Add required skills and progress output polish

Author: peterschmidt85

README.md

@@ -21,6 +21,7 @@ Program frontmatter:
 
 - `id` (optional)
 - `model` (optional)
+- `skills` (optional list of required provider-native skills)
 
 ## How `apply` Works
 
@@ -98,6 +99,16 @@ The confirmation preview includes a variables summary, for example: `variables:
 
 Program variables are defined in frontmatter under `variables` (`NAME: {}` for required, `NAME: { default: "..." }` for optional defaults) and referenced as `${{ var.NAME }}`.
 
+Required provider-native skills can be declared in frontmatter with `skills`, for example:
+
+```yaml
+---
+skills: [dstack]
+---
+```
+
+If a required skill is unavailable, the run fails early. Implementation details and provider-specific behavior are documented in `contrib/ARCHITECTURE.md` and `contrib/DEVELOPMENT.md`.
+
 Example output (will vary by session/model):
 
 ```text
@@ -106,7 +117,7 @@ Last session: 019d5843-eb2d-70b1-b49a-343033117944 (success, 43m ago)
   program: examples/smoke.md unchanged
   changes: 0 files
 Proceed? [y/N] y
-🧵 019d586b-aa65-78b2-8a0d-27b5543c59bb | workspace
+🧵 019d586b-aa65-78b2-8a0d-27b5543c59bb | workspace | claude:sonnet
 ✔ cat examples/smoke.md | 1ms | out
 💬 Verified `example-data/output-smoke.txt:1` already contains the required `SMOKE_OK` line with trailing newline. | msg
 turn 1 | tokens: in=117k out=1.6k cached=107k

contrib/ARCHITECTURE.md

@@ -9,7 +9,7 @@ Clawform runs agent work from markdown files instead of chat windows.
 
 A **program** is one markdown file (`*.md`) representing one task.
 
-- frontmatter is tool-owned and strict (`id`, `model`, `variables`)
+- frontmatter is tool-owned and strict (`id`, `model`, `skills`, `variables`)
 - markdown body is agent-facing and free-form
 
 ## 2) Implemented v0 Scope
@@ -63,6 +63,7 @@ Frontmatter (strict):
 
 - `id` (optional)
 - `model` (optional override)
+- `skills` (optional list of required provider-native skill names)
 - `variables` (optional map)
   - key: variable name (`[A-Za-z_][A-Za-z0-9_]*`)
   - value:
@@ -83,6 +84,16 @@ Variable rules:
 3. apply-time `--var NAME=VALUE` overrides frontmatter default
 4. variables without default are required at apply time
 
+Skill rules:
+
+1. `skills` is a shared frontmatter list, for example `skills: [dstack]`
+2. each listed skill name must be a single non-empty token without whitespace
+3. Clawform prepends explicit provider-native skill validation/invocation lines to the real session prompt before the normal apply contract
+4. skill command syntax is provider-specific (`$skill` for Codex, `/skill` for Claude)
+5. the injected validation line is phrased as `Fail if skill is not found.`
+6. the injected prelude also instructs the agent to write `.clawform/agent_result.json` with `status: failure` and `reason: program_blocked` before stopping
+7. if the provider still surfaces a missing-skill response without writing the result artifact, Clawform uses provider-specific fallback detection and fails the apply run
+
 ## 4) Apply Session Flow (Current Behavior)
 
 1. Load program + config and resolve provider + model (`-p/--provider` overrides default provider selection).
@@ -94,7 +105,7 @@ Variable rules:
    - variable diff vs last session variable snapshot (if available)
 5. Ask for confirmation (interactive default; skipped by `--yes`).
 6. Clear prior run protocol files in `.clawform/` and write runtime variables file (`.clawform/agent_variables.json`) when variables are present.
-7. Build runtime prompt; in `workspace` and `auto` modes include explicit verdict-gate rules for sandbox-vs-program blocking.
+7. Build runtime prompt; if the program declares `skills`, prepend provider-native skill invocation commands first. In `workspace` and `auto` modes include explicit verdict-gate rules for sandbox-vs-program blocking.
 8. Run provider in the current workspace (no temp workspace copy).
 9. Normalize provider-native events into shared progress categories, stream them to terminal, and during the run write session `commands/*` and `messages/*`. In debug mode, also write session `events.ndjson`.
 10. In `auto` mode, allow at most one retry in `full-access` mode only when current-run `.clawform/agent_result.json` reports `status=partial|failure` and `reason=sandbox_blocked` (no stdout/stderr heuristic fallback).
@@ -106,7 +117,7 @@ Variable rules:
 ## 4.1 Progress Rendering Semantics
 
 - In an interactive TTY, rich progress keeps a spinner plus a live `running` or `running: <activity>` status line.
-- The run-start line includes the session id, execution mode, and a compact `provider:model` suffix, for example `🧵 <session> | workspace | codex:gpt-5-codex`.
+- The run-start line includes the session id, execution mode, a compact `provider:model` suffix, and when applicable a compact `skills:` suffix, for example `🧵 <session> | workspace | codex:gpt-5-codex | skills:dstack`.
 - `running` means the provider run is still alive. It is a liveness indicator, not the same thing as reasoning text.
 - The live activity suffix is derived from the highest-priority active provider item, for example `search ...`, `fetch ...`, `use ...`, `edit ...`, or `update plan`.
 - Completed progress lines are normalized across Claude and Codex into shared categories such as reasoning (`💭`), assistant text (`💬`), search (`🔎`), fetch (`🌐`), command (`❱`), file change (`✏️`), plan update (`update plan | ...`), generic tool (`🔧`), and unknown provider event (`📦`).

contrib/DEVELOPMENT.md

@@ -108,7 +108,7 @@ Interactive progress UI is enabled automatically only when stdin/stdout are atta
 Current progress semantics:
 
 - Rich mode keeps a spinner plus a live `running` or `running: <activity>` status line.
-- The run-start line includes the session id, execution mode, and a compact `provider:model` suffix such as `🧵 <session> | workspace | claude:sonnet`.
+- The run-start line includes the session id, execution mode, a compact `provider:model` suffix, and when applicable a compact `skills:` suffix such as `🧵 <session> | workspace | claude:sonnet | skills:dstack`.
 - `running` is a liveness indicator. It does not mean the model is explicitly emitting reasoning.
 - Plain mode prints stable progress lines without the interactive spinner/status renderer.
 - Completed provider items are normalized across Claude and Codex into categories such as `💭`, `💬`, `🔎`, `🌐`, `❱`, `✏️`, `update plan | ...`, `🔧`, and `📦`.
@@ -124,6 +124,16 @@ cargo run -p clawform -- apply -f examples/smoke.md --debug
 
 That debug run will also persist `.clawform/programs/<program_id>/sessions/<session_id>/events.ndjson` for postmortem event inspection.
 
+Program frontmatter also supports required provider-native skills:
+
+```yaml
+---
+skills: [dstack]
+---
+```
+
+When present, Clawform prepends explicit provider-native skill validation/invocation lines to the real session prompt before the normal apply contract, for example `$dstack Fail if skill is not found.` or `/dstack Fail if skill is not found.`. The injected prelude also tells the agent to stop and write `.clawform/agent_result.json` with `reason: program_blocked` if a required skill is unavailable.
+
 Disable progress rendering entirely:
 
