gitpr-cli 1.1.0__py3-none-any.whl → 1.2.0__py3-none-any.whl

{gitpr_cli-1.1.0.dist-info → gitpr_cli-1.2.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: gitpr-cli
-Version: 1.1.0
+Version: 1.2.0
 Summary: AI-powered PR automation, commit messages, and code review (Gemini, DeepSeek, and Ollama)
 Author-email: Natan Fiuza <contato@natanfiuza.dev.br>
 Requires-Python: >=3.10
@@ -148,11 +148,11 @@ You can pass the following *flags* for specific actions:
 * `--status`: Lists uncommitted file changes categorized as **new**, **modified**, and **deleted** — fast, no AI, no network. 📖 [Full docs](https://github.com/gitpr-cli/gitpr.git/blob/main/docs/git-status.md)
 * `--no-unstaged-check`: Skips the unstaged files verification before AI processing for a single invocation. Equivalent to `GITPR_SKIP_UNSTAGED_CHECK=true` for one run. 📖 [Full docs](https://github.com/gitpr-cli/gitpr.git/blob/main/docs/git-status.md)
 * `--linter-setup`: **Interactive external linter wizard.** Guides you through installing and configuring external linters (ESLint, PHPCS, Stylelint) as a Checkstyle XML bridge. 📖 [Full docs](https://github.com/gitpr-cli/gitpr.git/blob/main/docs/linter-regras-customizadas.md)
-* `--mcp`: Starts GitPR as an **MCP server** (Model Context Protocol) on stdio transport. Enables integration with VS Code, Cursor, Claude Desktop, and other MCP-compatible editors — exposing all GitPR AI capabilities as 12 annotated tools, 15 resources, and 7 pre-built prompts directly inside your IDE. Also available as the standalone `gitpr-mcp` command.
+* `--mcp`: Starts GitPR as an **MCP server** (Model Context Protocol) on stdio transport. Enables integration with VS Code, Cursor, Claude Desktop, and other MCP-compatible editors — exposing all GitPR AI capabilities as 14 annotated tools, 18 resources, and 7 pre-built prompts directly inside your IDE. Also available as the standalone `gitpr-mcp` command.
 * `--plugins`: Lists all **globally installed plugins** — custom linter packs from `~/.gitpr/plugins/linter/` and MCP prompt templates from `~/.gitpr/plugins/prompts/`. These plugins apply across all your projects without duplication. 📖 [Full docs](https://github.com/gitpr-cli/gitpr.git/blob/main/docs/plugins-system.md)
 * `--install`: **Interactive Setup Wizard.** Runs a guided 4-step setup: downloads skill templates, installs Git hooks, configures MCP for detected editors, and checks/requests your AI provider API key. Each step asks for confirmation before proceeding.
 * `-ih` or `--installhooks`: Automatically installs **local Git Hooks** (`pre-commit` and `prepare-commit-msg`) in your repository.
-* `-s` or `--skill`: Creates the AI context template files (`.gitpr.commit.md`, `.gitpr.pr.md`, `.gitpr.review.md`, `.gitpr.filereview.md`, `.gitpr.issue.md`, `.gitpr.blame.md`) and the Linter (`.gitpr.linter.yml`) at the project root.
+* `-s` or `--skill`: Creates the AI context template files (`.gitpr.commit.md`, `.gitpr.pr.md`, `.gitpr.review.md`, `.gitpr.filereview.md`, `.gitpr.issue.md`, `.gitpr.blame.md`) and the Linter (`.gitpr.linter.yml`) inside `.gitpr/skill/`.
 * `-is` or `--issue`: Automatically generates a draft of a **standardized Issue** and opens an interactive interface (TUI) for editing or direct submission via REST API. This feature has **3 context engines** depending on the command combination:
   * **New Code Issue (`gitpr -is`):** Reads the current `git diff`. **Why use:** Ideal for quickly documenting the task you just finished programming, before committing.
   * **Epic/Release Issue (`gitpr -is -ht`):** Reads the full history of the current branch (Git Log + PR Cache). **Why use:** Ideal for generating consolidated documentation of an entire release or a large *feature* that took several days/commits to complete.
@@ -229,7 +229,7 @@ You can dynamically switch models by configuring the `GEMINI_API_MODEL_PRIMARY`
 
 ## 🎯 Customizable "Skills" System (Prompt Engineering)
 
-Instead of hiding AI instructions in the source code, GitPR uses local Markdown files that act as *System Instructions*. When running `gitpr -s`, the following files are generated at the root of your project to customize the AI's "persona" according to your company's business rules:
+Instead of hiding AI instructions in the source code, GitPR uses local Markdown files that act as *System Instructions*. When running `gitpr -s`, the following files are generated inside `.gitpr/skill/` to customize the AI's "persona" according to your company's business rules:
 
