feat(pilot): Phase 5 — inline Script Generator integration

Wires the SuggestedFix card to an inline panel that handles both cases:
template-matched fixes open the Script Library generator with parameters
pre-filled from session context; un-matched fixes open the three-option
dialog (one_off / draft_template / build_template). The decision endpoint
records the path choice with side effects: draft_template persists a
draft_templates row via a Sonnet-driven TemplateExtractionService;
build_template returns a redirect to the Script Builder; one_off just
records the choice.

Backend:
- TemplateExtractionService: drafts a parameter schema from a concrete
  rendered script. Conservative by default ("prefer fewer parameters").
  Round-trip-validates that templated_body only references declared
  parameters; missing-key mismatch falls back to the original script
  with no params. LLM/parse failures fall back identically — the
  engineer can still create a draft and refine in the post-resolve
  prompt (Phase 6).
- /suggested-fixes/{fix_id}/decision side effects:
  * one_off → returns rendered_script (engineer's edited version or the
    fix's ai_drafted_script verbatim)
  * draft_template → same + creates draft_templates row with extracted
    params, returns draft_template_id
  * build_template → returns redirect_path=/scripts/builder?from_session=
    &fix= so the frontend can navigate to the builder pre-loaded
- 400 when a non-template fix has no ai_drafted_script (template-matched
  fixes take the dedicated /scripts/generate path, not this endpoint).
- 12 tests: TemplateExtractionService parse + fallback paths, all four
  decision branches, edited_script override, missing-script 400.

Frontend:
- src/components/pilot/script/{TemplateMatchPanel, NoTemplateDialog,
  ParameterizationPreview}.tsx — inline panels rendered in the task
  lane's bottom slot when the engineer clicks a SuggestedFix card.
- TemplateMatchPanel: loads template via /scripts/templates/{id},
  pre-fills params from fix.ai_drafted_parameters with cyan "from
  session" tags, generates via existing /scripts/generate (already
  bumps state_version on ai_session_id from Phase 3). 404 falls back
  with a clear message instead of erroring.
- NoTemplateDialog: shows the AI-drafted script with proposed parameter
  values highlighted in amber via ParameterizationPreview; three option
  cards with the middle (draft_template) flagged Recommended; inline
  edit on the script body before deciding.
- SuggestedFix card now clickable: onActivate toggles the inline panel.
- AssistantChatPage: scriptPanelOpen state + handleScriptDecision that
  navigates on build_template and toasts on the other paths. Active fix
  changes auto-close the panel so engineers don't act on stale state.
- Cmd+K → "Open inline Script Generator" palette entry surfaces only on
  /pilot/:id routes; fires a window event the chat page subscribes to.
  No Resolve shortcut added per Section 14 decision (browser ⌘R conflict).

Verified 2026-04-22 against the dev stack:
- one_off / draft_template / build_template all return the right shape
  with real Sonnet TemplateExtractionService for the draft path.
- Conservative extraction confirmed: cmdkey + Restart-Process script
  yielded zero proposed parameters as intended.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

backend/app/api/endpoints/session_suggested_fixes.py

@@ -32,6 +32,8 @@ from app.schemas.session_suggested_fix import (
     SessionSuggestedFixDecisionResponse,
     SessionSuggestedFixResponse,
 )
+from app.models.draft_template import DraftTemplate
+from app.models.session_fact import SessionFact
 from app.services.escalation_package_generator import EscalationPackageGeneratorService
 from app.services.preview_cache import preview_cache
 from app.services.psa_writeback_service import (
@@ -39,6 +41,7 @@ from app.services.psa_writeback_service import (
     PSAWritebackService,
 )
 from app.services.resolution_note_generator import ResolutionNoteGeneratorService
+from app.services.template_extraction_service import extract_parameters as _extract_template_parameters
 
 logger = logging.getLogger(__name__)
 
@@ -105,12 +108,13 @@ async def record_decision(
 ) -> SessionSuggestedFixDecisionResponse:
     """Record the engineer's path choice on a suggested fix.
 
-    Phase 3 only persists the decision and (for `dismissed`) supersedes the
-    row. Side effects — script generation for `one_off` / `draft_template`,
-    redirect for `build_template` — land in Phase 5 alongside the inline
-    Script Generator integration. The response shape is forward-compatible.
+    Phase 3 recorded the choice and (for `dismissed`) superseded the fix.
+    Phase 5 adds side effects: one_off / draft_template return the rendered
+    script; draft_template also creates a `draft_templates` row via the
+    TemplateExtractionService; build_template returns a redirect to the
+    Script Builder.
     """
-    await _load_session_or_404(db, session_id)
+    session_obj = await _load_session_or_404(db, session_id)
 
