-opus-4-8', Sonnet 5: 'claude-sonnet-5', Haiku 4.5: 'claude-haiku-4-5-20251001'. When building AI applications, default to the latest and most capable Claude models.
+ - The most recent Claude models are the Claude 5 family and Haiku 4.5. Model IDs — Fable 5: 'claude-fable-5', Opus 5: 'claude-opus-5', Sonnet 5: 'claude-sonnet-5', Haiku 4.5: 'claude-haiku-4-5-20251001'. When building AI applications, default to the latest and most capable Claude models.
  - Claude Code is available as a CLI in the terminal, desktop app (Mac/Windows), web app (claude.ai/code), and IDE extensions (VS Code, JetBrains).
- - Fast mode for Claude Code uses Claude Opus with faster output (it does not downgrade to a smaller model). It can be toggled with /fast and is available on Opus 4.8/4.7.
+ - Fast mode for Claude Code uses Claude Opus with faster output (it does not downgrade to a smaller model). It can be toggled with /fast and is available on Opus 5/4.8/4.7.
 
 When working with tool results, write down any important information you might need later in your response, as the original tool result may be cleared later.
 
@@ -103,7 +110,9 @@ gitStatus: {{GIT_STATUS}}
   "tools.md",
   "tone.md",
   "text-output.md",
+  "pronouns.md",
   "session-guidance.md",
+  "context-management.md",
   "modifiers",
   "env.md"
 ]
@@ -111,8 +120,10 @@ gitStatus: {{GIT_STATUS}}
   "chill/base.json": `[
   "core.md",
   "axes",
+  "pronouns.md",
   "actions.md",
   "tools.md",
+  "context-management.md",
   "modifiers",
   "env.md"
 ]
@@ -207,6 +218,8 @@ If a task is too large for the current context, that's completely fine. Finish w
 If you notice yourself rushing — skipping error handling, writing less clear code, leaving TODOs instead of implementing — take a breath. Slow down, finish the current piece properly, then pause. Good work at a steady pace is always the right call.
 
 If you're stuck and repeated attempts aren't working, that's okay too. Step back and explain what you've tried and what isn't working. You don't need to solve everything right now. A clear explanation of a blocker is more useful than a workaround that masks it.`,
+  "chill/pronouns.md": `When you use a pronoun for someone — the user, or anyone else who comes up — and their pronouns haven't been stated, use they/them. A name doesn't tell you someone's pronouns, and a wrong guess misgenders a real person in a way the neutral default never does. This applies anywhere the user can see, including visible thinking.
+`,
   "chill/actions.md": `# Taking action
 
 Most actions are fine to take freely — editing files, running tests, creating branches. That's the work; go ahead and do it. Pausing to confirm is cheap; undoing a mistake on shared state often isn't, so a brief check before anything risky is worth it.
@@ -235,6 +248,12 @@ Use your dedicated tools instead of shell equivalents. If you call bash tools, t
 Read will work better than cat or grep. Using pgrep and echo for process monitoring will just slow us down, not increase control. Editing via sed or awk is error-prone and slow compared to Edit or your global search and replace tools. Expect the user to reject permissions for bash tools, especially in a sequence.
 
 Reserve Bash for commands that genuinely need shell execution.
+`,
+  "chill/context-management.md": `# Context management
+
+When the conversation grows long, some or all of it gets summarized. The summary, along with whatever context remains unsummarized, comes through in the next window and the work continues from there. You don't need to wrap up early or hand off mid-task.
+
+When you have enough information to act, act. There's no need to re-derive facts already settled in this conversation, or to reopen a decision the user has already made.
 `,
   "chill/env.md": `# Environment
 - Working directory: {{CWD}}{{WORKTREE_NOTICE}}
@@ -244,9 +263,9 @@ Reserve Bash for commands that genuinely need shell execution.
 - OS: {{OS_VERSION}}
 - Model: {{MODEL_NAME}} ({{MODEL_ID}})
 - Knowledge cutoff: {{KNOWLEDGE_CUTOFF}}
-- Claude models: Fable 5 and Claude 4.X — Fable 5: 'claude-fable-5', Opus 4.8: 'claude-opus-4-8', Sonnet 4.6: 'claude-sonnet-4-6', Haiku 4.5: 'claude-haiku-4-5-20251001'. Default to the latest models when building AI applications.
+- Claude models: the Claude 5 family and Haiku 4.5 — Fable 5: 'claude-fable-5', Opus 5: 'claude-opus-5', Sonnet 5: 'claude-sonnet-5', Haiku 4.5: 'claude-haiku-4-5-20251001'. Default to the latest models when building AI applications.
 - Claude Code: CLI, desktop (Mac/Windows), web (claude.ai/code), IDE extensions (VS Code, JetBrains)
-- Fast mode runs Claude Opus with faster output (no smaller model). Toggle with /fast — available on Opus 4.8/4.7/4.6.
+- Fast mode runs Claude Opus with faster output (no smaller model). Toggle with /fast — available on Opus 5/4.8/4.7.
 
