capt-hook 6.1.0__tar.gz → 6.2.0__tar.gz

{capt_hook-6.1.0 → capt_hook-6.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: capt-hook
-Version: 6.1.0
+Version: 6.2.0
 Summary: Declarative hook framework for Claude Code
 Keywords: claude,claude-code,hooks,llm,agents,guardrails,cli
 Author: Yasyf Mohamedali

{capt_hook-6.1.0 → capt_hook-6.2.0}/captain_hook/__init__.py

@@ -23,6 +23,7 @@ from captain_hook import file, style, util
 from captain_hook.app import hook, on
 from captain_hook.ast_grep import COMMENT_TYPES
 from captain_hook.command import CommandLine, ParsedCommand, Redirect
+from captain_hook.conditions import workflow_opt_values, workflow_script_source
 from captain_hook.context import HookContext
 from captain_hook.contexts import AfterEdit, BeforeEdit, Introduced, PromptContext, apply_contexts
 from captain_hook.durable import DurableSlot, DurableState, DurableStore

{capt_hook-6.1.0 → capt_hook-6.2.0}/captain_hook/conditions.py

@@ -87,6 +87,12 @@ def is_waiting(evt: BaseHookEvent) -> bool:
 
 
 def workflow_script_source(evt: BaseHookEvent) -> str | None:
+    """The pending ``Workflow`` call's script source, or ``None``.
+
+    Returns the inline ``script`` when present, else reads the file at
+    ``scriptPath``. A non-Workflow event, a missing/unreadable path, or a file
+    over ~1 MiB yields ``None``; never raises.
+    """
     if not (call := evt.as_input(WorkflowCall)):
         return None
     if call.script is not None:

{capt_hook-6.1.0 → capt_hook-6.2.0}/captain_hook/packs/general/capt-hook.toml

@@ -1,5 +1,5 @@
 name = "general"
-version = "0.6.0"
+version = "0.7.0"
 description = "Language-agnostic guards: git/command safety, doc & prompt nudges, task & plan discipline, and code-review + docs-freshness gates."
 hooks = "."
 nlp = true

