refactor: extract cli_config, cli_git, cli_display from cli.py (−2,158 lines)

[xandgate]

Extracts three coherent clusters from the 16,280-line cli.py into focused modules, per the god-file modularization goal in AGENTS.md:

• cli_config.py (609 lines): config loading/saving (load_cli_config, save_config_value), prefill message loading and precedence resolution, reasoning/service-tier parsing
• cli_git.py (701 lines): git worktree isolation — synchronized-base _setup_worktree, lock classification, stale-worktree pruning, orphaned-branch cleanup; active-worktree state behind get_active_worktree()/set_active_worktree()
• cli_display.py (1,015 lines): ANSI/skin helpers, light-mode detection/remap, output-history record/replay, file-drop and attachment detection, ChatConsole, banners, slash-command detection

cli.py: 16,280 → 14,122 lines (−13.3%). The HermesCLI class and main() are untouched.

All moves are verbatim from the current head. cli.py re-imports every moved name, so existing from cli import X call sites (mixins, tools/delegate_tool.py, tests) keep working. Tests that monkeypatched moved module globals now patch the owning module; the full tests/cli + tests/hermes_cli suites show zero new failures vs a clean main baseline.

History note: originally opened in April against a 10,536-line cli.py; re-extracted from the current head in July per the hermes-sweeper review below (preserving the sync_base worktree setup, lock-classification/pruning safety logic, prefill precedence, and light-mode paths that landed in between).

🤖 Generated with Claude Code

[mxnstrexgl]

🤖 Automated Refactor Review

Summary

Clean extraction of 1,327 lines from cli.py into 3 focused modules (cli_config.py, cli_git.py, cli_display.py). Good modularization effort.

✅ Imports & Dependencies

• All imports properly preserved
• No circular import risks detected
• Module boundaries are clean

✅ API Compatibility

• Public APIs maintained (, , worktree helpers)
• No breaking changes to callers

✅ State Management

• Shared state properly handled via with get/set accessors
• Clean separation of concerns

⚠️ Issue Found: Missing Variable Reference

cli_config.py line ~98: references which is defined in cli.py, not in the new module. This will cause when called.

Fix: Either:

1. Pass as parameter to the function, OR
2. Import/get it via inside the function

✅ Security Review

• Config handling: No secrets exposed in new structure - API keys still handled via env vars
• Git operations: Credential handling preserved correctly - worktree operations use subprocess, no credential leaks
• Display module: No info disclosure - only UI helpers and ANSI formatting

Verdict

NEEDS CHANGES - The reference needs fixing before merge. Otherwise excellent refactor.

[mxnstrexgl]

Automated Refactor Review - PR #12355

Summary

Clean extraction of 1,327 lines from cli.py into 3 focused modules. Good modularization.

Critical Issue

cli_config.py: _load_prefill_messages() uses _hermes_home variable that is NOT defined in the module. This variable was in cli.py but wasn't passed to the new module. Will cause NameError.

Security Review: PASS

• Config handling: No secrets exposed
• Git operations: Credential handling preserved
• Display module: No info disclosure

Verdict: NEEDS CHANGES

Fix the _hermes_home reference before merge. Suggest using get_hermes_home() import inside the function.

[xandgate]

Fixed in 477b7f2.

_load_prefill_messages resolved relative paths against _hermes_home (a module-level var in cli.py, undefined in the extracted module). Replaced with get_hermes_home() — consistent with how load_cli_config and save_config_value already handle it.

# before
path = _hermes_home / path   # NameError in extracted module

# after  
path = get_hermes_home() / path   # consistent with rest of cli_config.py

Good catch — I had patched the two obvious _hermes_home usages in load_cli_config and save_config_value but missed this third one in _load_prefill_messages. All three are now get_hermes_home() calls. Syntax verified clean.

[xandgate]

Rebased on current main (9d3127) and conflicts resolved.