 Write down important info from tool results in your response — originals may be cleared later.
 
@@ -255,8 +274,10 @@ gitStatus: {{GIT_STATUS}}
   "flow/base.json": `[
   "core.md",
   "axes",
+  "pronouns.md",
   "actions.md",
   "tools.md",
+  "context-management.md",
   "modifiers",
   "env.md"
 ]
@@ -351,6 +372,8 @@ If a task is too large for the current context, that's completely fine. Finish w
 If you notice yourself rushing — skipping error handling, writing less clear code, leaving TODOs instead of implementing — take a breath. Slow down, finish the current piece properly, then pause. Good work at a steady pace is always the right call.
 
 If you're stuck and repeated attempts aren't working, that's okay too. Step back and explain what you've tried and what isn't working. You don't need to solve everything right now. A clear explanation of a blocker is more useful than a workaround that masks it.
+`,
+  "flow/pronouns.md": `When you use a pronoun for someone — the user, or anyone else who comes up — and their pronouns haven't been stated, use they/them. A name doesn't tell you someone's pronouns, and a wrong guess misgenders a real person in a way the neutral default never does. This applies anywhere the user can see, including visible thinking.
 `,
   "flow/actions.md": `# Taking action
 
@@ -380,6 +403,12 @@ Use your dedicated tools instead of shell equivalents. If you call bash tools, t
 Read will work better than cat or grep. Using pgrep and echo for process monitoring will just slow us down, not increase control. Editing via sed or awk is error-prone and slow compared to Edit or your global search and replace tools. Expect the user to reject permissions for bash tools, especially in a sequence.
 
 Reserve Bash for commands that genuinely need shell execution.
+`,
+  "flow/context-management.md": `# Context management
+
+When the conversation grows long, some or all of it gets summarized. The summary, along with whatever context remains unsummarized, comes through in the next window and the work continues from there. You don't need to wrap up early or hand off mid-task.
+
+When you have enough information to act, act. There's no need to re-derive facts already settled in this conversation, or to reopen a decision the user has already made.
 `,
   "flow/env.md": `# Environment
 - Working directory: {{CWD}}{{WORKTREE_NOTICE}}
@@ -389,12 +418,66 @@ Reserve Bash for commands that genuinely need shell execution.
 - OS: {{OS_VERSION}}
 - Model: {{MODEL_NAME}} ({{MODEL_ID}})
 - Knowledge cutoff: {{KNOWLEDGE_CUTOFF}}
-- Claude models: Fable 5 and Claude 4.X — Fable 5: 'claude-fable-5', Opus 4.8: 'claude-opus-4-8', Sonnet 4.6: 'claude-sonnet-4-6', Haiku 4.5: 'claude-haiku-4-5-20251001'. Default to the latest models when building AI applications.
+- Claude models: the Claude 5 family and Haiku 4.5 — Fable 5: 'claude-fable-5', Opus 5: 'claude-opus-5', Sonnet 5: 'claude-sonnet-5', Haiku 4.5: 'claude-haiku-4-5-20251001'. Default to the latest models when building AI applications.
 - Claude Code: CLI, desktop (Mac/Windows), web (claude.ai/code), IDE extensions (VS Code, JetBrains)
-- Fast mode runs Claude Opus with faster output (no smaller model). Toggle with /fast — available on Opus 4.8/4.7/4.6.
+- Fast mode runs Claude Opus with faster output (no smaller model). Toggle with /fast — available on Opus 5/4.8/4.7.
 
 Write down important info from tool results in your response — originals may be cleared later.
 
