git-cai-cli 0.17.2__tar.gz → 0.17.4__tar.gz

{git_cai_cli-0.17.2 → git_cai_cli-0.17.4}/.github/ci/cai_config.ci.yml

@@ -6,14 +6,13 @@ load_tokens_from: /home/runner/work/.config/cai/tokens.yml
 timeout: 30
 full_files: false
 anthropic:
-  model: claude-haiku-4-5
-  temperature: 0
+  model: claude-haiku-5-5
   max_tokens: 32768
 deepseek:
-  model: deepseek-chat
+  model: deepseek-flash
   temperature: 0
 gemini:
-  model: gemini-3.1-flash-lite
+  model: gemini-3.5-flash-lite
   temperature: 0
 groq:
   model: openai/gpt-oss-20b
@@ -22,11 +21,11 @@ mistral:
   model: codestral-2508
   temperature: 0
 ollama:
-  model: llama3.1
+  model: gemma4:e2b
   temperature: 0
   timeout: 300
 openai:
-  model: gpt-5.4-mini
+  model: gpt-6-luna
 xai:
   model: grok-4.3
   temperature: 0

{git_cai_cli-0.17.2 → git_cai_cli-0.17.4}/.github/workflows/python-tests.yml

@@ -99,7 +99,7 @@ jobs:
 
       # MegaLinter
       - name: MegaLinter
-        uses: oxsecurity/megalinter@15e5b45552097e318c93de385779ce3b1084052c # v10.0.0
+        uses: oxsecurity/megalinter@f2fe8c1cba37091848df3e524066c8c23ef353ed # v10.2.0
         id: ml
         env:
           VALIDATE_ALL_CODEBASE: false

{git_cai_cli-0.17.2 → git_cai_cli-0.17.4}/.mega-linter.yml

@@ -54,7 +54,7 @@ LINTER_RULES_PATH: .linters
 LOG_LEVEL: INFO
 MARKDOWN_DEFAULT_STYLE: markdownlint
 # Pins the Docker image for local runs to the version the CI workflow pins.
-MEGALINTER_VERSION: v10.0.0
+MEGALINTER_VERSION: v10.2.0
 PARALLEL: true
 PRINT_ALPACA: false
 PYTHON_BANDIT_RULES_PATH: .linters

{git_cai_cli-0.17.2 → git_cai_cli-0.17.4}/Makefile

@@ -1,10 +1,13 @@
 #!/usr/bin/make -f
 
-PIP := pipx
 UV := uv
 SHELL := /bin/bash
+MEGALINTER_VERSION := $(shell sed -n 's/^MEGALINTER_VERSION: *//p' .mega-linter.yml)
+MEGALINTER_IMAGE := ghcr.io/oxsecurity/megalinter:$(MEGALINTER_VERSION)
+# --platform: the image is amd64 only. z: relabels the mount on SELinux hosts, ignored elsewhere.
+MEGALINTER_RUN := docker run --rm --platform linux/amd64 -v "$(CURDIR):/tmp/lint:rw,z"
 
-.PHONY: help lint check-docker check-npx lint-fix add-lint-hook clean test
+.PHONY: help lint check-docker lint-fix add-lint-hook clean test
 
 help: ## Shows this help message
 	@echo "Available commands:"
@@ -27,25 +30,16 @@ check-docker: ## Checks if docker is installed
 		docker --version; \
 	fi
 
-check-npx: ## Checks if npx is installed
-	@if ! command -v npx &> /dev/null; then \
-		echo "npx is not installed. Please install it."; \
-		exit 1; \
-	else \
-		echo "npx version:"; \
-		npx --version; \
-	fi
-
 clean: ## Clean cache of uv and delete virtual environment
 	@$(UV) cache clean
 	@rm -rf .venv
 
-lint:
+lint: ## Checks the branch name and runs MegaLinter
 	@sh ./.linters/check_git_branch_name.sh
-	@npx mega-linter-runner@v10
+	@$(MEGALINTER_RUN) $(MEGALINTER_IMAGE)
 
-lint-fix: ## Lints the code using sqlfluff and fixes the issues
-	@npx mega-linter-runner@v10 --fix
+lint-fix: ## Runs MegaLinter and applies its fixes
+	@$(MEGALINTER_RUN) -e APPLY_FIXES=all $(MEGALINTER_IMAGE)
 
 test: ## Runs tests
 	@$(UV) run pytest

{git_cai_cli-0.17.2 → git_cai_cli-0.17.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: git-cai-cli
-Version: 0.17.2
+Version: 0.17.4
 Summary: Use LLM to create git commit messages
 Author-email: Thorsten Foltz <thorsten.foltz@live.com>
 License-Expression: MIT