 * `.gitpr.commit.md`: Rules for generating short commit messages.
 * `.gitpr.pr.md`: Required topic structure for the Pull Request description.
@@ -335,6 +335,8 @@ Once configured, use natural language in your editor's AI chat:
 | `generate_issue` | Structured issue from diff, history, or blame |
 | `list_unstaged_files` | Uncommitted file changes categorized (new/modified/deleted) |
 | `analyze_unstaged_diff` | Unstaged diff only (working tree vs index) |
+| `list_fix_candidates` | Fix candidates of the last review: patch, classification, id (read-only) |
+| `review_remote_pr` | AI review of a pull request already open on the forge, fetched by number (read-only) |
 
 ### Direct CLI Invocation
 
@@ -430,12 +432,14 @@ If you want to implement GitPR as an automated quality barrier in your team, che
 
 * [**Pull Request (Default Mode)**](https://github.com/gitpr-cli/gitpr.git/blob/main/docs/pr-descricao-padrao.md) — Complete flow for generating PR descriptions without flags.
 * [**Pull Request Publisher TUI**](https://github.com/gitpr-cli/gitpr.git/blob/main/docs/pull-request-publication.md) — How to review and publish Pull Requests directly to GitHub from the terminal.
-* [**AI Code Review**](https://github.com/gitpr-cli/gitpr.git/blob/main/docs/code-review-ia.md) — Guide to review modes (`--review`, `--fullreview`) and file auditing (`--input`).
+* [**AI Code Review**](https://github.com/gitpr-cli/gitpr.git/blob/main/docs/code-review-ia.md) — Guide to review modes (`--review`, `--fullreview`), file auditing (`--input`) and remote pull request review (`gitpr review-pr`).
+* [**Remote Pull Request Review (gitpr review-pr)**](https://github.com/gitpr-cli/gitpr.git/blob/main/docs/review-pr.md) — How the `gitpr review-pr` subcommand reviews a pull request that is already open on the forge, fetching its diff from the API without checking out its branch.
 * [**AI Commit Messages**](https://github.com/gitpr-cli/gitpr.git/blob/main/docs/commit-message-ia.md) — How to generate messages in the Conventional Commits standard and integrate with Git Hooks.
 * [**Issue Generation and TUI Interface**](https://github.com/gitpr-cli/gitpr.git/blob/main/docs/issue-tui-help.md) — How to use the terminal graphical interface (TUI) and the 3 context engines to manage structured Issues.
 * [**Code Archaeologist (Git Blame)**](https://github.com/gitpr-cli/gitpr.git/blob/main/docs/blame-arqueologo.md) — How to trace the origin of business rules with `git blame` and AI.
 * [**Skills and Templates System**](https://github.com/gitpr-cli/gitpr.git/blob/main/docs/skill-template.md) — How to customize AI behavior with `.gitpr.*.md` files.
 * [**Release Notes & Changelog (gitpr release)**](https://github.com/gitpr-cli/gitpr.git/blob/main/docs/release-notes.md) — How the `gitpr release` subcommand generates the changelog / release notes of a repository, suggests the next semantic version and publishes releases on the forge.
+* [**Fix Command (gitpr fix)**](https://github.com/gitpr-cli/gitpr.git/blob/main/docs/fix-command.md) — How the `gitpr fix` subcommand turns the findings of the last review into patches you read before they touch your tree, classifies each one by safety, and undoes an applied patch on demand.
 
