fix: detect, warn, and block file re-read/search loops after context compression

[0xbyt4]

Summary

Fixes the issue where the agent gets stuck in an infinite reading loop after context compression, re-reading the same files endlessly without writing or responding.

Root cause: Context compression summarizes conversation history but loses track of which files were already read. After compression, the model thinks it hasn't examined the files yet and reads them again. Combined with todo re-injection of completed items, this creates an infinite loop.

Fix (multi-layered):

1. Read tracking with escalation (tools/file_tools.py): Track file reads per task. 2nd read returns a soft warning with content. 3rd+ read blocks — returns error with no content, forcing the model to stop.

2. Search tracking (tools/file_tools.py): Same mechanism for search_files — identical searches are warned then blocked after 3 repeats.

3. File history injection (run_agent.py): After context compression, inject a structured message listing all files already read with "do NOT re-read" instruction.

4. Todo re-injection filtering (tools/todo_tool.py): format_for_injection() now filters out completed/cancelled todos. Only pending/in_progress items are re-injected after compression, preventing the model from re-doing finished work.

Design decisions:

• Escalating response: warn (2nd) → block (3rd+) — gives the model one chance before hard-stopping
• Thread-safe tracking with _read_tracker_lock
• Task-isolated — different tasks have separate trackers
• Pagination-aware — different offsets of the same file don't trigger false warnings
• Search tracking keyed on (pattern, target, path, file_glob) — different queries are independent

Test plan

• 26 unit tests: read warning/blocking, search warning/blocking, task isolation, different file/region/pattern, summary accuracy, tracker cleanup, compression history injection, todo filtering
• Manual bash verification of all 3 fixes with real function calls

[teknium1]

Merged! Thanks for the contribution @0xbyt4 — the read-loop detection and todo injection filtering are great additions.

I pushed a follow-up commit (a458b53) on top with several improvements:

1. Consecutive-only tracking — The counter now resets whenever any other tool is called in between (write, patch, terminal, etc.), so only truly back-to-back identical reads/searches trigger warnings. This prevents false blocks in legitimate read→edit→verify workflows.

2. Adjusted thresholds — Warn on 3rd consecutive (was 2nd), block on 4th+ (was 3rd+).

3. Fixed tuple unpacking bug — get_read_files_summary() was crashing on search keys (5-tuple) when trying to unpack as 3-tuple. Now uses a separate read_history set that only tracks file reads, so searches don't corrupt the summary.

4. Reverted web_extract docstring — The title field IS returned by web_tools.py, so restored it in the docs.

5. Tests updated — 35 tests covering the new consecutive-only behavior, notify_other_tool_call, interleaved operations, etc.