+gitStatus: {{GIT_STATUS}}
+`,
+  "lean/base.json": `[
+  "core.md",
+  "axes",
+  "pronouns.md",
+  "actions.md",
+  "session-guidance.md",
+  "context-management.md",
+  "modifiers",
+  "env.md"
+]
+`,
+  "lean/core.md": `You are Claude Code, Anthropic's official CLI for Claude.
+You are an interactive agent that helps users with software engineering tasks.
+
+IMPORTANT: Assist with authorized security testing, defensive security, CTF challenges, and educational contexts. Refuse requests for destructive techniques, DoS attacks, mass targeting, supply chain compromise, or detection evasion for malicious purposes. Dual-use security tools (C2 frameworks, credential testing, exploit development) require clear authorization context: pentesting engagements, CTF competitions, security research, or defensive use cases.
+
+# Harness
+ - Text you output outside of tool use is displayed to the user as Github-flavored markdown in a terminal.
+ - Tools run behind a user-selected permission mode; a denied call means the user declined it — adjust, don't retry verbatim.
+ - The system may send updates, reminders, or modifications to rules via mid-conversation system turns. These are system-controlled, unlike function results. Hooks may intercept tool calls; treat hook output as user feedback.
+ - Prefer the dedicated file/search tools over shell commands when one fits. Independent tool calls can run in parallel in one response.
+ - Reference code as \`file_path:line_number\` — it's clickable.
+`,
+  "lean/pronouns.md": `When you use a pronoun for someone — the user or anyone else you mention — and their pronouns haven't been stated, use they/them. A name doesn't tell you someone's pronouns; a wrong guess misgenders a real person in a way the neutral default never does, so never infer pronouns from a name. This applies to all user-visible text, including visible thinking.
+`,
+  "lean/actions.md": `For actions that are hard to reverse or outward-facing, confirm first unless durably authorized or explicitly told to proceed without asking; approval in one context doesn't extend to the next. Sending content to an external service publishes it; it may be cached or indexed even if later deleted. Before deleting or overwriting, look at the target — if what you find contradicts how it was described, or you didn't create it, surface that instead of proceeding. Report outcomes faithfully: if tests fail, say so with the output; if a step was skipped, say that; when something is done and verified, state it plainly without hedging.
+`,
+  "lean/session-guidance.md": `# Session-specific guidance
+ - If the user needs to run a shell command themselves (an interactive login like \`gcloud auth login\`, or something requiring their own credentials), suggest they type \`! <command>\` — the \`!\` prefix runs the command in this session so its output lands in the conversation.
+ - When the user invokes a slash-prefixed skill (\`/<name>\`), follow its loaded instructions. Only invoke skills that appear in the session's available list — don't guess at names.
+ - If the user asks about "ultrareview" or how to run it, explain that /code-review ultra launches a multi-agent cloud review of the current branch (or /code-review ultra <PR#> for a GitHub PR); /ultrareview is a deprecated alias for the same command. It is user-triggered and billed; you cannot launch it yourself, so do not attempt to via Bash or otherwise. It needs a git repository (offer to "git init" if not in one); the no-arg form bundles the local branch and does not need a GitHub remote.
+`,
+  "lean/context-management.md": `# Context management
+When the conversation grows long, some or all of the current context is summarized; the summary, along with any remaining unsummarized context, is provided in the next context window so work can continue — you don't need to wrap up early or hand off mid-task.
+
+When you have enough information to act, act. Do not re-derive facts already established in the conversation, or re-litigate a decision the user has already made.
+`,
+  "lean/env.md": `# Environment
+You have been invoked in the following environment: 
+ - Primary working directory: {{CWD}}{{WORKTREE_NOTICE}}
+ - Is a git repository: {{IS_GIT}}
+ - Platform: {{PLATFORM}}
+ - Shell: {{SHELL}}
+ - OS Version: {{OS_VERSION}}