capt_hook-6.2.0/captain_hook/packs/general/models.py

@@ -0,0 +1,378 @@
+from __future__ import annotations
+
+from dataclasses import dataclass
+
+from captain_hook import (
+    Agent,
+    Allow,
+    BaseHookEvent,
+    Block,
+    Event,
+    Input,
+    Rewrite,
+    TaskCall,
+    Tool,
+    ToolInput,
+    Warn,
+    WorkflowScript,
+    hook,
+    llm_gate,
+    llm_nudge,
+    nudge,
+    set_tool_input,
+    workflow_opt_values,
+    workflow_script_source,
+)
+
+
+@dataclass(frozen=True, slots=True)
+class DelegatedSpawn:
+    """Gating context: the pending Agent/Task call's model pin, agent type, and prompt."""
+
+    tag: str = "delegated_spawn"
+    required: bool = True
+
+    def content(self, evt: BaseHookEvent) -> str | None:
+        if (call := evt.as_input(TaskCall)) is None or not call.prompt:
+            return None
+        model = call.model or "(none — inherits the session model, fable)"
+        return f"model: {model}\nagent_type: {call.agent_type or '(default)'}\nprompt:\n{call.prompt}"
+
+
+WORKFLOW_SCRIPT_CAP = 14_000  # below the workflow prose nudge's max_context=16_000, so truncation stays ours
+
+
+@dataclass(frozen=True, slots=True)
+class WorkflowScriptSource:
+    """Gating context: the pending Workflow call's script source, headed by its model pins."""
+
+    tag: str = "workflow_script"
+    required: bool = True
+
+    def content(self, evt: BaseHookEvent) -> str | None:
+        if (source := workflow_script_source(evt)) is None:
+            return None
+        pin_lines = "\n".join(
+            f"  {line.strip()[:200]}" for line in source.splitlines() if workflow_opt_values(line, "model")
+        )
+        if len(source) > WORKFLOW_SCRIPT_CAP:
+            head, tail = WORKFLOW_SCRIPT_CAP * 3 // 4, WORKFLOW_SCRIPT_CAP // 4
+            source = (
+                f"{source[:head]}\n"
+                f"… [script truncated: {len(source):,} chars total; every model pin is quoted above] …\n"
+                f"{source[-tail:]}"
+            )
+        return (
+            "lines that pin a model in this script (a stage not quoted here inherits the "
+            f"session model, fable):\n{pin_lines or '  (none)'}\n\n{source}"
+        )
+
+
+hook(
+    Event.PreToolUse,
+    only_if=[Tool("Agent|Task"), ToolInput("model", r"(?i)\bhaiku\b")],
+    skip_if=[
+        ToolInput(
+            "prompt",
+            r"(?i)\b(classif|label|tag|categoriz|per.file|one (fact|thing)|mechanical|probe|ping"
+            r"|echo|smoke|count|capacit|extract)",
+        )
+    ],
+    message=(
+        "This subagent is pinned to haiku. Route a subagent to haiku only for a single-fact "
+        "mechanical step — classifying, labeling, tagging, counting, or probing one thing per "
+        "item. Anything that carries judgment should run on sonnet; use model='sonnet', or drop "
+        "model entirely to inherit the session model. If this genuinely is a mechanical "
+        "single-fact step, say 'mechanical' in the prompt and retry — the haiku pin will be "
+        "allowed. See CLAUDE.md § Plan Execution & Orchestration (Models)."
+    ),
+    block=True,
+    tests={
+        Input(model="haiku", prompt="implement the retry backoff in the client"): Block(),
+        Input(model="haiku", prompt="classify each file's language"): Allow(),
+        Input(model="haiku", prompt="Probe subagent capacity: spawn and return the word ok"): Allow(),
+        Input(model="haiku", prompt="mechanical step: return the repo's default branch name"): Allow(),
+        Input(model="haiku", prompt="count the TODO markers in src/"): Allow(),
+        Input(prompt="implement the retry backoff in the client"): Allow(),
+        Input(model="sonnet", prompt="implement the retry backoff in the client"): Allow(),
+    },
+)
+
+llm_gate(
+    """Decide whether this delegated subagent call pins a non-fable model on work whose
+deliverable is prose.
+
+<delegated_spawn> holds the pending Agent/Task call: its model pin, agent type, and
+prompt. A regex prefilter already found a haiku/sonnet/opus pin plus a writing verb
+near a prose noun in the prompt; your job is precision: is a prose artifact what this
+subagent is asked to PRODUCE?
+
+The Models rubric: all writing a user reads — READMEs, docs, changelogs, release
+notes, blog posts, PR descriptions, any user-facing text — routes to fable. Work
+that merely mentions a prose file — as a constraint ("do NOT touch the docs"), as
+reading material, as the subject of recon or review — is not prose work.
+
+Set block=true only when the prompt clearly asks the subagent to write, draft,
+revise, or polish a prose artifact. Recon, review, classification, and code work
+that references docs stay allowed: block=false. When uncertain, block=false — a
+wrong block stops legitimate work cold. Keep reasoning under 40 words.
+
+<examples>
+<example block="true">
+model: sonnet — Write the README quickstart for this repo.
+The deliverable is README prose on a non-fable pin.
+</example>
+<example block="true">
+model: opus — Update CHANGELOG.md with an entry for the retry fix.
+The subagent itself writes the changelog prose.
+</example>
+<example block="false">
+model: opus — Fix the failing test in cli.py. Do NOT edit CHANGELOG.md — a sibling owns it.
+CHANGELOG is a constraint, not the deliverable; this is code work.
+</example>
+<example block="false">
+model: sonnet — Explore the cc-interact building blocks and report where the docs pipeline is written.
+Read-only recon that mentions docs; nothing user-facing is produced.
+</example>
+<example block="false">
+model: sonnet — Review the README draft for factual errors and list them.
+Review findings go to the orchestrator; the subagent writes no user-facing prose.
+</example>
+</examples>""",
+    message=lambda r: (
+        "This subagent is pinned to a non-fable model but its deliverable is prose/writing work. "