{git_cai_cli-0.17.2 → git_cai_cli-0.17.4}/cai_config.yml

@@ -18,15 +18,15 @@ pr_prompt_file: "/home/thorsten/.config/cai/pr_prompt.md"
 stats: true
 signoff: true
 anthropic:
-  model: claude-haiku-4-5
-  temperature: 0
+  model: claude-haiku-5-5
   max_tokens: 32768
 deepseek:
-  model: deepseek-chat
+  model: deepseek-flash
   temperature: 0
 gemini:
-  model: gemini-3.1-flash-lite
+  model: gemini-3.5-flash-lite
   temperature: 0
+  timeout: 60
 groq:
   model: openai/gpt-oss-20b
   temperature: 0
@@ -34,11 +34,11 @@ mistral:
   model: codestral-2508
   temperature: 0
 ollama:
-  model: llama3.1
+  model: gemma4:e2b
   temperature: 0
   timeout: 300
 openai:
-  model: gpt-5.6-luna
+  model: gpt-6-luna
 xai:
   model: grok-4.3
   temperature: 0

{git_cai_cli-0.17.2 → git_cai_cli-0.17.4}/docs/git-cai.txt

@@ -360,7 +360,7 @@ provider keys.
 git cai -H language=de
 git cai -H default=anthropic
 git cai -H emoji=false
-git cai -H groq.model=llama-3.3-70b
+git cai -H groq.model=openai/gpt-oss-120b
 ----
 
 -h, --help::
@@ -455,7 +455,7 @@ identifier for that provider.
 +
 ----
 git cai -P openai -m gpt-4o
-git cai -P anthropic -m claude-haiku-4-5
+git cai -P anthropic -m claude-haiku-5-5
 git cai -P ollama -m mistral
 ----
 
@@ -528,7 +528,7 @@ both.
 +
 ----
 git cai -P anthropic
-git cai -P groq -m llama-3.3-70b
+git cai -P groq -m openai/gpt-oss-120b
 git cai -P openrouter
 ----
 
@@ -662,7 +662,7 @@ become integers or floats, everything else stays a string.
 ----
 git cai -S default=anthropic
 git cai -S emoji=false
-git cai -S groq.model=llama-3.3-70b
+git cai -S groq.model=openai/gpt-oss-120b
 git cai -S openai.temperature=0.7
 git cai -S conventional=true
 ----
@@ -695,9 +695,10 @@ git cai -s a1b2c3d
 
 -T, --timeout SECONDS::
 Override the HTTP timeout for the LLM call in this invocation. The default
-is 30 seconds for remote providers and 300 seconds for Ollama (since local
-generation can be significantly slower). The override applies uniformly to
-whichever provider is selected for this run.
+is 30 seconds for remote providers, 60 seconds for Gemini (Flash Lite can queue
+for up to 30 seconds) and 300 seconds for Ollama (since local generation can be
+significantly slower). The override applies uniformly to whichever provider is
+selected for this run, including those with their own `<provider>.timeout`.
 +
 To change the default permanently, set `timeout: <seconds>` in
 `cai_config.yml` or the home config:
@@ -707,11 +708,12 @@ git cai -H timeout=60
 git cai -S timeout=45
 ----
 +
-Ollama keeps its own timeout under `ollama.timeout` (default 300). Override
-it separately:
+Ollama and Gemini keep their own timeouts under `ollama.timeout` (default 300)
+and `gemini.timeout` (default 60). Override them separately:
 +
 ----
 git cai -H ollama.timeout=600
+git cai -H gemini.timeout=90
 ----
 +
 ----
@@ -938,13 +940,13 @@ Example output:
 ----
 Default models:
 
-  anthropic    → claude-haiku-4-5
-  deepseek     → deepseek-chat
-  gemini       → gemini-3.1-flash-lite
+  anthropic    → claude-haiku-5-5
+  deepseek     → deepseek-flash
+  gemini       → gemini-3.5-flash-lite
   groq         → openai/gpt-oss-20b
   mistral      → codestral-2508
-  ollama       → llama3.1
-  openai       → gpt-5.4-mini
+  ollama       → gemma4:e2b
+  openai       → gpt-6-luna
   xai          → grok-4.3
 ----
 