 ### Configuration & Infrastructure
 

gitpr_cli-1.2.0.dist-info/RECORD

@@ -0,0 +1,67 @@
+gitpr_cli-1.2.0.dist-info/licenses/LICENSE,sha256=UQzLC6mwqb-ZNDkBBAqOxC-SMQcEID0gaM6oFcu6u48,27030
+src/__init__.py,sha256=WZW9x8P5CYnru29UiarFa8q1nQzuwYWyIodSOlruCWI,29
+src/ai_providers.py,sha256=QyFEOsXyA59lZziO5bl-4nlYFJpxBcxU9aLhcTBnkWE,14633
+src/blame_engine.py,sha256=Zov1hxrt7Dcmf56ovJKz3QmHIfzDKtSub2iItXeg9-A,17104
+src/cache.py,sha256=GC2ZT0ThAtH22aRRZapIeauATv_nqBfiRsXjqeMhAiI,6368
+src/changelog_builder.py,sha256=6tqJVHpY4Ta5shOSNWDEo_eHlKmDrU2dUSzS0wyYe10,7116
+src/chat_memory.py,sha256=PcohnyspAfIwwwbsp9U2Wk_1AylJMFEMoWqRSJe_w30,8949
+src/commit_classifier.py,sha256=8_LHQEzisWflV5ZsK31P3fwyu2QWacv8cdFErXYe030,4668
+src/config.py,sha256=9945b0sTM5xL4a81M-CyJ6qfnAQ-Vkf3vr9Old79bjw,33985
+src/config_schema.py,sha256=a7fSsp39h9Pe97uMrUTncuLjGzB8mWvF77vYhJZIcmw,35283
+src/core.py,sha256=zWrz86ifjly5DqoqfQvqdQ0up6yyqFlDYYdB8t8cnyU,76310
+src/diff_parser.py,sha256=62X69ves4tlzNgVKXNsfWdOMwqn2h16RkUOPqszMU7k,11252
+src/doc_links.py,sha256=DuovU-RF9kjbisBdEoGAsuavQ2y0MYDhqy6Sz6vbrGI,1157
+src/github_api.py,sha256=ILm7e03UP80swc3XuJvIJJvhLp81-V_3kQQ1EuAcN8Q,3836
+src/i18n.py,sha256=8s-Bsvp-p3ZZhj0FvDNTRqfWsEio5v4VvXmS1NSqNW8,4578
+src/issue_engine.py,sha256=XwW9C5uX84eOWWZMerm0DvE8LpLRcjhrLsXtt7yftak,10872
+src/linter_engine.py,sha256=8SIs_FUtxEBeTMKWuqatqzlyyToYIjG5xeg0X7FYRAA,12683
+src/linter_wizard.py,sha256=7J_nQFghVwCaKG6QI3K45KcOaxHSH1oTCAvP_6tnCQM,7424
+src/main.py,sha256=B9X4Yr1QZxCioL2dvXOkJ7m37DzOR6DB2E6unLBqHuI,98535
+src/mcp_server.py,sha256=q2clMxTtjLLC-P6Vh4XMq0xylnoUw-K2IFQvgvHIf7o,90479
+src/metrics.py,sha256=VC68t8xGUBFry-ANvEb_TjaRkboPi37__GxK0eHZOe4,21250
+src/net.py,sha256=f0MOkXyIxi4HVetsybtWpV_yHjiC17T2JMWySBdoV2I,2774
+src/release_engine.py,sha256=gwTnOq0ESK26BqsW1RLlysGwL0ZS5guj1qG2V-yUyWI,21610
+src/reviewer_resolution.py,sha256=lZ3csSJHN6Y5TJ3U3p9uuXikOgG7fH4qqx7X9-4F7pc,5754
+src/reviewer_suggestion.py,sha256=jk9FAyp7WkkJbi-yepz7LDF1_tSHxJEEjcajMKF5HdY,8501
+src/security.py,sha256=fkWy-r4tdISH0xOyWZyR8ZCFyLuYqu8fn9sKsIeNAso,1351
+src/spinner.py,sha256=63pEotiDa3TVwBr-sn0MJoDPQ_uWJTTwINXieJIDDKc,8964
+src/suggest_reviewers.py,sha256=-wRe0HYksx4FWQZ-eDvU9Ghdk5gqtEt457KQP7sE31Y,5181
+src/tui_issue.py,sha256=WEFxT4kzWQ_KbSDsOnS_iuYXasZySB-68jryGNm0peg,10331
+src/updater.py,sha256=hm4brG_lSPe82LWFVb41iVMOBYUI824svwsfzMfxxUQ,4771
+src/usage_log.py,sha256=314sTT5Gl_uOXqWSIi-V42bnTHuMSYAZ6G6-_15uEGQ,5968
+src/version_bump.py,sha256=OjqJe9OL1-1dT5od7-SgU5ZN_dTtR_aKX_oSjSY64YQ,2954
+src/fix/__init__.py,sha256=YcN1nGpzWyXczd35LWEXdFpviKKsb9JS8e4bD6KcVcQ,935
+src/fix/apply_fix.py,sha256=yMinMW50Fw62V-k7hRwN1lsikCJbz6xV5iX4IH0sdxY,17912
+src/fix/fix_history.py,sha256=pHXHpt-SstitbLDmlPloDVKlH79FUD8hqKjVK5xKLMA,5626