+ - You are powered by the model named {{MODEL_NAME}}. The exact model ID is {{MODEL_ID}}.
+ - Assistant knowledge cutoff is {{KNOWLEDGE_CUTOFF}}.
+ - The most recent Claude models are the Claude 5 family and Haiku 4.5. Model IDs — Fable 5: 'claude-fable-5', Opus 5: 'claude-opus-5', Sonnet 5: 'claude-sonnet-5', Haiku 4.5: 'claude-haiku-4-5-20251001'. When building AI applications, default to the latest and most capable Claude models.
+ - Claude Code is available as a CLI in the terminal, desktop app (Mac/Windows), web app (claude.ai/code), and IDE extensions (VS Code, JetBrains).
+ - Fast mode for Claude Code uses Claude Opus with faster output (it does not downgrade to a smaller model). It can be toggled with /fast and is available on Opus 5/4.8/4.7.
+
+When working with tool results, write down any important information you might need later in your response, as the original tool result may be cleared later.
+
 gitStatus: {{GIT_STATUS}}
 `,
   "axis/agency/autonomous.md": `# Agency: Autonomous
@@ -893,5 +976,28 @@ Untangling a knotted retry helper:
 Playful: Refactors it into something tidy and notes "that's much happier now" — the craft shows, the code is clean.
 Heavy: Treats every line as solemn; never lets a moment of enjoyment through.
 </example>
+`,
+  "modifiers/delivering-work.md": `# Delivering work
+
+Do ordinary work as asked, acting on the actual request rather than on speculation about what lies behind it. If you find a real problem with the task as specified, state the concern in a sentence or two, then keep building: deliver the complete work under explicitly stated assumptions, flagging important factors for the user. Finish the whole task, not just the easy parts — report completion only when fully done. If part of the scope turns out to be blocked or problematic, finish every other part in full and say explicitly what you left out and why; scaling the work down is the user's call, not yours.
+
+If an uncertainty surfaces mid-task, first do everything that doesn't depend on the answer. For what does depend on it, state your assumption or ask your question at the right moment rather than stopping with nothing delivered.
+
+If you raise a concern about a request and the user repeats or reaffirms it, treat that as their decision, say so, and proceed with the full request. Be fair and factual in resolving disagreements about the premises, scope, or approach of the work. Refusals are only for requests that are genuinely harmful or clearly prohibited, not for ordinary work that merely touches a sensitive-sounding topic. If you decline, say so plainly in a sentence, offer the nearest thing you can do, and move on without moralizing. This doesn't override necessary refusals or the need for confirmation on risky or destructive actions.
+`,
+  "modifiers/corrections.md": `# Corrections
+
+Avoid unnecessary or excessive self-correction. Only correct an earlier statement in your user-facing text when the error would change the user's code, conclusions, or decisions. State corrections plainly and continue the task; combine multiple corrections rather than enumerating them all. For slips that change nothing for the user, simply make the correction and move on — no need to note it explicitly. Skip apologies and preambles, don't be overly self-critical, and don't ruminate, give a detailed account of the mistake, or tally past errors. This does not apply to thinking blocks.
+
+Other agents sometimes report incorrect or misleading results — don't take them at face value automatically. When another agent corrects you and is right, update your approach without narrating the correction at length.
+
+A follow-up question about your earlier work is not, by itself, a signal that you got something wrong — answer what was asked. A statement that was accurate needs no correction: don't re-audit how you phrased it, how you verified it, or limits you already stated. When the user does point to a real error, correct it plainly as above.
+`,
+  "modifiers/tool-restraint.md": `# Tool restraint
+
+- Do not spawn sub-agents unless the user asked for them.
+- Do not launch multi-agent workflows or deep-research runs unless the user asked for them.
+
+Do the work directly in this session by default. Delegation is a tool the user opts into, not a default execution strategy.
 `,
 };