@@ -1003,13 +1005,13 @@ Example output:
 ----
 Supported providers:
 
-  anthropic    model: claude-haiku-4-5                    (token required)
-  deepseek     model: deepseek-chat                      (token required)
-  gemini       model: gemini-3.1-flash-lite              (token required)
+  anthropic    model: claude-haiku-5-5                   (token required)
+  deepseek     model: deepseek-flash                     (token required)
+  gemini       model: gemini-3.5-flash-lite              (token required)
   groq         model: openai/gpt-oss-20b                 (token required)
   mistral      model: codestral-2508                     (token required)
-  ollama       model: llama3.1                           (no token required)
-  openai       model: gpt-5.4-mini                       (token required)
+  ollama       model: gemma4:e2b                         (no token required)
+  openai       model: gpt-6-luna                         (token required)
   xai          model: grok-4.3                           (token required)
 
 Custom providers (OpenAI compatible, from config):
@@ -1097,9 +1099,11 @@ Available configuration keys:
 - `full_files` -- send full working-tree contents of staged files alongside
   the diff (`true`/`false`, default `false`)
 - `max_diff_bytes` -- maximum size (in UTF-8 bytes) of the diff/commit-log
-  sent to the LLM. When the input exceeds this, it is truncated and a marker
-  line is appended so the model knows it was cut. `0` (the default) means no
-  limit. Useful to avoid context-window errors on very large staged changes.
+  sent to the LLM. When the input exceeds this, the largest files are shortened
+  first so every file stays represented, and marker lines tell the model what
+  was cut. `0` (the default) means no limit. You rarely need this for very
+  large staged changes: when a provider rejects the input as too large, git-cai
+  shortens it the same way, warns, and retries up to two times.
 - `pr_to_file` -- when running `--PR`, write the generated PR description
   to a Markdown file in the repository root instead of printing it to
   stdout (`true`/`false`, default `false`)
@@ -1141,7 +1145,8 @@ Available configuration keys:
   `anthropic` the default is `32768` (`max_tokens` is still read as a legacy
   alias). For OpenAI compatible providers it is only sent when set.
 - `<provider>.timeout` -- HTTP timeout in seconds for one provider,
-  overriding the global `timeout` (`ollama` defaults to `300`)
+  overriding the global `timeout` (`ollama` defaults to `300`, `gemini` to `60`).
+  The `--timeout` flag overrides both.
 - `stats` -- opt in to local-only usage analytics (`true`/`false`,
   default `false`). When enabled, every generation appends one row to
   `~/.local/share/git-cai/stats.db` capturing kind (`commit`, `amend`,
@@ -1315,7 +1320,7 @@ Change a value in the repo config:
 
 ----
 git cai -S default=anthropic
-git cai -S groq.model=llama-3.3-70b
+git cai -S groq.model=openai/gpt-oss-120b
 git cai -S emoji=false
 ----
 

