Feature/add skill recovery

[sah1234567]

Summary

Add skill invocation tracking with deduplication and automatic post-compaction re-injection — the model no longer loses loaded skill content when context is compacted.

What changed

• Session tracking (session.rs): new InvokedSkillRecord struct and invoked_skills: HashMap<String, InvokedSkillRecord> field on Session. Repeated load_skill calls for the same skill name overwrite the previous entry — the HashMap key is the dedup mechanism.

• Turn loop recording (turn_loop.rs): when a load_skill tool call succeeds, the full raw body is recorded via Session::record_skill_invocation() so it survives compaction.

• Compaction injection (compaction.rs): inject_invoked_skills() builds a single synthetic user message from every recorded skill, placed at the front of the post-compact message list. The format follows Claude Code's layout: a leading sentence (The following skills were invoked…), ### Skill: {name} markdown headers, --- separators, most-recent-first ordering, and no artificial token budget.

• Cycle advance cleanup (engine.rs): invoked_skills is cleared alongside compaction_summary_prompt on cycle boundaries.

• All 4 compact_messages_safe call sites (engine.rs, turn_loop.rs, capacity_flow.rs) pass invoked_skills through.

Files changed

File	Change
crates/tui/src/core/session.rs	+InvokedSkillRecord, +invoked_skills field, +record_skill_invocation() method
crates/tui/src/core/engine/turn_loop.rs	Record on load_skill success
crates/tui/src/compaction.rs	+inject_invoked_skills(), new param on compact_messages / compact_messages_safe
crates/tui/src/core/engine.rs	Pass invoked_skills to compaction, clear on cycle advance
crates/tui/src/core/engine/capacity_flow.rs	Pass invoked_skills to compaction
crates/tui/src/core/engine/tests.rs	+7 engine-level tests
crates/tui/tests/skill_compaction_recovery.rs	+7 integration tests
crates/tui/tests/test_skills/	5 minimal skill fixtures for integration tests

Testing

• cargo test -p deepseek-tui --bin deepseek-tui (2170 passed; 12 pre-existing repl::runtime failures unrelated to this PR)
• cargo test -p deepseek-tui --test skill_compaction_recovery (7/7 passed)
• Verified cargo build produces zero new warnings
• Manual reasoning: traced end-to-end flow for dedup, compaction ordering, summary non-interference, and cycle clearing

Checklist

• Updated docs or comments as needed
• Added or updated tests where relevant
• Verified TUI behavior manually if UI changes (N/A — no UI surface, pure engine logic)

[gemini-code-assist]

Code Review

This pull request implements a system to preserve invoked skill content during message compaction by re-injecting it as a synthetic user message. It introduces the InvokedSkillRecord struct and a HashMap in the Session to track and deduplicate skill invocations, ensuring the latest content is maintained. The engine's turn loop now records successful load_skill tool calls, and the compaction logic integrates these records into the post-compaction message stream. Comprehensive unit and integration tests have been added to verify the deduplication, ordering, and recovery of skill content. I have no feedback to provide.

[Hmbown]

Thanks for working on this. This PR is too large or scope-mixed for review in this repo's current maintainer model. Please split it into one PR per accepted issue, each under 500 changed lines and with a concrete test plan. Changes to core compaction/session behavior need maintainer approval before implementation.

This PR also needs the admission details from the contributor guide before review: a linked accepted issue, AI-use disclosure, a human-owner statement that you have read and can explain every changed file, and a clean test plan with commands that pass. I am closing this PR, but a smaller PR against an accepted issue is welcome.