 ```bash

crates/clawform-cli/src/lib.rs

@@ -4,6 +4,7 @@ use std::ffi::OsString;
 use std::fs;
 use std::io::{self, IsTerminal, Write};
 use std::path::{Path, PathBuf};
+use std::sync::atomic::{AtomicBool, Ordering};
 
 use anyhow::{anyhow, Context, Result};
 use clap::{ArgAction, Parser, Subcommand, ValueEnum};
@@ -20,6 +21,7 @@ const MAX_REPORTED_FILES_DISPLAY: usize = 20;
 const AGENT_RESULT_REL: &str = ".clawform/agent_result.json";
 const HELP_RENDER_WIDTH: usize = 100;
 const HELP_SPEC_WIDTH: usize = 30;
+static DETAILED_TERMINAL_VERDICT_PRINTED: AtomicBool = AtomicBool::new(false);
 const TOP_LEVEL_HELP_ROWS: &[(&str, &str)] = &[
     ("-h, --help", "print help"),
     ("-V, --version", "print version"),
@@ -205,13 +207,18 @@ enum Commands {
 }
 
 pub fn main_entry() {
+    DETAILED_TERMINAL_VERDICT_PRINTED.store(false, Ordering::Relaxed);
     if let Err(err) = real_main() {
         if is_user_cancelled_error(&err) {
-            print_canceled(true);
+            if !DETAILED_TERMINAL_VERDICT_PRINTED.load(Ordering::Relaxed) {
+                print_canceled(true);
+            }
             std::process::exit(130);
         }
         if is_blocked_error(&err) {
-            print_blocked();
+            if !DETAILED_TERMINAL_VERDICT_PRINTED.load(Ordering::Relaxed) {
+                print_blocked();
+            }
             std::process::exit(2);
         }
         eprintln!("error: {:#}", err);
@@ -875,6 +882,7 @@ fn maybe_print_agent_status_from_result_file(workspace_root: &Path) {
         line.push_str(&result_suffix);
     }
     eprintln!("{}", line);
+    DETAILED_TERMINAL_VERDICT_PRINTED.store(true, Ordering::Relaxed);
 }
 
 fn status_file_separator(use_color: bool) -> &'static str {