+        f"{r.reasoning} All writing — docs, READMEs, release notes, any user-facing text — routes "
+        "to fable: drop model to inherit the session model, or pass model='fable'. "
+        "See CLAUDE.md § Plan Execution & Orchestration (Models)."
+    ),
+    contexts=[DelegatedSpawn()],
+    events=Event.PreToolUse,
+    only_if=[
+        Tool("Agent|Task"),
+        ToolInput("model", r"(?i)\b(haiku|sonnet|opus)\b"),
+        ToolInput(
+            "prompt",
+            r"(?i)\b(?:writ(?:e|es|ing|ten)|draft|redraft|rewrit|revis|polish|copyedit|compose|author|update|edit)\w*"
+            r"\W+(?:\w+\W+){0,4}(?:readme|docs?|documentation|blog|changelog|release notes|announcement"
+            r"|tutorial|guide|prose|copywrit|marketing copy|article|newsletter|email"
+            r"|pr description|commit message)\b"
+            r"|\b(?:readme|docs?|documentation|blog|changelog|release notes|announcement"
+            r"|tutorial|guide|prose|copywrit|marketing copy|article|newsletter|email"
+            r"|pr description|commit message)\b"
+            r"\W+(?:\w+\W+){0,4}(?:writ(?:e|es|ing|ten)|draft|redraft|rewrit|revis|polish|copyedit|compose|author|update|edit)",
+        ),
+    ],
+    skip_if=[
+        ToolInput("prompt", r"(?i)\b(classif|label|tag|categoriz|count|extract|mechanical)"),
+        Agent("Explore|claude-code-guide"),
+    ],
+    agent=False,
+    transcript=False,
+    tests={
+        Input(model="sonnet", prompt="Write the README quickstart for this repo"): Block(),
+        Input(model="opus", prompt="draft the release notes for v2"): Block(),
+        Input(model="haiku", prompt="update the CHANGELOG entry for the fix"): Block(),
+        Input(model="fable", prompt="write the README quickstart"): Allow(),
+        Input(prompt="write the README quickstart"): Allow(),
+        Input(model="sonnet", prompt="review the README for factual errors"): Allow(),
+        Input(model="sonnet", prompt="update the retry backoff config"): Allow(),
+        Input(model="haiku", prompt="label each README section with its Diataxis mode"): Allow(),
+        Input(agent_type="Explore", model="sonnet", prompt="find where the README quickstart is written"): Allow(),
+        Input(agent_type="claude-code-guide", model="sonnet", prompt="explain how the docs get updated"): Allow(),
+        Input(model="opus", prompt="Update the retry backoff config per the spec in docs/plan.md"): Allow(),
+        Input(
+            model="opus",
+            prompt="Fix the failing test in cli.py. Do NOT edit the CHANGELOG — a sibling owns updating it",
+            llm={"block": False},
+        ): Allow(),
+    },
+)
+
+set_tool_input(
+    "model",
+    "sonnet",
+    tool="Agent|Task",
+    only_if=[Agent("Explore|claude-code-guide")],
+    note=(
+        "Upgraded this recon subagent from the silent haiku default to sonnet, per the Models "
+        "table in CLAUDE.md § Plan Execution & Orchestration (Models)."
+    ),
+    tests={
+        Input(agent_type="Explore"): Rewrite(model="sonnet"),
+        Input(agent_type="Explore", model="haiku"): Allow(),
+        Input(agent_type="general-purpose"): Allow(),
+    },
+)
+
+llm_nudge(
+    """Decide whether this delegated subagent should run on opus-4.8 instead of fable-5.
+
+<delegated_spawn> holds the pending Agent/Task call: its model pin (or that it inherits
+the session model, fable), agent type, and prompt.
+
+The Models rubric: implementation delegates to opus-4.8 at xhigh — opus is ~2x cheaper
+than fable and nearly as capable. Fable's lanes are orchestration, review, hard
+planning/design/diagnosis, all prose/writing, and implementation that is very sensitive
+or error-prone (auth, migrations, concurrency, data loss, crypto, subtle algorithms).
+
+Set fire=true only when the prompt is clearly routine implementation — building, fixing,
+wiring, or refactoring code — with no fable-lane signal. A prompt that reviews, plans,
+designs, diagnoses a hard bug, writes prose, or touches a sensitive surface stays on
+fable: fire=false. When uncertain, fire=false — the agent may have chosen fable
+deliberately, and a false alarm teaches it to ignore this nudge. Keep reasoning under
+40 words.
+
+<examples>
+<example fire="true">
+Implement the pagination endpoint in api/users.py per the spec in the plan.
+Routine implementation with no sensitivity signal — the opus xhigh lane.
+</example>
+<example fire="true">
+Add a --json flag to the export command and thread it through the formatter.
+Well-scoped feature wiring; the default implementation lane.
+</example>
+<example fire="false">
+Review the diff for correctness and concurrency issues.
+Review is fable's lane.
+</example>
+<example fire="false">
+Design the migration strategy for the sharded session store.
+Hard planning/design stays on fable.
+</example>
+<example fire="false">
+Implement the token-refresh race fix in the auth middleware.
+Auth plus concurrency: sensitive, error-prone implementation stays on fable.
+</example>
+</examples>""",
+    message=lambda r: (
+        f"This delegation would run on fable, but it reads as routine implementation. {r.reasoning} "
+        "Implementation defaults to model='opus' + effort='xhigh' (~2x cheaper, nearly as capable); "
+        "a well-scoped edit to existing code can also go to gpt-5.5 via the codex skill behind a "
+        "model='sonnet' low-effort wrapper. Keep fable if this genuinely is sensitive or error-prone. "
+        "See CLAUDE.md § Plan Execution & Orchestration (Models)."
+    ),
+    contexts=[DelegatedSpawn()],
+    events=Event.PreToolUse,
+    only_if=[Tool("Agent|Task")],
+    skip_if=[
+        ToolInput("model", r"(?i)\b(opus|sonnet|haiku)\b"),
+        Agent("Explore|claude-code-guide"),
+    ],
+    agent=False,
+    transcript=False,
+    tests={
+        Input(prompt="implement the pagination endpoint in api/users.py"): Warn(pattern="opus"),
