Models no longer silently over-claim file edits. Two layers, one PR:

Layer 1 — salvaged from #15673 (@briandevans): Advertises per-mode required params in the patch schema's descriptions so strict tool callers (kimi-k2.x in #15524, grok-4.1-fast reported separately) stop omitting old_string/new_string/patch. JSON Schema's required array stays ["mode"] because anyOf/oneOf at parameters level breaks Anthropic, Fireworks, and Kimi — description-level REQUIRED when mode='…' hints are the only cross-provider-compatible mechanism.

Layer 2 (new): Per-turn verifier that catches over-claiming after the fact. Every write_file / patch call's outcome is tracked per resolved path; failures that are never superseded by a later successful write to the same path are surfaced in a footer appended to the assistant's final response:

⚠️ File-mutation verifier: 3 file(s) were NOT modified this turn despite any wording above that may suggest otherwise. Run `git status` or `read_file` to confirm.
  • concepts/automatic-organization.md — [patch] Could not find match for old_string
  • concepts/lora.md — [patch] Could not find match for old_string
  • concepts/rag-pipeline.md — [patch] Could not find match for old_string

Regression target: Ben Eng's llm-wiki session where grok-4.1-fast batched parallel patches, half failed with 'Could not find old_string', and the model summarised the turn claiming every file was edited — forcing the user to manually run git status every turn.

Changes

• tools/file_tools.py — PATCH_SCHEMA description + per-property REQUIRED when mode='…' hints (cherry-pick from fix(patch-tool): advertise per-mode required params in schema descriptions #15673)
• run_agent.py — _FILE_MUTATING_TOOLS const, _extract_file_mutation_targets (handles write_file, patch-replace, and patch-v4a single + multi-file), _extract_error_preview, AIAgent._record_file_mutation_result, AIAgent._file_mutation_verifier_enabled, AIAgent._format_file_mutation_failure_footer. Recording wired into both _execute_tool_calls_concurrent and _execute_tool_calls_sequential so parallel and sequential batches are both covered. Footer emission runs at turn end, before transform_llm_output / post_llm_call plugin hooks, so plugins still see (and can modify) the augmented text.
• hermes_cli/config.py — display.file_mutation_verifier default True
• tests/run_agent/test_file_mutation_verifier.py — 31 unit tests (target extraction, error-preview extraction, state transitions, footer rendering, env/config precedence)
• tests/tools/test_file_tools.py — 9 TestPatchSchemaShape assertions from fix(patch-tool): advertise per-mode required params in schema descriptions #15673
• website/docs/user-guide/configuration.md + website/docs/reference/environment-variables.md — display.file_mutation_verifier + HERMES_FILE_MUTATION_VERIFIER

Validation

Notes

• Zero cost: no LLM calls, no tool schema changes (schema layer is just descriptions).
• Prompt-cache safe: no messages injected into history, footer is appended to the returned string only.
• Message-alternation safe: no synthetic user/tool messages introduced.
• Opt-out: set display.file_mutation_verifier: false in config.yaml or HERMES_FILE_MUTATION_VERIFIER=0.
• V4A multi-file patches: parsed for *** Update File: / *** Add File: / *** Delete File: headers so each file in a multi-file patch is tracked separately.

Closes #15524
Cherry-picks and credits @briandevans (PR #15673).