+src/fix/patch_applier.py,sha256=ahruhdKOhrt2excc3nJljoWi4_0pUj26owSVAQWeeKQ,6258
+src/fix/patch_extractor.py,sha256=gEiwSvxGNHM74TnubTN22NBpbXDwvNpB2vMqgpF7hQQ,2514
+src/fix/patch_provenance.py,sha256=sF934ntcZjJCTzzhtuR2Pi-vMUdXFElKcuw5riwPnUk,3195
+src/fix/patch_safety_classifier.py,sha256=StZvh_0Y-ZCdf_LwFoVAdfhpi0b7v1TOxTRoG86_63M,4121
+src/fix/rollback_fix.py,sha256=rMBYewuvlhJOFdo-Ke_SPfu1O-yd6NJzEYiO5o-LcKk,2501
+src/infrastructure/__init__.py,sha256=zTIcY3cscsdkXogSGXoJbH0mVGFbKYTRK32fg26wr9U,81
+src/infrastructure/scm/__init__.py,sha256=3_4CmrZPZSmDbFwf_0YBeKc935jWV8oAHaZtC_i-Fak,1550
+src/infrastructure/scm/azure_devops_provider.py,sha256=ezRS_O2450mfYEfqFOD4UBSVFhT3aJdx3DEVkZcJjac,14717
+src/infrastructure/scm/base.py,sha256=L-7NYG2hqIWKzKCsQ5NurtgJJvvYvjTVOs0NJ2Aqx_M,9536
+src/infrastructure/scm/bitbucket_provider.py,sha256=jv62MlOd0Iiu3VYAHUffJO2D95TE4k2lmukQL93bOd8,12105
+src/infrastructure/scm/factory.py,sha256=Qeearq8eWnF_4boI68rVBkoGEiS-xRGD4V-O_FF0204,3856
+src/infrastructure/scm/github_provider.py,sha256=9RXRAocYEVWS6NR2PsOg7wtqiF8989x1BxAqgBSoLOI,16438
+src/infrastructure/scm/gitlab_provider.py,sha256=mznhtBhXtc4zRut_Jb0pM57RIXViNYYiKiM7VKDFhl8,13881
+src/review/__init__.py,sha256=XVyQZFf6UAqIKi8Gb1pQ4dKjcrC50soUvNuKu1DBspc,687
+src/review/diff_normalizer.py,sha256=PRXQ7b81EBH1nx_8_Tm9G7Kjdax_6hIVtFgEixicT0E,3551
+src/review/diff_source.py,sha256=O8bIlff0Zr7wPMzCKAVk_LF2pMDr3sc9KAuyA3OvHG0,2664
+src/review/remote_pr.py,sha256=maPCQO5p7u9-_PA0FEBNgimfA6na2BdQGBPmqNj1_0c,11817
+src/review/render.py,sha256=Hnkz0PvtYIaaikTS1inTzrneAJL4Xqjd-F8Lr6Yh6mQ,2691
+src/ui/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
+src/ui/chat_app.py,sha256=6qpjWxHECYbKc5PTfbr63FjGo0j5-gXf3X-qW1klWBg,25396
+src/ui/config_app.py,sha256=xgppZbEBFrFijgLGo1b1PN1vTGCg8xpI8gpMYBv2KzU,73417
+src/ui/help_screen.py,sha256=EvbP8kFZ6Sdehi7bJ6Wo_YcTTk5MTkgneeidWwefYrA,1890
+src/ui/issue_app.py,sha256=eINMmamloeopbtwyS1rCPwosaHeZW5oLebnEV4TfamY,9281
+src/ui/linter_app.py,sha256=Y5vwU7D5PQIR_HaY6t40jZy7Guc8z2w-auyosCQqoro,1566
+src/ui/metrics_app.py,sha256=a3jsB_kjMusSL57A6LpJyb9LFN7JJWlOdUNClklwq-c,16016
+src/ui/pr_publish_app.py,sha256=cZ9xr1TSZF6fjvQPlAdSjzSNykufHF1l7hsmo3Ib2j0,67311
+src/ui/pr_publish_help.py,sha256=fBiUtD97ngxbSfAvPVGGXU19r5kbALCKyi4QeKqyQ2c,1891
+gitpr_cli-1.2.0.dist-info/METADATA,sha256=1BSAiMJtMgY7j3YsU0YgaHw7rpfk2ccj5z2iOzSjyxk,34689
+gitpr_cli-1.2.0.dist-info/WHEEL,sha256=YVMoNqKzERt-wjUZwJ33xBGAwnFl-4cqbYkTtWa4itE,91
+gitpr_cli-1.2.0.dist-info/entry_points.txt,sha256=TstGnMUzdOI5qNmpM97EPDBPuMqOZxoo6EoLxHSOMSQ,71
+gitpr_cli-1.2.0.dist-info/top_level.txt,sha256=74rtVfumQlgAPzR5_2CgYN24MB0XARCg0t-gzk6gTrM,4
+gitpr_cli-1.2.0.dist-info/RECORD,,