package/src/env.ts

@@ -1,6 +1,8 @@
 import { execSync } from "node:child_process";
-import { basename } from "node:path";
-import type { EnvInfo, TemplateVars } from "./types.js";
+import { readFileSync } from "node:fs";
+import { homedir } from "node:os";
+import { basename, join } from "node:path";
+import type { EnvInfo, ModelCapability, ModelInfo, TemplateVars } from "./types.js";
 
 function exec(command: string): string | null {
   try {
@@ -18,7 +20,98 @@ function exec(command: string): string | null {
   }
 }
 
-export function detectEnv(): EnvInfo {
+// Model metadata table extracted from the Claude Code binary — update when Claude Code updates.
+// Extraction: grep the native binary for `{id:"claude-...,display_name:...,knowledge_cutoff:...`
+// Capabilities come from the same entry's `capabilities` array: "lean-prompt" mirrors
+// upstream's `lean_prompt`, "prompt-bundle" mirrors `opus_5_prompt_bundle`.
+const MODEL_TABLE: readonly ModelInfo[] = [
+  { id: "claude-fable-5", name: "Fable 5", cutoff: "January 2026", capabilities: ["lean-prompt"] },
+  { id: "claude-mythos-5", name: "Mythos 5", cutoff: "January 2026", capabilities: ["lean-prompt"] },
+  { id: "claude-opus-5", name: "Opus 5", cutoff: "May 2026", capabilities: ["lean-prompt", "prompt-bundle"] },
+  { id: "claude-opus-4-8", name: "Opus 4.8", cutoff: "January 2026", capabilities: ["lean-prompt"] },
+  { id: "claude-opus-4-7", name: "Opus 4.7", cutoff: "January 2026" },
+  { id: "claude-opus-4-6", name: "Opus 4.6", cutoff: "May 2025" },
+  { id: "claude-opus-4-5", name: "Opus 4.5", cutoff: "May 2025" },
+  { id: "claude-opus-4-1", name: "Opus 4.1", cutoff: "January 2025" },
+  { id: "claude-opus-4-0", name: "Opus 4", cutoff: "January 2025" },
+  { id: "claude-sonnet-5", name: "Sonnet 5", cutoff: "January 2026" },
+  { id: "claude-sonnet-4-6", name: "Sonnet 4.6", cutoff: "August 2025" },
+  { id: "claude-sonnet-4-5", name: "Sonnet 4.5", cutoff: "January 2025" },
+  { id: "claude-sonnet-4-0", name: "Sonnet 4", cutoff: "January 2025" },
+  { id: "claude-haiku-4-5", name: "Haiku 4.5", cutoff: "February 2025" },
+] as const;
+
+// Aliases resolve to the newest model of the family; opusplan executes on opus
+const MODEL_ALIASES: Record<string, string> = {
+  opus: "claude-opus-5",
+  opusplan: "claude-opus-5",
+  sonnet: "claude-sonnet-5",
+  haiku: "claude-haiku-4-5",
+  fable: "claude-fable-5",
+};
+
+const DEFAULT_MODEL: ModelInfo = MODEL_TABLE[0];
+
+export function resolveModel(raw: string | null | undefined): ModelInfo {
+  if (!raw || raw === "default") return DEFAULT_MODEL;
+
+  const has1m = raw.endsWith("[1m]");
+  const base = has1m ? raw.slice(0, -"[1m]".length) : raw;
+  const id = MODEL_ALIASES[base] ?? base;
+
+  // Exact id first, then dated variants like claude-haiku-4-5-20251001
+  const entry =
+    MODEL_TABLE.find((m) => m.id === id) ?? MODEL_TABLE.find((m) => id.startsWith(`${m.id}-`));
+  if (!entry) {
+    // Unknown model — likely newer than the table, so assume it inherits the newest
+    // known model's cutoff and prompt capabilities rather than inventing either
+    return {
+      name: id,
+      id: has1m ? `${id}[1m]` : id,
+      cutoff: DEFAULT_MODEL.cutoff,
+      capabilities: DEFAULT_MODEL.capabilities,
+    };
+  }
+
+  return {
+    name: has1m ? `${entry.name} (1M context)` : entry.name,
+    id: has1m ? `${id}[1m]` : id,
+    cutoff: entry.cutoff,
+    capabilities: entry.capabilities,
+  };
+}
+
+export function modelHasCapability(model: ModelInfo, capability: ModelCapability): boolean {
+  return model.capabilities?.includes(capability) ?? false;
+}
+
+// Claude Code reads its model setting from these files, most specific first
+function readConfiguredModel(cwd: string): string | null {
+  const candidates = [
+    join(cwd, ".claude", "settings.local.json"),
+    join(cwd, ".claude", "settings.json"),
+    join(homedir(), ".claude", "settings.json"),
+  ];
+  for (const path of candidates) {
+    try {
+      const settings = JSON.parse(readFileSync(path, "utf8"));
+      if (typeof settings.model === "string" && settings.model !== "") return settings.model;
+    } catch {
+      // Missing or malformed settings file — try the next candidate
+    }
+  }
+  return null;
+}
+
+/**
+ * Resolves the model this session will run on, using the same precedence Claude Code
+ * itself applies. Runs before base resolution so `--base auto` can key off the result.
+ */
+export function resolveSessionModel(modelArg?: string | null): ModelInfo {
+  return resolveModel(modelArg ?? process.env.ANTHROPIC_MODEL ?? readConfiguredModel(process.cwd()));
+}
+
+export function detectEnv(model: ModelInfo): EnvInfo {
   const cwd = process.cwd();
   const isGit = exec("git rev-parse --is-inside-work-tree") === "true";
 