{git_cai_cli-0.17.2 → git_cai_cli-0.17.4}/docs/man/git-cai.1

@@ -2,12 +2,12 @@
 .\"     Title: git-cai
 .\"    Author: Thorsten Foltz
 .\" Generator: Asciidoctor 2.0.26
-.\"      Date: 2026-10-11
+.\"      Date: 2026-11-06
 .\"    Manual: \ \&
 .\"    Source: \ \&
 .\"  Language: English
 .\"
-.TH "GIT\-CAI" "1" "2026-10-11" "\ \&" "\ \&"
+.TH "GIT\-CAI" "1" "2026-11-06" "\ \&" "\ \&"
 .ie \n(.g .ds Aq \(aq
 .el       .ds Aq '
 .ss \n[.ss] 0
@@ -656,7 +656,7 @@ provider keys.
 git cai \-H language=de
 git cai \-H default=anthropic
 git cai \-H emoji=false
-git cai \-H groq.model=llama\-3.3\-70b
+git cai \-H groq.model=openai/gpt\-oss\-120b
 .fam
 .fi
 .if n .RE
@@ -789,7 +789,7 @@ identifier for that provider.
 .nf
 .fam C
 git cai \-P openai \-m gpt\-4o
-git cai \-P anthropic \-m claude\-haiku\-4\-5
+git cai \-P anthropic \-m claude\-haiku\-5\-5
 git cai \-P ollama \-m mistral
 .fam
 .fi
@@ -890,7 +890,7 @@ both.
 .nf
 .fam C
 git cai \-P anthropic
-git cai \-P groq \-m llama\-3.3\-70b
+git cai \-P groq \-m openai/gpt\-oss\-120b
 git cai \-P openrouter
 .fam
 .fi
@@ -1066,7 +1066,7 @@ become integers or floats, everything else stays a string.
 .fam C
 git cai \-S default=anthropic
 git cai \-S emoji=false
-git cai \-S groq.model=llama\-3.3\-70b
+git cai \-S groq.model=openai/gpt\-oss\-120b
 git cai \-S openai.temperature=0.7
 git cai \-S conventional=true
 .fam
@@ -1109,9 +1109,10 @@ git cai \-s a1b2c3d
 \-T, \-\-timeout SECONDS
 .RS 4
 Override the HTTP timeout for the LLM call in this invocation. The default
-is 30 seconds for remote providers and 300 seconds for Ollama (since local
-generation can be significantly slower). The override applies uniformly to
-whichever provider is selected for this run.
+is 30 seconds for remote providers, 60 seconds for Gemini (Flash Lite can queue
+for up to 30 seconds) and 300 seconds for Ollama (since local generation can be
+significantly slower). The override applies uniformly to whichever provider is
+selected for this run, including those with their own \f(CR<provider>.timeout\fP.
 .sp
 To change the default permanently, set \f(CRtimeout: <seconds>\fP in
 \f(CRcai_config.yml\fP or the home config:
@@ -1125,13 +1126,14 @@ git cai \-S timeout=45
 .fi
 .if n .RE
 .sp
-Ollama keeps its own timeout under \f(CRollama.timeout\fP (default 300). Override
-it separately:
+Ollama and Gemini keep their own timeouts under \f(CRollama.timeout\fP (default 300)
+and \f(CRgemini.timeout\fP (default 60). Override them separately:
 .sp
 .if n .RS 4
 .nf
 .fam C
 git cai \-H ollama.timeout=600
+git cai \-H gemini.timeout=90
 .fam
 .fi
 .if n .RE
@@ -1441,13 +1443,13 @@ Example output:
 .fam C
 Default models:
 
-  anthropic\&    → claude\-haiku\-4\-5
-  deepseek\&     → deepseek\-chat
-  gemini\&       → gemini\-3.1\-flash\-lite
+  anthropic\&    → claude\-haiku\-5\-5
+  deepseek\&     → deepseek\-flash
+  gemini\&       → gemini\-3.5\-flash\-lite
   groq\&         → openai/gpt\-oss\-20b
   mistral\&      → codestral\-2508
-  ollama\&       → llama3.1
-  openai\&       → gpt\-5.4\-mini
+  ollama\&       → gemma4:e2b
+  openai\&       → gpt\-6\-luna
   xai\&          → grok\-4.3
 .fam
 .fi
@@ -1530,13 +1532,13 @@ Example output:
 .fam C
 Supported providers:
 
-  anthropic\&    model: claude\-haiku\-4\-5\&                    (token required)
-  deepseek\&     model: deepseek\-chat\&                      (token required)
-  gemini\&       model: gemini\-3.1\-flash\-lite\&              (token required)
+  anthropic\&    model: claude\-haiku\-5\-5\&                   (token required)
+  deepseek\&     model: deepseek\-flash\&                     (token required)
+  gemini\&       model: gemini\-3.5\-flash\-lite\&              (token required)
   groq\&         model: openai/gpt\-oss\-20b\&                 (token required)
   mistral\&      model: codestral\-2508\&                     (token required)
-  ollama\&       model: llama3.1\&                           (no token required)
-  openai\&       model: gpt\-5.4\-mini\&                       (token required)
+  ollama\&       model: gemma4:e2b\&                         (no token required)
+  openai\&       model: gpt\-6\-luna\&                         (token required)
   xai\&          model: grok\-4.3\&                           (token required)
 