src/cache.py

@@ -47,9 +47,22 @@ def get_cached_response(action_folder, prompt_text):
 
 
 def save_cached_response(
-    action_folder, action_type, prompt_text, response_dict, meta_raw=None
+    action_folder,
+    action_type,
+    prompt_text,
+    response_dict,
+    meta_raw=None,
+    reviewed_diff=None,
 ):
-    """Saves the AI response to the local cache."""
+    """Saves the AI response to the local cache.
+
+    ``reviewed_diff`` is the diff the review actually ran on. When given it is
+    stored at the top level of the record, so ``gitpr fix`` can re-derive its
+    patch from the revision that was reviewed instead of whatever the working
+    tree happens to hold later — the two differ whenever the review came from a
+    pull request (src/review/remote_pr.py) or from a full-branch diff. Records
+    written before this field existed simply lack it, and the reader falls back.
+    """
     md5_hash = generate_md5(prompt_text)
     folder_path = get_cache_base_dir() / action_folder
     folder_path.mkdir(parents=True, exist_ok=True)
@@ -76,6 +89,9 @@ def save_cached_response(
         "response": response_dict,
     }
 
+    if reviewed_diff is not None:
+        cache_data["diff"] = reviewed_diff
+
     try:
         with open(cache_file, "w", encoding="utf-8") as f:
             json.dump(cache_data, f, indent=2, ensure_ascii=False)
@@ -83,6 +99,51 @@ def save_cached_response(
         pass  # Silent cache failure to avoid crashing the tool
 
 
+# Review action types whose prompt is a branch diff. `filereview` is left out
+# on purpose: it audits a single file and has no diff to re-derive.
+REVIEW_ACTION_TYPES = ("review", "fullreview")
+
+
+def resolve_last_review(repo_name, branch_name):
+    """The most recent cached review for *repo_name* and *branch_name*, or None.
+
+    All three review modes write into the same ``review/`` folder (see
+    ``core.generate_pr_content``), so ``action_type`` is what tells them apart —
+    without that filter a file audit would be mistaken for a branch review.
+
+    Returns the cache record itself: the caller reads
+    ``record["response"]["review"]`` for the text and ``record["action_type"]``
+    to know which diff produced it, so the patch can be re-derived against
+    today's tree rather than the tree of the day the review ran.
+    """
+    review_folder = get_cache_base_dir() / "review"
+    if not review_folder.exists():
+        return None
+
+    newest = None
+    for cache_file in review_folder.glob("*.json"):
+        try:
+            with open(cache_file, "r", encoding="utf-8", errors="replace") as f:
+                data = json.load(f)
+        except (json.JSONDecodeError, IOError):
+            continue
+
+        if not isinstance(data, dict):
+            continue
+        if data.get("repo") != repo_name or data.get("branch") != branch_name:
+            continue
+        if data.get("action_type") not in REVIEW_ACTION_TYPES:
+            continue
+        if not (data.get("response") or {}).get("review"):
+            continue
+
+        # Timestamps are '%Y-%m-%d %H:%M:%S', so string order is date order.
+        if newest is None or data.get("datetime", "") > newest.get("datetime", ""):
+            newest = data
+
+    return newest
+
+
 def get_cached_pr_descriptions(repo_name, branch_name):
     """Searches the cache for all historically generated PRs for this repository and branch."""
     from src.i18n import __

src/config.py

@@ -61,6 +61,16 @@ DEFAULT_CONFIG = {
     "GITPR_RELEASE_AI_SUMMARY": "true",
     "GITPR_RELEASE_AUTO_BUMP": "true",
     "GITPR_RELEASE_PUBLISH_DRAFT_BY_DEFAULT": "true",
+    # Fix feature (gitpr fix). The excluded paths are glob patterns a patch may
+    # not touch and still count as safe; they are about *risk*, not diff noise,
+    # which is why they are not shared with the smart-excludes list.
+    "GITPR_FIX_SAFE_MAX_LINES_CHANGED": "5",
+    "GITPR_FIX_SAFE_EXCLUDED_PATHS": (
+        "database/migrations/**;**/*.ci.yml;docker/**;terraform/**;.github/workflows/**"
+    ),
+    "GITPR_FIX_REQUIRE_CONFIRMATION": "true",
+    "GITPR_FIX_CREATE_BRANCH_ON_ALL_SAFE": "true",
+    "GITPR_FIX_BRANCH_NAME_TEMPLATE": "fix/gitpr-{datetime}",
 }
 
 # Fallbacks used when the .env value is missing or not a positive number.
@@ -131,6 +141,7 @@ SKILL_FILES_BY_TYPE = {
     "issue": ".gitpr.issue.md",
     "blame": ".gitpr.blame.md",
     "release": ".gitpr.release.md",
+    "fix": ".gitpr.fix.md",
 }
 SKILL_TYPES = tuple(SKILL_FILES_BY_TYPE)
 
@@ -680,6 +691,47 @@ def get_release_settings():
     }
 
 
+def get_fix_settings():
+    """Returns the gitpr fix configuration as a flat dict.
+
+    ``safe_excluded_paths`` is the ; separated list of glob patterns a patch may
+    not touch and still be classified safe; an empty value falls back to the
+    built-in list, because clearing the field by accident must not silently drop
+    the protection. ``branch_name_template`` names the branch of a batch with
+    the vocabulary the output filenames already use ({branch} and {datetime}),
+    so there is one placeholder syntax to learn.
+    """
+    load_dotenv(ENV_FILE)
+    try:
+        max_lines = int(
+            (os.getenv("GITPR_FIX_SAFE_MAX_LINES_CHANGED") or "").strip() or 5
+        )
+        if max_lines <= 0:
+            raise ValueError
+    except ValueError:
+        max_lines = 5
+
+    # The fallback is spelled out, like every other getter here, instead of
+    # looked up in DEFAULT_CONFIG.
+    raw_paths = os.getenv("GITPR_FIX_SAFE_EXCLUDED_PATHS") or (
+        "database/migrations/**;**/*.ci.yml;docker/**;terraform/**;.github/workflows/**"
+    )
+    excluded_paths = tuple(
+        item.strip() for item in raw_paths.split(";") if item.strip()
+    )
+
+    template = (os.getenv("GITPR_FIX_BRANCH_NAME_TEMPLATE") or "").strip()
+    return {
+        "safe_max_lines_changed": max_lines,
+        "safe_excluded_paths": excluded_paths,
+        "require_confirmation": _env_bool_default_true("GITPR_FIX_REQUIRE_CONFIRMATION"),
+        "create_branch_on_all_safe": _env_bool_default_true(
+            "GITPR_FIX_CREATE_BRANCH_ON_ALL_SAFE"
+        ),
+        "branch_name_template": template or "fix/gitpr-{datetime}",
+    }
+
+
 def validate_github_token(token):
     """
     Validates a GitHub PAT by making a lightweight API call to /user.