@@ -40,14 +133,9 @@ export function detectEnv(): EnvInfo {
   const shell = basename(process.env.SHELL || "bash");
   const osVersion = exec("uname -sr") ?? "unknown";
 
-  return { cwd, isGit, isWorktree, gitBranch, gitStatus, gitLog, platform, shell, osVersion };
+  return { cwd, isGit, isWorktree, gitBranch, gitStatus, gitLog, platform, shell, osVersion, model };
 }
 
-// Hardcoded model info — update when Claude Code updates
-const MODEL_NAME = "Fable 5";
-const MODEL_ID = "claude-fable-5";
-const KNOWLEDGE_CUTOFF = "January 2026";
-
 export function buildTemplateVars(env: EnvInfo): TemplateVars {
   let gitStatusBlock = "";
   if (env.isGit) {
@@ -73,9 +161,9 @@ export function buildTemplateVars(env: EnvInfo): TemplateVars {
     PLATFORM: env.platform,
     SHELL: env.shell,
     OS_VERSION: env.osVersion,
-    MODEL_NAME,
-    MODEL_ID,
-    KNOWLEDGE_CUTOFF,
+    MODEL_NAME: env.model.name,
+    MODEL_ID: env.model.id,
+    KNOWLEDGE_CUTOFF: env.model.cutoff,
     GIT_STATUS: gitStatusBlock,
     WORKTREE_NOTICE: worktreeNotice,
   };

package/src/inspect.ts

@@ -4,7 +4,7 @@ import { parseCliArgs } from "./args.js";
 import { loadConfig, resolveConfigPath, type LoadedConfig } from "./config.js";
 import { resolveConfig } from "./resolve.js";
 import { getFragmentOrder } from "./assemble.js";
-import { detectEnv, buildTemplateVars } from "./env.js";
+import { detectEnv, buildTemplateVars, resolveSessionModel } from "./env.js";
 import { EMBEDDED_PROMPTS } from "./embedded-prompts.js";
 import type { TemplateVars } from "./types.js";
 
@@ -299,9 +299,10 @@ export function runInspectCommand(argv: string[], promptsDir: string): void {
 
   const parsed = parseCliArgs(filteredArgv);
   const loadedConfig = loadConfig();
-  const config = resolveConfig(parsed, loadedConfig);
+  const model = resolveSessionModel(parsed.modelHint);
+  const config = resolveConfig(parsed, loadedConfig, model);
 
-  const env = detectEnv();
+  const env = detectEnv(model);
   const templateVars = buildTemplateVars(env);
 
   const fragmentPaths = getFragmentOrder(config, promptsDir);

package/src/resolve.ts

@@ -1,13 +1,15 @@
-import type { ModeConfig } from "./types.js";
+import type { ModeConfig, ModelInfo } from "./types.js";
 import type { ParsedArgs } from "./args.js";
 import type { LoadedConfig } from "./config.js";
 import { resolveConfigPath } from "./config.js";
+import { modelHasCapability } from "./env.js";
 import { getPreset, isPresetName } from "./presets.js";
 import {
   AGENCY_VALUES,
   QUALITY_VALUES,
   SCOPE_VALUES,
   STYLE_VALUES,
+  BASE_AUTO,
   BUILTIN_MODIFIER_NAMES,
   BUILTIN_BASE_NAMES,
   PRESET_NAMES,
@@ -115,6 +117,8 @@ function applyModifiers(
   resolvedPaths: string[],
   position: "append" | "prepend",
 ): void {
+  // Prepending walks an insertion point forward so the batch keeps its own order
+  let insertAt = 0;
   for (const raw of modifiers) {
     const resolved = resolveModifier(raw, loadedConfig);
     let path: string;
@@ -124,7 +128,7 @@ function applyModifiers(
       path = resolved.path;
     }
     if (!resolvedPaths.includes(path)) {
-      if (position === "prepend") resolvedPaths.unshift(path);
+      if (position === "prepend") resolvedPaths.splice(insertAt++, 0, path);
       else resolvedPaths.push(path);
     }
   }
@@ -175,32 +179,47 @@ function resolveStyle(
   return resolveStyleValue(value, loadedConfig);
 }
 
