Codebase quality pass: security, performance, maintainability fixes

[thromel]

Summary

A tracked, end-to-end code-quality pass over the codebase, addressing review findings in priority order (security → performance → maintainability → practices). Every change is committed as an isolated, logical step and the full quality gate is green.

Quality gate: ruff check, ruff format --check, mypy src (292 files), and pytest (757 passed, 1 skipped) all pass.

Security (P0)

• Run project hooks via argv with shell=False — removes a shell-injection vector from repo-provided hook config; commands are parsed with shlex.split and missing executables / unparseable commands are handled explicitly (regression tests added).
• Stop persisting instruction/profile text in transcripts — only metadata (*_chars) is recorded; full text is rehydrated from disk on session resume, keeping secrets/proprietary content out of logs.
• Add timeouts to git subprocess invocations — clone/checkout/diff/query calls now have named timeouts and handle TimeoutExpired.

Performance (P1)

• Per-run file-text cache + code-graph memoization (mtime/size fingerprints).
• In-session transcript event cache with append-aware invalidation.
• Single-pass RuntimeTraceSnapshot shared by feedback/attempts, replacing repeated reverse scans.
• Avoid a git diff subprocess on every retry attempt by preferring the agent's final_diff.
• Cache successful model preflight per (model, endpoint) for the session.
• Replace O(n²) membership scans with set lookups.
• Hoist hot-path regexes, the JSON repair schema, and the route table to module level.

Maintainability (P1)

• Consolidate duplicated trace parsers and retrieval helpers.
• Deduplicate CLI logic: shared emit_summary/SummaryProtocol, shared open_url HTTP helper, and a single budget_limit_label.
• Promote leaky cross-module helpers: is_test_path/path_terms are now public in retrieval_features, and an identical _is_test_path duplicate in context_packing was removed.
• Split the runtime.attempts god module: extracted the independent sandbox-attempt execution into runtime/sandbox_attempt.py (clean separation of concerns), with attempts.py re-exporting for API stability. The retry-feedback logic is intentionally kept together as a single cohesive unit.

Readability & Practices (P2)

• Convert the session-resume if/elif dispatch chain into a handler registry.
• Guard /help against command-registry drift with a test.
• Always clean up the workspace restorer via try/finally.
• Centralize magic numbers as named constants (sandbox/git timeouts, retry thresholds, pids-limit).
• Structured error handling reviewed: re-raises already use exception chaining (raise ... from), and broad boundary handlers already record structured errors into trace/result objects — no redundant logging framework added.

Deferred (recommended as isolated follow-up PRs)

These are large, purely-organizational reorganizations with non-trivial regression risk; keeping them out of this PR preserves reviewability:

• Move deepagents_*.py into a patchsmith/deepagents/ subpackage with a facade.
• Consolidate root evaluation_* / public_issue_* module navigation.
• Extract large static prompt/HTML content into data files (high risk to prompt-dependent tests).

Testing

• uv run ruff check src tests — clean
• uv run ruff format --check src tests — clean
• uv run mypy src — clean (292 files)
• uv run pytest — 757 passed, 1 skipped