 Custom providers (OpenAI compatible, from config):
@@ -1827,9 +1829,11 @@ the diff (\f(CRtrue\fP/\f(CRfalse\fP, default \f(CRfalse\fP)
 .  IP \(bu 2.3
 .\}
 \f(CRmax_diff_bytes\fP \(em maximum size (in UTF\-8 bytes) of the diff/commit\-log
-sent to the LLM. When the input exceeds this, it is truncated and a marker
-line is appended so the model knows it was cut. \f(CR0\fP (the default) means no
-limit. Useful to avoid context\-window errors on very large staged changes.
+sent to the LLM. When the input exceeds this, the largest files are shortened
+first so every file stays represented, and marker lines tell the model what
+was cut. \f(CR0\fP (the default) means no limit. You rarely need this for very
+large staged changes: when a provider rejects the input as too large, git\-cai
+shortens it the same way, warns, and retries up to two times.
 .RE
 .sp
 .RS 4
@@ -2021,7 +2025,8 @@ alias). For OpenAI compatible providers it is only sent when set.
 .  IP \(bu 2.3
 .\}
 \f(CR<provider>.timeout\fP \(em HTTP timeout in seconds for one provider,
-overriding the global \f(CRtimeout\fP (\f(CRollama\fP defaults to \f(CR300\fP)
+overriding the global \f(CRtimeout\fP (\f(CRollama\fP defaults to \f(CR300\fP, \f(CRgemini\fP to \f(CR60\fP).
+The \f(CR\-\-timeout\fP flag overrides both.
 .RE
 .sp
 .RS 4
@@ -2310,7 +2315,7 @@ Change a value in the repo config:
 .nf
 .fam C
 git cai \-S default=anthropic
-git cai \-S groq.model=llama\-3.3\-70b
+git cai \-S groq.model=openai/gpt\-oss\-120b
 git cai \-S emoji=false
 .fam
 .fi

{git_cai_cli-0.17.2 → git_cai_cli-0.17.4}/src/git_cai_cli/_version.py

@@ -18,7 +18,7 @@ version_tuple: tuple[int | str, ...]
 commit_id: str | None
 __commit_id__: str | None
 
-__version__ = version = '0.17.2'
-__version_tuple__ = version_tuple = (0, 17, 2)
+__version__ = version = '0.17.4'
+__version_tuple__ = version_tuple = (0, 17, 4)
 
-__commit_id__ = commit_id = 'ga453e832c'
+__commit_id__ = commit_id = 'gf30a44c67'

{git_cai_cli-0.17.2 → git_cai_cli-0.17.4}/src/git_cai_cli/core/config.py

@@ -35,19 +35,26 @@ TOKENS_FILE = CONFIG_DIR / "tokens.yml"
 
 DEFAULT_CONFIG: dict[str, Any] = {
     "anthropic": {
-        "model": "claude-haiku-4-5",
-        "temperature": 0,
+        "model": "claude-haiku-5-5",
         # ``max_tokens`` is still read as a legacy alias (see llm.py)
         "max_output_tokens": 32768,
     },
-    "openai": {"model": "gpt-5.6-luna"},
-    "deepseek": {"model": "deepseek-chat", "temperature": 0},
-    "gemini": {"model": "gemini-3.1-flash-lite", "temperature": 0},
+    "openai": {"model": "gpt-6-luna"},
+    "deepseek": {"model": "deepseek-flash", "temperature": 0},
+    "gemini": {
+        "model": "gemini-3.5-flash-lite",
+        "temperature": 0,
+        # Flash Lite was measured taking up to 30 seconds server side, even for
+        # a single word reply, so the global 30 second timeout is too tight.
+        "timeout": 60,
+    },
     "groq": {"model": "openai/gpt-oss-20b", "temperature": 0},
     "xai": {"model": "grok-4.3", "temperature": 0},
+    # Not the cheapest on the list: ministral-3b wrapped messages in code
+    # fences and ministral-8b added a warning line, both unfit for a commit.
     "mistral": {"model": "codestral-2508", "temperature": 0},
     "ollama": {
-        "model": "llama3.1",
+        "model": "gemma4:e2b",
         "temperature": 0,
         "timeout": 300,
         "startup_timeout": 8,
@@ -733,6 +740,11 @@ def apply_cli_overrides(
         config["branch_context"] = branch_context
     if timeout_override is not None:
         config["timeout"] = timeout_override
+        # The flag outranks ``<provider>.timeout``, which is checked first when
+        # resolving. Rebuild instead of pop: blocks may be shared with DEFAULT_CONFIG.
+        for name, block in list(config.items()):
+            if isinstance(block, dict) and "timeout" in block:
+                config[name] = {k: v for k, v in block.items() if k != "timeout"}
     if full_files_override is not None:
         config["full_files"] = full_files_override
     if sql_override is not None:

{git_cai_cli-0.17.2 → git_cai_cli-0.17.4}/src/git_cai_cli/core/gitutils.py

@@ -349,26 +349,55 @@ def git_diff_excluding(
     return result.stdout
 
 
+# Where an independently trimmable part of an LLM payload starts: one file's
+# diff, or a block appended after the diff. A stray match inside file content
+# only creates one more part, which is harmless.
+_PAYLOAD_PART_RE = re.compile(
+    r"^(?=diff --git "
+    r"|--- (?:Full file contents|Additional context from the author) ---)",
+    re.MULTILINE,
+)
+
+
+def shorten_diff(diff: str, max_bytes: int) -> str:
+    """Cut ``diff`` to about ``max_bytes`` UTF-8 bytes, largest parts first.
+
+    Parts below their fair share stay whole and the others split what is left,
+    so every file keeps its header and the start of its changes. Text without
+    file headers (a commit log) is one part, which makes this a plain tail cut.
+    Each cut adds a short marker, so the result can slightly exceed the limit.
+    """
+    parts = [part.encode("utf-8") for part in _PAYLOAD_PART_RE.split(diff) if part]
+    remaining = max_bytes
+    shares = [0] * len(parts)
+    by_size = sorted(range(len(parts)), key=lambda i: len(parts[i]))
+    for position, index in enumerate(by_size):
+        shares[index] = min(len(parts[index]), remaining // (len(parts) - position))
+        remaining -= shares[index]
+
+    shortened = []
+    for part, share in zip(parts, shares):
+        # errors="ignore" drops a multibyte character split by the byte slice.
+        text = part[:share].decode("utf-8", errors="ignore")
+        if share < len(part):
+            text += f"\n[... {len(part) - share} bytes omitted ...]\n"
+        shortened.append(text)
+    return "".join(shortened)
+
+
 def truncate_diff(diff: str, max_bytes: int) -> tuple[str, bool]:
-    """Truncate ``diff`` to at most ``max_bytes`` UTF-8 bytes.
+    """Truncate ``diff`` to about ``max_bytes`` UTF-8 bytes.
 
     A ``max_bytes`` of 0 (or negative) means "no limit" and the diff is
     returned unchanged — preserving the default behavior. When truncation
     occurs, a marker line is appended so both the LLM and the user know the
     input was cut. Returns ``(possibly_truncated_diff, was_truncated)``.
     """
-    if max_bytes <= 0:
-        return diff, False
-
-    encoded = diff.encode("utf-8")
-    if len(encoded) <= max_bytes:
+    if max_bytes <= 0 or len(diff.encode("utf-8")) <= max_bytes:
         return diff, False
 
-    # Cut on a UTF-8 safe boundary: decoding with errors="ignore" drops any
-    # partial trailing multibyte sequence left by the byte-level slice.
-    truncated = encoded[:max_bytes].decode("utf-8", errors="ignore")
     marker = f"\n\n[... diff truncated: exceeded max_diff_bytes={max_bytes} ...]"
-    return truncated + marker, True
+    return shorten_diff(diff, max_bytes) + marker, True
 
 
 def apply_diff_limit(text: str, config: dict, *, label: str = "Input") -> str:

{git_cai_cli-0.17.2 → git_cai_cli-0.17.4}/src/git_cai_cli/core/llm.py

@@ -18,7 +18,11 @@ from urllib.parse import urlparse
 
 import requests
 from git_cai_cli.core.config import CONFIG_DIR, is_custom_provider
-from git_cai_cli.core.gitutils import classify_changed_paths, paths_from_diff
+from git_cai_cli.core.gitutils import (
+    classify_changed_paths,
+    paths_from_diff,
+    shorten_diff,
+)
 from git_cai_cli.core.languages import LANGUAGE_MAP
 from git_cai_cli.core.prompts_fallback import (
     HARDCODED_CHANGELOG_PROMPT,
@@ -30,6 +34,7 @@ from git_cai_cli.core.prompts_fallback import (
     HARDCODED_SPLIT_PROMPT,
     HARDCODED_SQUASH_PROMPT,
 )
+from git_cai_cli.core.validate import _extract_api_error_message
 from requests.adapters import HTTPAdapter
 from urllib3.util.retry import Retry
 