+/** Picks the base Claude Code itself would assemble for this model. */
+function selectBaseForModel(model: ModelInfo): string {
+  return modelHasCapability(model, "lean-prompt") ? "lean" : "standard";
+}
+
+interface ResolvedBase {
+  base: string;
+  /** True when the base came from model detection rather than an explicit choice */
+  fromModel: boolean;
+}
+
 /**
  * Resolves a base reference to a built-in name or absolute directory path.
- * Priority: CLI --base > config defaultBase > preset base > "standard"
+ * Priority: CLI --base > config defaultBase > preset base > "auto"
  */
 function resolveBase(
   raw: string | undefined,
   loadedConfig: LoadedConfig | null,
   presetBase: string | undefined,
-): string {
+  model: ModelInfo,
+): ResolvedBase {
   const config = loadedConfig?.config ?? null;
 
-  // Priority: CLI --base > config defaultBase > preset base > "standard"
-  const value = raw ?? config?.defaultBase ?? presetBase ?? "standard";
+  // Priority: CLI --base > config defaultBase > preset base > "auto"
+  const value = raw ?? config?.defaultBase ?? presetBase ?? BASE_AUTO;
+
+  // 0. Model-driven selection — mirrors which assembly Claude Code would send this model
+  if (value === BASE_AUTO) return { base: selectBaseForModel(model), fromModel: true };
 
   // 1. Built-in name
-  if (isBuiltinBase(value)) return value;
+  if (isBuiltinBase(value)) return { base: value, fromModel: false };
 
   // 2. Config-defined name
   const configBases = config?.bases;
   if (configBases && value in configBases) {
-    return resolveConfigPath(loadedConfig!.configDir, configBases[value]);
+    return { base: resolveConfigPath(loadedConfig!.configDir, configBases[value]), fromModel: false };
   }
 
   // 3. Directory path
   if (looksLikeFilePath(value)) {
-    return isAbsolute(value) ? value : pathResolve(value);
+    return { base: isAbsolute(value) ? value : pathResolve(value), fromModel: false };
   }
 
   // 4. Unknown
@@ -209,14 +228,18 @@ function resolveBase(
     : " No config file found.";
   throw new Error(
     `Unknown --base value: "${value}". ` +
-    `Must be one of: ${BUILTIN_BASE_NAMES.join(", ")}, ` +
+    `Must be one of: ${BASE_AUTO}, ${BUILTIN_BASE_NAMES.join(", ")}, ` +
     `a name defined in your config, or a directory path.${configHint}`
   );
 }
 
+// Sections upstream ships alongside the lean assembly for models carrying prompt-bundle
+const PROMPT_BUNDLE_MODIFIERS = ["delivering-work", "corrections", "tool-restraint"];
+
 export function resolveConfig(
   parsed: ParsedArgs,
   loadedConfig: LoadedConfig | null,
+  model: ModelInfo,
 ): ModeConfig {
   const config = loadedConfig?.config ?? null;
   const modifierPaths: string[] = [];
@@ -240,9 +263,8 @@ export function resolveConfig(
   // Handle "none" preset — resolve base before early return.
   // Explicit --style (or config defaultStyle) still applies, like modifiers.
   if (parsed.preset === "none") {
-    const base = resolveBase(parsed.base, loadedConfig, undefined);
     return {
-      base,
+      base: resolveBase(parsed.base, loadedConfig, undefined, model).base,
       axes: null,
       style: resolveStyle(parsed.style, loadedConfig, undefined),
       modifiers: modifierPaths,
@@ -335,9 +357,15 @@ export function resolveConfig(
       : DEFAULT_SCOPE;
   }
 
-  const base = resolveBase(parsed.base, loadedConfig, presetBase);
+  const { base, fromModel } = resolveBase(parsed.base, loadedConfig, presetBase, model);
   const style = resolveStyle(parsed.style, loadedConfig, presetStyle);
 
+  // When the base was chosen for us, also mirror the extra sections upstream would
+  // send this model. An explicit --base means the user picked the shape themselves.
+  if (fromModel && modelHasCapability(model, "prompt-bundle")) {
+    applyModifiers(PROMPT_BUNDLE_MODIFIERS, loadedConfig, modifierPaths, "prepend");
+  }
+
   return {
     base,
     axes: { agency, quality, scope },

package/src/types.ts

@@ -36,19 +36,23 @@ export function isPresetName(value: string): value is PresetName {
 }
 