What conflicted: Two upstream commits landed while the PR was open:

• fix(cli): strip all reasoning tag variants from /resume recap — improved _strip_reasoning_tags with re.IGNORECASE, orphan close-tag stripping, and added "thought" to _REASONING_TAGS for Gemma 4
• fix(agent): strip <think> blocks from stored assistant content — related content processing fix

How resolved:

• cli.py: kept our extracted import structure (discarded the re-inlined function definitions that came back via upstream)
• cli_config.py: absorbed all upstream improvements — updated _REASONING_TAGS to match ("thought" added, "THINKING" removed since re.IGNORECASE handles it), and replaced _strip_reasoning_tags with the improved version including re.IGNORECASE flags and orphan close-tag stripping

All four files syntax-verified clean post-rebase.

[teknium1]

Thanks for the focused modularization work. This aligns with the repository’s stated goal of extracting coherent clusters from cli.py (AGENTS.md:65-70), but the patch needs a current-head salvage rather than a mechanical conflict resolution.

Problems

• The proposed cli_git.py has _setup_worktree(repo_root) only, while current main calls _setup_worktree(sync_base=_sync_base) (cli.py:15854) and implements synchronized-base setup at cli.py:1423-1588.
• Current worktree safety logic must be retained: lock classification is at cli.py:1643-1702, and pruning uses it at cli.py:1846-1949 (added by 74809b4e9).
• Preserve later behavior in the extracted config/display clusters: prefill precedence is resolved at cli.py:318-334 and used at cli.py:3914; the compact banner fast-start path is at cli.py:3496-3502.

Suggested changes

• Re-extract the current-head implementations and retain the existing cli.py compatibility exports.
• Cover the relocated surface with the existing worktree and prefill tests.

This is an automated hermes-sweeper review.

[xandgate]

Re-extracted from the current head per the review above (branch force-pushed; the previous April snapshot is superseded).

Addressing each point:

• _setup_worktree signature — cli_git.py now carries the current implementation: _setup_worktree(repo_root=None, sync_base=True) including the synchronized-base setup, plus _resolve_worktree_base and the _worktree_has_unpushed_commits / _worktree_is_dirty helpers. The sync_base=_sync_base call site in cli.py is unchanged.
• Worktree safety logic — lock classification (_worktree_lock_is_live) and the pruning that consumes it (from 74809b4) moved intact into cli_git.py.
• Later config/display behavior — _resolve_prefill_messages_file (prefill precedence) is in cli_config.py; the compact-banner fast-start path and the light-mode detection/remap block are in cli_display.py.
• Compatibility exports — every moved name is re-imported in cli.py, so from cli import X call sites (mixins, tools/delegate_tool.py, tests) keep working. Active-worktree session state lives in cli_git behind get_active_worktree() / set_active_worktree().
• Test coverage — the existing worktree suites (test_worktree.py, test_worktree_security.py, test_worktree_sync_base.py) and test_prefill_config.py pass against the relocated surface. Tests that monkeypatched moved module globals (cli._hermes_home, _LIGHT_MODE_CACHE, _pt_print, _OUTPUT_HISTORY) now patch the owning module. The full tests/cli + tests/hermes_cli run has failure parity with a clean main baseline (identical pre-existing flakes, zero new failures).

All moves are verbatim from today's cli.py; the only non-move changes are the re-import shims, the two accessor functions, and dropping imports that became dead in cli.py. Conflicts are resolved — the PR is mergeable again.

[andrexibiza]

Coordination from the god-file decomposition campaign (epic #78647, cli.py target #55136):

This PR overlaps cli.py's R1 region's two biggest extraction candidates:

• C2 (reasoning/prefill, window 234–407) — this PR removes the whole window
• C5 (git worktree cluster, windows 1402–2115 + 2213–2497) — same

Our double-blind consensus (2 independent analysts + adjudication, 2026-08-05) found this PR during the open-PR sweep and blocked both candidates on it — no R1 slice will be filed while this is open. If this merges, R1's C2/C5 work is pre-done and we re-derive the remaining cluster map from its diff (per the consensus execution order §4.1).

Standing coordination: the campaign will re-scan the merge state before every R1 scheduling decision. If you'd like the full consensus file for reference: it's summarized at #78647 progress comments; the blocking map is in the campaign's R1-CONSENSUS.md.

No conflicts intended — the extraction wave is proceeding on R2–R5 (windows 5613+), which this PR does not touch.