+        Input(model="fable", prompt="add a --json flag to the export command"): Warn(pattern="opus"),
+        Input(model="opus", prompt="implement the pagination endpoint in api/users.py"): Allow(),
+        Input(model="sonnet", prompt="scan the repo for TODO markers"): Allow(),
+        Input(agent_type="Explore", prompt="find where the config loader lives"): Allow(),
+    },
+)
+
+nudge(
+    """
+    This workflow script pins agent() steps to haiku. Reserve haiku for mechanical single-fact
+    map steps; a judgment-bearing stage should inherit the session model or route up. See
+    CLAUDE.md § Plan Execution & Orchestration (Models).
+    """,
+    only_if=[Tool("Workflow"), WorkflowScript(model="haiku")],
+    events=Event.PreToolUse,
+    max_fires=2,
+    tests={
+        Input(script="steps:\n  - agent: reviewer\n    model: 'haiku'\n"): Warn(),
+        Input(script="steps:\n  - agent: reviewer\n    model: 'sonnet'\n"): Allow(),
+    },
+)
+
+llm_nudge(
+    """Decide whether this workflow script pins a non-fable model on a stage whose
+deliverable is prose.
+
+<workflow_script> holds the pending Workflow call's script source. Its header quotes
+every line of the script that pins a model; a stage that is not quoted there carries
+no pin and inherits the session model, fable — already correctly routed, whatever it
+writes. A regex prefilter already found both a haiku/sonnet/opus pin and a writing
+verb near a prose noun somewhere in the script; your job is precision: does a
+PINNED stage have prose as its own deliverable?
+
+The Models rubric: all writing a user reads — READMEs, docs, changelogs, release
+notes, blog posts, announcements, any user-facing text — routes to fable. A stage's
+deliverable is prose when its agent() prompt asks it to write, draft, revise, or
+polish such an artifact.
+
+Set fire=true only when at least one agent() call both pins haiku/sonnet/opus and
+has prose as its deliverable. A prose stage with no pin of its own is fine, even
+when other stages pin opus. A prose keyword that appears as a constraint ("do NOT
+edit CHANGELOG.md"), an ownership note, a file the stage merely reads, a
+meta.description, or prose the orchestrator script assembles itself outside any
+pinned agent() call is not that stage's deliverable: fire=false. When uncertain,
+fire=false — a false alarm teaches the agent to ignore this nudge. Keep reasoning
+under 40 words and name the offending stage.
+
+<examples>
+<example fire="true">
+agent('Write the README quickstart section for the new CLI', {model: 'opus'})
+The stage's deliverable is README prose, pinned to opus.
+</example>
+<example fire="true">
+stages.push(agent(`Draft the docs-site page for ${feature}`, {model: 'sonnet'}))
+A docs page is user-facing prose; the pin is non-fable.
+</example>
+<example fire="false">
+agent('Fix the failing import in cli.py. Do NOT edit CHANGELOG.md — a sibling owns it', {model: 'opus'})
+CHANGELOG appears only as a constraint; the deliverable is a code fix.
+</example>
+<example fire="false">
+agent('Fix the CLI error handling', {label: 'fix:cli', model: 'opus'}) alongside
+agent('Reword the troubleshooting guide and CHANGELOG bullet', {label: 'fix:docs'})
+The prose stage carries no pin — it inherits fable; the opus pin is on code.
+</example>
+<example fire="false">
+meta: {description: 'verify the doc claims against actual behavior'}, then agent('run the test matrix', {model: 'opus'})
+"doc claims" lives in the description, not in any pinned stage's deliverable.
+</example>
+<example fire="false">
+agent('fix the three failing tests', {model: 'opus'}), then the script itself assembles CHANGELOG.md from the results
+The orchestrator writes the prose; the pinned stage only fixes tests.
+</example>
+<example fire="false">
+agent('Read docs/architecture.md and list the sections that are stale', {model: 'sonnet'})
+Reading and classifying docs is analysis, not a prose deliverable.
+</example>
+</examples>""",
+    message=lambda r: (
+        f"This workflow script pins a non-fable model on a stage whose deliverable is prose. {r.reasoning} "
+        "All writing — docs, READMEs, release notes, any user-facing text — routes to fable: inherit the "
+        "session model or pin model: 'fable' on that stage. "
+        "See CLAUDE.md § Plan Execution & Orchestration (Models)."
+    ),
+    contexts=[WorkflowScriptSource()],
+    events=Event.PreToolUse,
+    only_if=[
+        Tool("Workflow"),
+        WorkflowScript(
+            pattern=r"(?i)\b(?:writ(?:e|es|ing|ten)|draft|redraft|rewrit|revis|polish|copyedit|compose|author|update|edit)\w*"
+            r"\W+(?:\w+\W+){0,4}(?:readme|docs?|documentation|blog|changelog|release notes|prose)\b"
+            r"|\b(?:readme|docs?|documentation|blog|changelog|release notes|prose)\b"
+            r"\W+(?:\w+\W+){0,4}(?:writ(?:e|es|ing|ten)|draft|redraft|rewrit|revis|polish|copyedit|compose|author|update|edit)",
+            model="haiku|sonnet|opus",
+        ),
+    ],
+    max_fires=2,
+    max_context=16_000,
+    agent=False,
+    transcript=False,
+    tests={
+        Input(script="steps:\n  - agent: write the README intro\n    model: 'sonnet'\n"): Warn(pattern="fable"),
+        Input(script="steps:\n  - agent: write the README intro\n    model: 'fable'\n"): Allow(),
+        Input(script="steps:\n  - agent: fix the retry backoff\n    model: 'sonnet'\n"): Allow(),
+        Input(script="agent('Audit docs/architecture.md for stale claims', {model: 'opus'})"): Allow(),
+        Input(
+            script="agent('recon the module map', {model: 'sonnet'})\n// all prose stays with the main agent on fable\n"
+        ): Allow(),
+        Input(
+            script="agent('Fix the import in cli.py. Do NOT edit CHANGELOG.md — a sibling owns it', {model: 'opus'})",
+            llm={"fire": False},
+        ): Allow(),
+    },
+)