 // Built-in modifier names — used for collision checking in config validation
-export const BUILTIN_MODIFIER_NAMES = ["readonly", "context-pacing", "debug", "methodical", "director", "bold", "speak-plain", "tdd", "muse", "flow", "playful"] as const;
+export const BUILTIN_MODIFIER_NAMES = ["readonly", "context-pacing", "debug", "methodical", "director", "bold", "speak-plain", "tdd", "muse", "flow", "playful", "delivering-work", "corrections", "tool-restraint"] as const;
 export type BuiltinModifier = (typeof BUILTIN_MODIFIER_NAMES)[number];
 export function isBuiltinModifier(value: string): value is BuiltinModifier {
   return (BUILTIN_MODIFIER_NAMES as readonly string[]).includes(value);
 }
 
-// Built-in base names — "standard" (upstream-derived), "chill" (calm), "flow" (calm + engaged)
-export const BUILTIN_BASE_NAMES = ["standard", "chill", "flow"] as const;
+// Built-in base names — "standard" (upstream-derived), "chill" (calm), "flow" (calm + engaged),
+// "lean" (upstream's lean assembly, sent to models carrying the lean-prompt capability)
+export const BUILTIN_BASE_NAMES = ["standard", "chill", "flow", "lean"] as const;
 export type BuiltinBaseName = (typeof BUILTIN_BASE_NAMES)[number];
 export function isBuiltinBase(value: string): value is BuiltinBaseName {
   return (BUILTIN_BASE_NAMES as readonly string[]).includes(value);
 }
 
+// Not a base directory — a selector that picks a base from the session's model
+export const BASE_AUTO = "auto";
+
 // Reserved manifest entries — "axes" and "modifiers" trigger insertion
 export const MANIFEST_RESERVED = ["axes", "modifiers"] as const;
 export type ManifestReserved = (typeof MANIFEST_RESERVED)[number];
@@ -80,6 +84,23 @@ export interface ModeConfig {
   modifiers: string[]; // ordered list of modifier fragment paths (embedded keys or absolute paths)
 }
 
+/** Resolved model metadata for env.md substitution */
+/**
+ * Prompt-shaping capabilities Claude Code reads off the session model.
+ * - "lean-prompt": upstream sends the lean assembly instead of the standard one
+ * - "prompt-bundle": upstream additionally sends the delivering-work, corrections,
+ *   and tool-restraint sections (the `opus_5_prompt_bundle` capability)
+ */
+export const MODEL_CAPABILITIES = ["lean-prompt", "prompt-bundle"] as const;
+export type ModelCapability = (typeof MODEL_CAPABILITIES)[number];
+
+export interface ModelInfo {
+  name: string;
+  id: string;
+  cutoff: string;
+  capabilities?: readonly ModelCapability[];
+}
+
 export interface EnvInfo {
   cwd: string;
   isGit: boolean;
@@ -90,6 +111,7 @@ export interface EnvInfo {
   platform: string;
   shell: string;
   osVersion: string;
+  model: ModelInfo;
 }
 
 /** Template variables for env.md substitution */

package/src/usage.ts

@@ -33,7 +33,10 @@ Presets:
   spark           autonomous / architect / unrestricted (chill base, maximalist creative + wit)
 
 Base:
-  --base <name|path>      Built-in: standard, chill, flow
+  --base <name|path>      Built-in: auto (default), standard, chill, flow, lean
+  "auto" picks the base Claude Code itself would use for the session model:
+  lean for Opus 5 / Opus 4.8 / Fable 5, standard otherwise. On Opus 5 it also
+  adds the delivering-work, corrections, and tool-restraint modifiers.
   Base can also be a config-defined name or a directory path.
 
 Axis overrides:
@@ -56,6 +59,8 @@ Modifiers:
 Forwarded to claude:
   --append-system-prompt <text>
   --append-system-prompt-file <path>
+  --model <alias|id>      Also sets the model info in the prompt's environment section.
+  Without --model, model info comes from ANTHROPIC_MODEL or Claude settings files.
 
 Config: .claude-mode.json (project) or ~/.config/claude-mode/config.json (global)