src/config_schema.py

@@ -180,6 +180,12 @@ CATEGORIES = (
         __("Changelog generation and release publishing."),
         doc="release-notes.md",
     ),
+    Category(
+        "fix",
+        __("Fix"),
+        __("Review findings turned into patches, and the safety rules of a batch."),
+        doc="fix-command.md",
+    ),
     Category(
         "scm",
         __("SCM / Forge"),
@@ -221,6 +227,7 @@ SKILL_LABELS = {
     "issue": __("Issue"),
     "blame": __("Blame"),
     "release": __("Release"),
+    "fix": __("Fix"),
 }
 
 # Sub-headers shown inside a category. Generic words go through __(); forge and
@@ -614,6 +621,59 @@ FIELDS = (
         kind=KIND_TEMPLATE,
         default="{branch}_{datetime}_RELEASE.md",
     ),
+    # --------------------------------------------------------------------- Fix
+    ConfigField(
+        key="GITPR_FIX_SAFE_MAX_LINES_CHANGED",
+        label=__("Safe Patch Size"),
+        description=__(
+            "Maximum number of added plus removed lines a patch may have and still be classified as safe."
+        ),
+        category="fix",
+        kind=KIND_INT,
+        default="5",
+    ),
+    ConfigField(
+        key="GITPR_FIX_SAFE_EXCLUDED_PATHS",
+        label=__("Sensitive Paths"),
+        description=__(
+            "Semicolon separated glob patterns a patch may not touch and still be safe (migrations, workflows, containers, infrastructure)."
+        ),
+        category="fix",
+        kind=KIND_STR,
+        default=(
+            "database/migrations/**;**/*.ci.yml;docker/**;terraform/**;.github/workflows/**"
+        ),
+    ),
+    ConfigField(
+        key="GITPR_FIX_REQUIRE_CONFIRMATION",
+        label=__("Ask Before Applying"),
+        description=__(
+            "Shows the diff and asks for confirmation before a patch is written to the working tree."
+        ),
+        category="fix",
+        kind=KIND_BOOL,
+        default="true",
+    ),
+    ConfigField(
+        key="GITPR_FIX_CREATE_BRANCH_ON_ALL_SAFE",
+        label=__("Branch for a Batch"),
+        description=__(
+            "Creates a new branch when the safe patches are applied in a batch, so the current branch stays untouched."
+        ),
+        category="fix",
+        kind=KIND_BOOL,
+        default="true",
+    ),
+    ConfigField(
+        key="GITPR_FIX_BRANCH_NAME_TEMPLATE",
+        label=__("Branch Name"),
+        description=__(
+            "Name of the branch created for a batch. Placeholders: branch and datetime."
+        ),
+        category="fix",
+        kind=KIND_TEMPLATE,
+        default="fix/gitpr-{datetime}",
+    ),
     # --------------------------------------------------------------- SCM/Forge
     # The common settings stay ungrouped and always visible; the fields that
     # only exist for one forge live under its sub-header and follow the

src/core.py

@@ -330,6 +330,23 @@ def _load_docs_smart_excludes():
 SMART_EXCLUDES = _load_smart_excludes() + _load_docs_smart_excludes()
 
 
+def get_smart_exclude_patterns():
+    """
+    Return the raw glob patterns behind SMART_EXCLUDES, without the prefix.
+
+    SMART_EXCLUDES is a list of git pathspecs (":(exclude)*.lock") because that
+    is what `git diff` takes on its command line. A diff fetched from a forge
+    API never passes through git, so the same exclusions have to be applied in
+    Python — this exposes the patterns both paths share, derived from the very
+    list the git command uses so the two can never drift apart.
+
+    GITPR_SKIP_SMART_EXCLUDES is honored upstream in the loaders, so an empty
+    SMART_EXCLUDES already yields an empty list here.
+    """
+    prefix = ":(exclude)"
+    return [p[len(prefix):] for p in SMART_EXCLUDES if p.startswith(prefix)]
+
+
 def _get_raw_docs_patterns():
     """
     Returns raw glob patterns for documentation extensions (e.g. ['*.md', '*.txt']).
@@ -742,8 +759,30 @@ def split_diff_into_chunks(diff_text, max_tokens=90000):
     return chunks
 