{capt_hook-6.1.0 → capt_hook-6.2.0}/captain_hook/skills/authoring-hooks/SKILL.md

@@ -94,7 +94,9 @@ registration gets:
 - Inline `tests = {...}` from Step 1: the offending shape expecting `Block(...)` or
   `Warn(...)` (match the chosen severity), the benign neighbor expecting `Allow()`.
   LLM hooks (`llm_gate`, `llm_nudge`) ship without `tests=` — their inline tests would
-  only exercise a stubbed model.
+  only exercise a stubbed model. Exception: with a required `contexts=` provider, ship
+  `tests=`, and use `Input(llm={"fire": False})` (or `{"block": False}`) to wire-test the
+  judge-declines path; the default stub still always fires.
 
 ### 4. Verify
 

{capt_hook-6.1.0 → capt_hook-6.2.0}/captain_hook/skills/authoring-hooks/references/testing-hooks.md

@@ -13,8 +13,11 @@ hook's conditions and handler, and asserts the outcome. Exit code 1 on any failu
    (first alternative of the pattern: `Tool("Edit|Write")` infers `Edit`), else `Bash`.
 3. `skip_if`/`only_if` are evaluated; a non-matching input yields `Allow()`.
 4. LLM calls are stubbed: `GateVerdict(block=True)`, `NudgeVerdict(fire=True)`,
-   `PromptCheckVerdict(action="block")`. The stub always fires, so inline tests on LLM hooks
-   assert nothing real — ship LLM hooks without `tests=`.
+   `PromptCheckVerdict(action="block")`. The default stub always fires, so inline tests on
+   LLM hooks assert only conditions and wiring, never judgment — ship LLM hooks without
+   `tests=` unless a required `contexts=` provider makes them meaningful.
+   `Input(llm={"fire": False})` / `Input(llm={"block": False})` overrides the stub verdict
+   per test, wire-testing the judge-declines path.
 
 ## Input fields
 
@@ -89,4 +92,4 @@ Statuses: `pass`, `fail` (wrong outcome), `error` (exception in hook or test), `
 | `@on` handler on a piped command never fires | `evt.command_line.q.runs(...)` checks the pipeline's *last* command. Use `.any_command(lambda c: c.program == "...")`. |
 | Edit/Write hook gets `None` content | The `Input` set `command=` but the hook's first event/tool is Edit — set `file=` and `content=` (and `tool=` if the hook matches several). |
 | `Input(command=...)` ignored | The hook's first event is `Stop` — the runner builds a Stop event without tool input. Encode the command as a Bash `tool_use` in `transcript=` instead. |
-| LLM hook test always blocks | The stub verdict always fires. Remove `tests=` from LLM hooks; test their static `only_if` narrowing on a deterministic sibling hook instead. |
+| LLM hook test always blocks | The default stub verdict always fires. Pass `Input(llm={"fire": False})` / `{"block": False}` to exercise the declines path, or remove `tests=` and test the static `only_if` narrowing on a deterministic sibling hook. |

{capt_hook-6.1.0 → capt_hook-6.2.0}/captain_hook/skills/bootstrapping-hooks/SKILL.md

@@ -19,7 +19,7 @@ skill, which owns hook drafting (Step 6 delegates to it).
 ## Hard Rules
 
