✨ Add subagent frontmatter to agent action definitions

[monkut]

Closes via review: refs #80

Summary

Adopts a Claude Code subagent–style YAML frontmatter on each *_SYSTEM_PROMPT.md template so per-action tool surface, model, and reasoning effort are declarative instead of uniform.

• AgentConfig gains optional fields: tools, disallowed_tools, model, effort, max_thinking_tokens, max_turns
• Stdlib-only frontmatter parser (no new runtime deps); validates enum fields (model, effort) at load time with clear errors
• load_agent_config splits frontmatter from body; templates without frontmatter continue to work unchanged (back-compat)
• ClaudeRunner translates AgentConfig fields into --model, --allowedTools, --disallowedTools, --max-turns
• CLI precedence (highest wins): explicit CLI flag > env var > template frontmatter > built-in default
• Per-action defaults match the issue's table:

Action	tools	model	effort
prepare	Read, Grep, Glob, Bash(gh:*)	sonnet	medium
plan	Read, Grep, Glob, Bash(gh:*)	opus	high
develop	Read, Write, Edit, Bash, Grep, Glob	opus	max
issue-review	Read, Grep, Glob, Bash(gh:*)	sonnet	medium
pr-review	Read, Grep, Glob, Bash(gh:*,git:*)	opus	high
explore	Read, Grep, Glob, Bash(gh:*)	sonnet	high
diagnose	Read, Grep, Glob, Bash(gh:*,git:*)	sonnet	high
fix-ci	Read, Write, Edit, Bash, Grep, Glob	sonnet	high

• Bumps version to 0.2.6

Notes

• permission_mode is intentionally not exposed in AgentConfig or per-action defaults. askcc runs claude with --dangerously-skip-permissions so the one-shot subprocess can execute unattended; if any permission check were to trigger, the run would deadlock with no human to answer the prompt. The per-action tools allowlist is the safety boundary instead.
• A code comment near DEVELOP_AGENT_PROMPT and FIXCI_AGENT_PROMPT explains why those write-capable agents specifically need the broad Edit/Write/Bash allowlist.
• Frontmatter parser is ~30 lines of stdlib code; project's dependencies = [] stance preserved.

Key Flows

flowchart TD
    A[load_agent_config action] --> B[load_template *_SYSTEM_PROMPT.md]
    B --> C{starts with '---'?}
    C -- no --> D[return AgentConfig with body as system_prompt]
    C -- yes --> E[parse_frontmatter]
    E --> F{valid model/effort?}
    F -- no --> G[raise ValueError with allowed values]
    F -- yes --> H[override AgentConfig fields]
    H --> I[return AgentConfig with body and frontmatter overrides]

Loading

flowchart TD
    A[CLI arg --effort] --> B{explicit?}
    B -- yes --> Z[use CLI value]
    B -- no --> C{env ASKCC_CLAUDE_EFFORT_LEVEL set and valid?}
    C -- yes --> Y[use env value]
    C -- no --> D{frontmatter effort set?}
    D -- yes --> X[use frontmatter value]
    D -- no --> W[use DEFAULT_EFFORT_LEVEL]

Loading

Verification

• `uv run poe check` (ruff) — passed
• `uv run poe typecheck` (pyright) — 0 errors
• `uv run poe test` — 178 passed (155 existing + 23 new across TestParseFrontmatter, TestLoadAgentConfigFrontmatter, TestEffortPrecedence, TestRunnerFrontmatterFlags)

Test plan

• Frontmatter parser: happy path, list/int coercion, invalid model, invalid effort, invalid int, unclosed delimiter, unknown key warning, blank/comment lines
• load_agent_config: default templates yield frontmatter fields, no-frontmatter back-compat, user-frontmatter override, invalid frontmatter raises at load
• CLI precedence: CLI > env > frontmatter > default
• Runner emits --model, --allowedTools, --disallowedTools, --max-turns from AgentConfig; no flags emitted when fields unset
• Existing tests for effort/thinking flow still pass