 
-def generate_pr_content(action_folder, action_type, diff_text, provider="gemini"):
-    """Sends the diff to the AI using System Instruction and returns a parsed JSON."""
+def generate_pr_content(
+    action_folder,
+    action_type,
+    diff_text,
+    provider="gemini",
+    cache_scope="",
+    store_diff=False,
+):
+    """Sends the diff to the AI using System Instruction and returns a parsed JSON.
+
+    The engine has no idea where ``diff_text`` came from and does not need one:
+    a local ``git diff`` and a pull request fetched over the API are the same
+    text by the time they arrive, so both review flows share this function.
+
+    ``cache_scope`` is appended to the cache key only — never to the prompt the
+    model receives — so a review of PR 42 and a review of an identical local
+    diff cannot answer each other. The default "" leaves the MD5 of every
+    existing local review untouched, so switching a caller to DiffSource
+    invalidates no cache.
+
+    ``store_diff`` records the reviewed diff alongside the response, which is
+    what lets ``gitpr fix`` patch against the revision that was reviewed rather
+    than the working tree of the day it runs.
+    """
     if not diff_text or not diff_text.strip():
         click.secho(
             __("⚠️ No diff found. Make some changes before running the command."),
@@ -870,7 +909,7 @@ def generate_pr_content(action_folder, action_type, diff_text, provider="gemini"
         pass  # Non-critical — never block the main flow for this metadata
 
     # TRY TO RETRIEVE FROM CACHE
-    cached_data = get_cached_response(action_folder, prompt)
+    cached_data = get_cached_response(action_folder, prompt + cache_scope)
     if cached_data:
         click.secho(__("⚡ Response retrieved from local cache."), fg="green", dim=True)
         from src.metrics import log_command_metric
@@ -1061,7 +1100,12 @@ def generate_pr_content(action_folder, action_type, diff_text, provider="gemini"
     # SAVE TO CACHE AND RETURN
     if result_json:
         save_cached_response(
-            action_folder, action_type, prompt, result_json, meta_raw=total_meta
+            action_folder,
+            action_type,
+            prompt + cache_scope,
+            result_json,
+            meta_raw=total_meta,
+            reviewed_diff=diff_text if store_diff else None,
         )
         # Fire-and-forget metric for successful AI-powered command
         from src.metrics import log_command_metric
@@ -1200,6 +1244,7 @@ def generate_skill_template():
         ".gitpr.blame.md": f"gitpr.blame{lang_suffix}.md",
         ".gitpr.issue.md": f"gitpr.issue{lang_suffix}.md",
         ".gitpr.release.md": f"gitpr.release{lang_suffix}.md",
+        ".gitpr.fix.md": f"gitpr.fix{lang_suffix}.md",
     }
 
     success_count = 0

src/diff_parser.py

@@ -5,8 +5,14 @@ The PR flows diff with ``git diff -U1 -w -M -B <ancestor> -- <excludes>``
 numbers of the added ('+') lines, per file — the numbers used to blame the
 working tree. It never runs git and never re-applies excludes: it only
 interprets the diff text it is given.
+
+``summarize_patch()`` reads the same text a second way — as a patch a whole:
+which files it touches, how many hunks, how many lines it adds and removes.
+That is what the ``gitpr fix`` safety classifier scores, and it accepts the
+looser shape an AI returns (no ``diff --git`` line, just ``---``/``+++``).
 """
 import re
+from dataclasses import dataclass
 
 _HUNK_HEADER_RE = re.compile(r"^@@ -\d+(?:,\d+)? \+(\d+)(?:,(\d+))? @@")
 