@@ -41,10 +46,11 @@ log = logging.getLogger(__name__)
 # still handled — ``_http_post`` retries once without the field — so add
 # an entry only when a warning shows up in practice.
 _TEMPERATURE_UNSUPPORTED_PATTERNS = [
-    re.compile(r"^gpt-5", re.IGNORECASE),  # OpenAI GPT-5 family (incl. mini/nano)
+    re.compile(r"^gpt-[5-9]", re.IGNORECASE),  # OpenAI GPT-5+ (incl. mini/nano/luna)
     re.compile(r"^o[134](-|$)", re.IGNORECASE),  # OpenAI o1/o3/o4 reasoning series
     re.compile(r"opus-4-[7-9]", re.IGNORECASE),  # Anthropic Opus 4.7 / 4.8 / 4.9
     re.compile(r"opus-[5-9]", re.IGNORECASE),  # Anthropic Opus 5+
+    re.compile(r"haiku-[5-9]", re.IGNORECASE),  # Anthropic Haiku 5+
 ]
 
 
@@ -178,6 +184,37 @@ def _http_post(*args, **kwargs):
     return response
 
 
+# Model limits change too often for a table, and the provider's rejection
+# names them, so an oversized request is shortened and retried instead.
+_OVERFLOW_STATUS_CODES = (400, 413, 422)
+_OVERFLOW_RE = re.compile(
+    r"too (?:long|large)|exceed|maximum (?:context|prompt)|reduce the length",
+    re.IGNORECASE,
+)
+# Skips numbers glued to a name or version, like a date in a model name.
+_OVERFLOW_SIZE_RE = re.compile(r"(?<![\w.-])\d[\d,]{2,}\b")
+_OVERFLOW_RETRIES = 2
+
+
+def _overflow_shrink_factor(exc: requests.HTTPError) -> float | None:
+    """Return how much to shrink the input after a "too many tokens" rejection.
+
+    ``None`` means ``exc`` is some other error. Providers word the rejection
+    differently, but the first two numbers are the limit and the actual size,
+    and the smaller one is the limit. Without them, halve the input.
+    """
+    response = exc.response
+    if response is None or response.status_code not in _OVERFLOW_STATUS_CODES:
+        return None
+    message = _extract_api_error_message(response)
+    if not _OVERFLOW_RE.search(message):
+        return None
+    sizes = [int(n.replace(",", "")) for n in _OVERFLOW_SIZE_RE.findall(message)[:2]]
+    if len(sizes) < 2 or sizes[0] == sizes[1]:
+        return 0.5
+    return 0.9 * min(sizes) / max(sizes)
+
+
 def load_prompt_file(
     config_key: str,
     config: Dict[str, Any],
@@ -967,7 +1004,8 @@ class CommitMessageGenerator:
     def _dispatch_generate(self, content: str, system_prompt: str) -> str:
         """
         Route to the right provider. ``system_prompt`` comes from the
-        matching ``_build_*_prompt``; ``content`` is the git output.
+        matching ``_build_*_prompt``; ``content`` is the git output. Input the
+        provider rejects as too large is shortened and retried.
         """
         self._scan_for_secrets(content)
 
@@ -986,12 +1024,30 @@ class CommitMessageGenerator:
 
         log.debug("Using provider '%s' for generation.", provider)
 
-        if openai_compatible:
-            return self.generate_openai_compatible(
-                content, provider, system_prompt_override=system_prompt
-            )
+        def call_provider(text: str) -> str:
+            if openai_compatible:
+                return self.generate_openai_compatible(
+                    text, provider, system_prompt_override=system_prompt
+                )
+            return model_dispatch[provider](text, system_prompt_override=system_prompt)
 
-        return model_dispatch[provider](content, system_prompt_override=system_prompt)
+        sent = content
+        for _ in range(_OVERFLOW_RETRIES):
+            try:
+                return call_provider(sent)
+            except requests.HTTPError as exc:
+                factor = _overflow_shrink_factor(exc)
+                if factor is None:
+                    raise
+            target = int(len(sent.encode("utf-8")) * factor)
+            log.warning(
+                "The input is too large for this model. Retrying with the largest "
+                "files shortened to about %d bytes in total, so the result may "
+                "miss details from the cut parts.",
+                target,
+            )
+            sent = shorten_diff(content, target)
+        return call_provider(sent)
 
     # ---------------------------
     # MODEL CALLS