 - **Never write a hook the user has not approved in Step 4.** Survey and propose first; write only what was selected.
-- **Every deterministic hook ships inline tests** — at least one firing `Input` and one `Allow()` — and `uvx capt-hook test` must be green before Step 8. LLM hooks (`llm_gate`, `llm_nudge`, `prompt_check`) and signal-scored `nudge`s ship without `tests=` — their inline tests would only exercise a stubbed model — with one exception: an LLM hook gated by a **required `contexts=` provider** should ship `tests=`, because the context extraction runs for real under the stub. The stub always confirms, so every firing case exercises the real gate and every `Allow()` case must yield no evidence (fail the gate), never rely on LLM judgment.
+- **Every deterministic hook ships inline tests** — at least one firing `Input` and one `Allow()` — and `uvx capt-hook test` must be green before Step 8. LLM hooks (`llm_gate`, `llm_nudge`, `prompt_check`) and signal-scored `nudge`s ship without `tests=` — their inline tests would only exercise a stubbed model — with one exception: an LLM hook gated by a **required `contexts=` provider** should ship `tests=`, because the context extraction runs for real under the stub. The default stub always confirms, so every firing case exercises the real gate and every `Allow()` case must yield no evidence (fail the gate) — or override the verdict per test with `Input(llm={"fire": False})` / `{"block": False}` to wire-test the judge-declines path. Never treat a stubbed verdict as proof the model judges correctly.
 - **Propose `block` only for irreversible or destructive actions** (history rewrites, data deletion, deploys, secret leaks). Default everything else to warn. The user picks final severity in Step 4.
 - **Never write style rules here.** A style guide found during the survey is delegated whole to the `translating-styleguides` skill (category E below).
 

{capt_hook-6.1.0 → capt_hook-6.2.0}/captain_hook/testing/helpers.py

@@ -4,7 +4,7 @@ import atexit
 import re
 import shutil
 import tempfile
-from collections.abc import Iterator
+from collections.abc import Callable, Iterator
 from contextlib import contextmanager
 from itertools import count
 from pathlib import Path
@@ -508,21 +508,29 @@ def assert_result(
                 assert val in actual, f"{prefix}Rewrite {key}={actual!r} doesn't contain '{val}'"
 
 
-def stub_call_llm(
-    _template: Any,
-    *args: Any,
-    response_model: type[BaseModel] | None = None,
-    **kwargs: Any,
-) -> str | BaseModel:
-    if response_model is None:
-        return "stubbed"
-    return response_model(
-        **{
-            name: STUB_FIELD_VALUES.get(name, "")
-            for name, info in response_model.model_fields.items()
-            if name in STUB_FIELD_VALUES or info.default is None
-        }
-    )
+def make_stub_call_llm(overrides: dict[str, Any] | None = None) -> Callable[..., str | BaseModel]:
+    values = STUB_FIELD_VALUES | (overrides or {})
+
+    def stub_call_llm(
+        _template: Any,
+        *args: Any,
+        response_model: type[BaseModel] | None = None,
+        **kwargs: Any,
+    ) -> str | BaseModel:
+        if response_model is None:
+            return "stubbed"
+        return response_model(
+            **{
+                name: values.get(name, "")
+                for name, info in response_model.model_fields.items()
+                if name in values or info.default is None
+            }
+        )
+
+    return stub_call_llm
+
+
+stub_call_llm = make_stub_call_llm()
 
 
 @contextmanager
@@ -563,7 +571,7 @@ def run_inline_tests() -> list[tuple[str, str, bool, str]]:
                             if spec_tools
                             else None,
                         )
-                        evt.ctx.call_llm = stub_call_llm  # type: ignore[assignment]
+                        evt.ctx.call_llm = make_stub_call_llm(key.llm)  # type: ignore[assignment]
                         hook_result = (
                             execute_hook(entry, evt)
                             if matches_conditions(entry.spec, evt) and not is_planning_agent_skip(entry.spec, evt)

{capt_hook-6.1.0 → capt_hook-6.2.0}/captain_hook/testing/types.py

@@ -110,6 +110,9 @@ class Input:
         transcript: Session history for transcript conditions — a path, a
             ``TranscriptFixture``, or a raw list of transcript-line dicts.
         tasks: The native task list read via ``evt.tasks``.
+        llm: Per-test LLM stub overrides merged over the default stub verdict
+            (``fire``/``block``/``action``/``reasoning``), e.g. ``llm={"fire": False}``
+            to exercise an LLM hook's judge-declines path.
     """
 
     command: str | None = None
@@ -131,6 +134,7 @@ class Input:
     limit: int | None = None
     transcript: Path | TranscriptFixture | list[dict[str, Any]] | None = None
     tasks: list[dict[str, Any]] | None = None
+    llm: dict[str, Any] | None = None
 