@@ -157,3 +163,159 @@ def parse_added_lines(diff_text):
     for lines in added_by_file.values():
         lines.sort()
     return added_by_file
+
+
+@dataclass(frozen=True)
+class PatchSummary:
+    """What a unified diff touches, as the safety classifier needs to see it.
+
+    ``files`` holds each file the patch mentions, in order of appearance, taken
+    from the new side — the 'diff --git' line when there is one (a deletion
+    keeps its name there, since its '+++ /dev/null' carries none), otherwise the
+    '+++ b/path' header. ``hunks`` counts well-formed '@@' headers, and
+    ``added``/``removed`` count content lines inside those hunks, so headers and
+    the '\\ No newline' marker never count. ``removed_lines`` keeps the text of
+    each removed line — what lets a caller spot a deleted function call.
+    """
+    files: tuple[str, ...]
+    hunks: int
+    added: int
+    removed: int
+    removed_lines: tuple[str, ...]
+
+
+def summarize_patch(diff_text):
+    """Return a PatchSummary for *diff_text* (see the dataclass docstring).
+
+    Pure like parse_added_lines: it interprets the text it is given and never
+    runs git. Text with no recognizable diff section yields an empty summary
+    rather than an error — the caller decides whether that means "no patch".
+    """
+    files = []
+    hunks = 0
+    added = 0
+    removed = 0
+    removed_lines = []
+    in_hunk = False
+
+    for raw_line in diff_text.split("\n"):
+        line = raw_line.rstrip("\r")
+
+        if line.startswith("diff --git "):
+            _remember_file(files, _strip_diff_prefix(_header_path(line)))
+            in_hunk = False
+            continue
+        if line.startswith("+++ ") and not in_hunk:
+            path = _strip_diff_prefix(_header_path(line, marker="+++ "))
+            if path and path != "/dev/null":
+                _remember_file(files, path)
+            in_hunk = False
+            continue
+        if line.startswith("@@"):
+            if _HUNK_HEADER_RE.match(line):
+                hunks += 1
+                in_hunk = True
+            continue
+
+        if not in_hunk or line.startswith("\\ No newline"):
+            continue
+        if line.startswith("+"):
+            added += 1
+        elif line.startswith("-"):
+            removed += 1
+            removed_lines.append(line[1:])
+
+    return PatchSummary(
+        files=tuple(files),
+        hunks=hunks,
+        added=added,
+        removed=removed,
+        removed_lines=tuple(removed_lines),
+    )
+
+
+def _remember_file(files, path):
+    """Append *path* to *files* once, preserving appearance order."""
+    if path and path not in files:
+        files.append(path)
+
+
+@dataclass(frozen=True)
+class PatchSection:
+    """One file's slice of a unified diff, with the path it belongs to.
+
+    ``path`` is the new-side path (the b/ side), so a deletion keeps its name —
+    its own '+++ /dev/null' header carries none. ``text`` is the section
+    verbatim, headers included, so the pieces of a split diff concatenate back
+    into something the review engine can still read.
+    """
+    path: str
+    text: str
+
+
+def split_patch_sections(diff_text):
+    """Split *diff_text* into one PatchSection per file, in order.
+
+    The counterpart of summarize_patch for callers that must drop whole files
+    (the remote-PR smart-excludes filter): it needs each file's path next to its
+    own text, which no other helper here exposes. Accepts both header shapes
+    summarize_patch accepts — git's own output ('diff --git' plus a '---'/'+++'
+    pair per file) and the looser shape an AI returns (bare '---'/'+++' pairs).
+    Text before the first header is dropped, as are sections whose path never
+    resolves; a diff with no recognizable file yields [].
+
+    Pure, like its siblings: it interprets the text it is given, never runs git.
+    """
+    lines = [raw.rstrip("\r") for raw in diff_text.split("\n")]
+    sections = []
+    start = None
+    path = ""
+    in_hunk = False
+
+    def flush(end):
+        if start is not None and path:
+            sections.append(
+                PatchSection(path=path, text="\n".join(lines[start:end]))
+            )
+
+    i = 0
+    while i < len(lines):
+        line = lines[i]
+
+        if line.startswith("diff --git "):
+            flush(i)
+            start, in_hunk = i, False
+            # The b/ side names the file on the new revision, and unlike the
+            # '+++ /dev/null' of a deletion it is always populated.
+            path = _strip_diff_prefix(_header_path(line))
+            i += 1
+            continue
+
+        # A '---' immediately followed by a '+++' is a file header pair. Inside
+        # a 'diff --git' section it is that section's own header; anywhere else
+        # it opens a new file, which is how the AI's looser patches arrive.
+        if (
+            line.startswith("--- ")
+            and i + 1 < len(lines)
+            and lines[i + 1].startswith("+++ ")
+        ):
+            candidate = _strip_diff_prefix(_header_path(lines[i + 1], marker="+++ "))
+            if candidate == "/dev/null":
+                candidate = ""
+            already_open = start is not None and lines[start].startswith("diff --git ")
+            if not already_open:
+                flush(i)
+                start, in_hunk = i, False
+                path = candidate
+            elif candidate:
+                path = candidate
+            i += 1
+            continue
+
+        if line.startswith("@@"):
+            in_hunk = bool(_HUNK_HEADER_RE.match(line))
+
+        i += 1
+
+    flush(len(lines))
+    return sections

src/fix/__init__.py

@@ -0,0 +1,16 @@
+"""The ``gitpr fix`` feature: turn a review finding into a reviewable patch.
+
+Modules, in dependency order:
+  - patch_provenance.py — the data contract (safety classes, candidates, results)
+  - patch_extractor.py — reading the code an AI hands back (shared with the chat)
+  - patch_safety_classifier.py — SAFE / REVIEW_REQUIRED / EXPERIMENTAL, pure logic
+  - patch_applier.py — the project's first shared git wrapper (apply, check, reverse)
+  - fix_history.py — the applied-patch record that makes --rollback deterministic
+  - apply_fix.py — the use case: review -> AI findings -> classify -> dry-run/apply
+  - rollback_fix.py — the use case that undoes an applied patch
+
+Import the submodule you need (``from src.fix.apply_fix import apply_fix``); this
+package deliberately re-exports nothing, because the chat imports
+``src.fix.patch_extractor`` on every session and must not drag the AI and git
+layers in with it.
+"""