fix: batch config wiring quick wins (#262-#266)

[laynepenney]

Summary

Five quick fixes to close gaps left after PR #261's config wiring:

• with_provider/with_provider_and_path constructors bypass config wiring #262 — Deprecate with_provider()/with_provider_and_path() constructors that bypass config. Added #[deprecated] attributes pointing to the correct with_project_path() + set_config() + set_provider() pattern.
• on_compaction callback is None — no visual feedback during context compaction #264 — Wire on_compaction callback so the TUI displays "Compacting context..." / "Context compacted" status messages. Added Compaction(bool) variant to AppEvent.
• Non-interactive mode (-P) bypasses systemPromptAdditions and projectContext #266 — Extract build_system_prompt_from_config() as a standalone public function and use it in -P non-interactive mode, so config-driven systemPromptAdditions and projectContext are applied consistently.
• dangerousPatterns config field is not wired #263 — Wire dangerousPatterns config through to tool confirmation. Added dangerous_patterns field to AgentConfig, matches_dangerous_pattern() method, and unified maybe_confirm() that checks both built-in destructive tools and custom patterns in a single pass.
• estimate_tokens re-serializes JSON tool inputs on every iteration #265 — Cache running_char_count on AgentState to avoid re-serializing all message JSON on every iteration of the agent loop. Incremented on each message push, recalculated after compaction.

Files changed (6)

File	Fixes
codi-rs/src/tui/app.rs	#262, #264, #266, #263
codi-rs/src/tui/mod.rs	#266
codi-rs/src/main.rs	#266, #263
codi-rs/src/agent/types.rs	#263, #265
codi-rs/src/agent/mod.rs	#263, #265
codi-rs/src/orchestrate/child_agent.rs	#263

Test plan

• cargo build — clean, zero warnings
• cargo test — 463 tests pass (7 new)
• New tests: test_dangerous_pattern_match, test_dangerous_pattern_empty, test_dangerous_pattern_invalid_regex_skipped, test_agent_state_default_running_char_count, test_build_system_prompt_from_config_none, test_build_system_prompt_from_config_with_additions, test_app_event_compaction_variant

Review notes

• maybe_confirm() replaces the old should_confirm() + confirm_tool() two-method pattern to avoid double-serializing tool input JSON
• Invalid regex patterns in dangerousPatterns config emit a tracing::warn! rather than silently skipping
• messages_mut() is a known escape hatch that bypasses running_char_count — no current callers, but noted for future awareness

🤖 Generated with Claude Code

[laynepenney]

Self-Review

Build & Tests

• ✅ cargo build — clean, zero warnings
• ✅ cargo test — 463 tests pass (7 new), 0 failures
• ✅ CI: Type Check pass, Test (macos) pass, Test (ubuntu) pass

Fix-by-fix verification

#262 — Deprecate bypass constructors

• ✅ #[deprecated] attributes on both with_provider() and with_provider_and_path()
• ✅ Functions remain functional for backwards compat
• ✅ Deprecation message points to correct pattern

#264 — Wire on_compaction callback

• ✅ Compaction(bool) variant added to AppEvent
• ✅ Callback wired in set_provider() with Arc + tx.clone() pattern matching other callbacks
• ✅ Match arm in process_app_events() sets status messages
• ✅ New test test_app_event_compaction_variant

#266 — Fix -P mode system prompt

• ✅ build_system_prompt_from_config() extracted as standalone public function
• ✅ App::build_system_prompt() delegates to it
• ✅ Re-exported from tui/mod.rs
• ✅ main.rs handle_prompt() uses build_system_prompt_from_config(Some(config)) instead of hardcoded string
• ✅ Existing tests still pass, 2 new tests for standalone function

#263 — Wire dangerousPatterns

• ✅ dangerous_patterns: Vec<String> on AgentConfig with Vec::new() default
• ✅ matches_dangerous_pattern() method with tracing::warn! for invalid regex
• ✅ Unified maybe_confirm() replaces old should_confirm()+confirm_tool() two-method pattern — serializes input once
• ✅ Wired in build_agent_config() (app.rs), handle_prompt() (main.rs), child_agent.rs
• ✅ 3 new tests: pattern match, empty patterns, invalid regex graceful skip

#265 — Cache running char count

• ✅ running_char_count: usize on AgentState with 0 default
• ✅ estimate_tokens() uses cached count instead of iterating all messages
• ✅ Incremented at all 3 message push sites (user msg, assistant msg, tool results)
• ✅ Recalculated from remaining messages after compact_context() drain
• ✅ Reset to 0 via AgentState::default() in clear()
• ✅ 1 new test for default state

Review findings addressed during PR

• Fixed double serialization: merged should_confirm()+confirm_tool() into single maybe_confirm() method
• Added tracing::warn! for invalid regex patterns instead of silent skip
• Noted messages_mut() escape hatch as future-awareness item (no current callers)

No issues found. Ready to merge.