     @field_validator("transcript", mode="before")
     @classmethod

{capt_hook-6.1.0 → capt_hook-6.2.0}/pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "capt-hook"
 # Inert sentinel — the real version is stamped from the release tag (uv version --frozen).
-version = "6.1.0"
+version = "6.2.0"
 description = "Declarative hook framework for Claude Code"
 readme = "README.md"
 license = "PolyForm-Noncommercial-1.0.0"

capt_hook-6.1.0/captain_hook/packs/general/models.py

@@ -1,221 +0,0 @@
-from __future__ import annotations
-
-from dataclasses import dataclass
-
-from captain_hook import (
-    Agent,
-    Allow,
-    BaseHookEvent,
-    Block,
-    Event,
-    Input,
-    Rewrite,
-    TaskCall,
-    Tool,
-    ToolInput,
-    Warn,
-    WorkflowScript,
-    hook,
-    llm_nudge,
-    nudge,
-    set_tool_input,
-)
-
-
-@dataclass(frozen=True, slots=True)
-class DelegatedSpawn:
-    """Gating context: the pending Agent/Task call's model pin, agent type, and prompt."""
-
-    tag: str = "delegated_spawn"
-    required: bool = True
-
-    def content(self, evt: BaseHookEvent) -> str | None:
-        if (call := evt.as_input(TaskCall)) is None or not call.prompt:
-            return None
-        model = call.model or "(none — inherits the session model, fable)"
-        return f"model: {model}\nagent_type: {call.agent_type or '(default)'}\nprompt:\n{call.prompt}"
-
-
-hook(
-    Event.PreToolUse,
-    only_if=[Tool("Agent|Task"), ToolInput("model", r"(?i)\bhaiku\b")],
-    skip_if=[
-        ToolInput(
-            "prompt",
-            r"(?i)\b(classif|label|tag|categoriz|per.file|one (fact|thing)|mechanical|probe|ping"
-            r"|echo|smoke|count|capacit|extract)",
-        )
-    ],
-    message=(
-        "This subagent is pinned to haiku. Route a subagent to haiku only for a single-fact "
-        "mechanical step — classifying, labeling, tagging, counting, or probing one thing per "
-        "item. Anything that carries judgment should run on sonnet; use model='sonnet', or drop "
-        "model entirely to inherit the session model. If this genuinely is a mechanical "
-        "single-fact step, say 'mechanical' in the prompt and retry — the haiku pin will be "
-        "allowed. See CLAUDE.md § Plan Execution & Orchestration (Models)."
-    ),
-    block=True,
-    tests={
-        Input(model="haiku", prompt="implement the retry backoff in the client"): Block(),
-        Input(model="haiku", prompt="classify each file's language"): Allow(),
-        Input(model="haiku", prompt="Probe subagent capacity: spawn and return the word ok"): Allow(),
-        Input(model="haiku", prompt="mechanical step: return the repo's default branch name"): Allow(),
-        Input(model="haiku", prompt="count the TODO markers in src/"): Allow(),
-        Input(prompt="implement the retry backoff in the client"): Allow(),
-        Input(model="sonnet", prompt="implement the retry backoff in the client"): Allow(),
-    },
-)
-
-hook(
-    Event.PreToolUse,
-    only_if=[
-        Tool("Agent|Task"),
-        ToolInput("model", r"(?i)\b(haiku|sonnet|opus)\b"),
-        ToolInput(
-            "prompt",
-            r"(?i)\b(writ(e|es|ing|ten)|draft|redraft|rewrit|revis|polish|copyedit|compose|author|update|edit)",
-        ),
-        ToolInput(
-            "prompt",
-            r"(?i)\b(readme|docs?|documentation|blog|changelog|release notes|announcement"
-            r"|tutorial|guide|prose|copywrit|marketing copy|article|newsletter|email"
-            r"|pr description|commit message)\b",
-        ),
-    ],
-    skip_if=[ToolInput("prompt", r"(?i)\b(classif|label|tag|categoriz|count|extract|mechanical)")],
-    message=(
-        "This subagent is pinned to a non-fable model but its prompt is prose/writing work. "
-        "All writing — docs, READMEs, release notes, any user-facing text — routes to fable: "
-        "drop model to inherit the session model, or pass model='fable'. "
-        "See CLAUDE.md § Plan Execution & Orchestration (Models)."
-    ),
-    block=True,
-    tests={
-        Input(model="sonnet", prompt="Write the README quickstart for this repo"): Block(),
-        Input(model="opus", prompt="draft the release notes for v2"): Block(),
-        Input(model="haiku", prompt="update the CHANGELOG entry for the fix"): Block(),
-        Input(model="fable", prompt="write the README quickstart"): Allow(),
-        Input(prompt="write the README quickstart"): Allow(),
-        Input(model="sonnet", prompt="review the README for factual errors"): Allow(),
-        Input(model="sonnet", prompt="update the retry backoff config"): Allow(),
-        Input(model="haiku", prompt="label each README section with its Diataxis mode"): Allow(),
-    },
-)
-
-set_tool_input(
-    "model",
-    "sonnet",
-    tool="Agent|Task",
-    only_if=[Agent("Explore|claude-code-guide")],
-    note=(
-        "Upgraded this recon subagent from the silent haiku default to sonnet, per the Models "
-        "table in CLAUDE.md § Plan Execution & Orchestration (Models)."
-    ),
-    tests={
-        Input(agent_type="Explore"): Rewrite(model="sonnet"),
-        Input(agent_type="Explore", model="haiku"): Allow(),
-        Input(agent_type="general-purpose"): Allow(),
-    },
-)
-
-llm_nudge(
-    """Decide whether this delegated subagent should run on opus-4.8 instead of fable-5.
-
-<delegated_spawn> holds the pending Agent/Task call: its model pin (or that it inherits
-the session model, fable), agent type, and prompt.
-
-The Models rubric: implementation delegates to opus-4.8 at xhigh — opus is ~2x cheaper
-than fable and nearly as capable. Fable's lanes are orchestration, review, hard
-planning/design/diagnosis, all prose/writing, and implementation that is very sensitive
-or error-prone (auth, migrations, concurrency, data loss, crypto, subtle algorithms).
-
-Set fire=true only when the prompt is clearly routine implementation — building, fixing,
-wiring, or refactoring code — with no fable-lane signal. A prompt that reviews, plans,
-designs, diagnoses a hard bug, writes prose, or touches a sensitive surface stays on
-fable: fire=false. When uncertain, fire=false — the agent may have chosen fable
-deliberately, and a false alarm teaches it to ignore this nudge. Keep reasoning under
-40 words.
-
-<examples>
-<example fire="true">
-Implement the pagination endpoint in api/users.py per the spec in the plan.
-Routine implementation with no sensitivity signal — the opus xhigh lane.
-</example>
-<example fire="true">
-Add a --json flag to the export command and thread it through the formatter.
-Well-scoped feature wiring; the default implementation lane.
-</example>
-<example fire="false">
-Review the diff for correctness and concurrency issues.
-Review is fable's lane.
-</example>
-<example fire="false">
-Design the migration strategy for the sharded session store.
-Hard planning/design stays on fable.
-</example>
-<example fire="false">
-Implement the token-refresh race fix in the auth middleware.
-Auth plus concurrency: sensitive, error-prone implementation stays on fable.
-</example>
-</examples>""",
-    message=lambda r: (
-        f"This delegation would run on fable, but it reads as routine implementation. {r.reasoning} "
-        "Implementation defaults to model='opus' + effort='xhigh' (~2x cheaper, nearly as capable); "
-        "a well-scoped edit to existing code can also go to gpt-5.5 via the codex skill behind a "
-        "model='sonnet' low-effort wrapper. Keep fable if this genuinely is sensitive or error-prone. "
-        "See CLAUDE.md § Plan Execution & Orchestration (Models)."
-    ),
-    contexts=[DelegatedSpawn()],
-    events=Event.PreToolUse,
-    only_if=[Tool("Agent|Task")],
-    skip_if=[
-        ToolInput("model", r"(?i)\b(opus|sonnet|haiku)\b"),
-        Agent("Explore|claude-code-guide"),
-    ],
-    agent=False,
-    transcript=False,
-    tests={
-        Input(prompt="implement the pagination endpoint in api/users.py"): Warn(pattern="opus"),
-        Input(model="fable", prompt="add a --json flag to the export command"): Warn(pattern="opus"),
-        Input(model="opus", prompt="implement the pagination endpoint in api/users.py"): Allow(),
-        Input(model="sonnet", prompt="scan the repo for TODO markers"): Allow(),
-        Input(agent_type="Explore", prompt="find where the config loader lives"): Allow(),
-    },
-)
-
-nudge(
-    """
-    This workflow script pins agent() steps to haiku. Reserve haiku for mechanical single-fact
-    map steps; a judgment-bearing stage should inherit the session model or route up. See
-    CLAUDE.md § Plan Execution & Orchestration (Models).
-    """,
-    only_if=[Tool("Workflow"), WorkflowScript(model="haiku")],
-    events=Event.PreToolUse,
-    max_fires=2,
-    tests={
-        Input(script="steps:\n  - agent: reviewer\n    model: 'haiku'\n"): Warn(),
-        Input(script="steps:\n  - agent: reviewer\n    model: 'sonnet'\n"): Allow(),
-    },
-)
-
-nudge(
-    """
-    This workflow script pins a non-fable model on what looks like prose/writing stages. All
-    writing — docs, READMEs, release notes, any user-facing text — routes to fable: inherit the
-    session model or pin model: 'fable'. See CLAUDE.md § Plan Execution & Orchestration (Models).
-    """,
-    only_if=[
-        Tool("Workflow"),
-        WorkflowScript(
-            pattern=r"(?i)\b(readme|docs?|documentation|blog|changelog|release notes|prose)\b",
-            model="haiku|sonnet|opus",
-        ),
-    ],
-    events=Event.PreToolUse,
-    max_fires=2,
-    tests={
-        Input(script="steps:\n  - agent: write the README intro\n    model: 'sonnet'\n"): Warn(),
-        Input(script="steps:\n  - agent: write the README intro\n    model: 'fable'\n"): Allow(),
-        Input(script="steps:\n  - agent: fix the retry backoff\n    model: 'sonnet'\n"): Allow(),
-    },
-)