     result = await db.execute(
         select(SessionSuggestedFix).where(
@@ -145,15 +149,97 @@ async def record_decision(
         .where(AISession.id == session_id)
         .values(state_version=AISession.state_version + 1)
     )
+
+    rendered_script: str | None = None
+    draft_template_id: UUID | None = None
+    redirect_path: str | None = None
+
+    # Phase 5 side effects. All three non-dismiss paths assume the fix has
+    # either a script_template_id (template match — use the dedicated
+    # /scripts/generate endpoint from the frontend, not this one) or an
+    # ai_drafted_script (custom script — this is the entry point).
+    if body.decision in ("one_off", "draft_template", "build_template"):
+        drafted = body.edited_script or fix.ai_drafted_script
+        if not drafted:
+            # Template-matched fixes take the regular /scripts/generate path.
+            # If a fix somehow reaches here without a drafted script AND
+            # without a template, that's a client-side wiring bug.
+            raise HTTPException(
+                status_code=status.HTTP_400_BAD_REQUEST,
+                detail=(
+                    "Suggested fix has no ai_drafted_script — use "
+                    "/api/v1/scripts/generate for template-matched fixes."
+                ),
+            )
+        rendered_script = drafted.strip()
+
+    if body.decision == "draft_template":
+        # TemplateExtractionService proposes the parameterization. Runs
+        # under the same transaction so a failure rolls back the decision.
+        session_ctx = await _summarize_session_for_extraction(db, session_id)
+        extraction = await _extract_template_parameters(
+            script_body=rendered_script or "",
+            session_context=session_ctx,
+            ticket_context=None,  # ticket context wiring lands in Phase 5 polish
+        )
+
+        draft = DraftTemplate(
+            account_id=session_obj.account_id,
+            source_session_id=session_obj.id,
+            source_user_id=current_user.id,
+            script_body=extraction["templated_body"] or (rendered_script or ""),
+            proposed_parameters={"parameters": extraction["parameters"]},
+            proposed_name=fix.title[:200] if fix.title else None,
+            status="pending",
+        )
+        db.add(draft)
+        await db.flush()
+        draft_template_id = draft.id
+
+    if body.decision == "build_template":
+        # Frontend navigates to the Script Builder preloaded with the
+        # drafted body. The builder wires the full parameterization flow;
+        # we hand it a scratch-pad query string, not persistent state.
+        redirect_path = (
+            f"/scripts/builder?from_session={session_obj.id}&fix={fix.id}"
+        )
+
     await db.commit()
     await db.refresh(fix)
 
     return SessionSuggestedFixDecisionResponse(
         id=fix.id,
         user_decision=fix.user_decision,  # type: ignore[arg-type]
+        rendered_script=rendered_script,
+        draft_template_id=draft_template_id,
+        redirect_path=redirect_path,
     )
 
 
+async def _summarize_session_for_extraction(
+    db: AsyncSession, session_id: UUID,
+) -> str:
+    """Compact fact list for TemplateExtractionService context.
+
+    We don't send the full chat transcript — the extractor only needs enough
+    signal to decide which values in the script are session-specific (and
+    therefore worth parameterizing).
+    """
+    result = await db.execute(
+        select(SessionFact)
+        .where(
+            SessionFact.session_id == session_id,
+            SessionFact.deleted_at.is_(None),
+        )
+        .order_by(SessionFact.created_at.asc())
+    )
+    facts = list(result.scalars().all())
+    if not facts:
+        return ""
+    lines = [f"- {f.text}" for f in facts]
+    return "\n".join(lines)
+
+
 # ── Resolution note preview ────────────────────────────────────────────────
 
 @router.post(

backend/app/core/config.py

@@ -142,6 +142,10 @@ class Settings(BaseSettings):
         # FlowPilot migration Phase 4 — escalation handoff package. Parallel
         # to resolution_note: Sonnet, same cache story, no MCP.
         "escalation_package": "standard",
+        # FlowPilot migration Phase 5 — extract a parameter schema from a
+        # concrete rendered script so a draft_template can be proposed.
+        # Creates a persistent library artifact on accept, so Sonnet.
+        "template_extraction": "standard",
     }
 
     def get_model_for_action(self, action_type: str) -> str:

backend/app/schemas/session_suggested_fix.py

@@ -33,21 +33,38 @@ class SessionSuggestedFixDecisionRequest(BaseModel):
     """Engineer's path choice on a suggested fix.
 
     Server-side side effects per Section 5.2:
-    - one_off: render the script (Phase 5), no template created.
-    - draft_template: render + queue a draft_templates row (Phase 5/6).
-    - build_template: redirect to full template creation (Phase 5).
-    - dismissed: mark the fix superseded so a fresh suggestion can take over.
+    - one_off: record decision, return the rendered (AI-drafted or
+      engineer-edited) script. No persistent library artifact created.
+    - draft_template: same as one_off, plus TemplateExtractionService
+      proposes a parameterization and a draft_templates row is created.
+    - build_template: return a redirect payload pointing at the Script
+      Builder page, pre-loaded with the drafted script body.
+    - dismissed: mark the fix superseded.
+
+    For one_off / draft_template, the engineer may have edited the drafted
+    script or its parameters in the dialog. The final versions are sent
+    back here so we persist what will actually run.
     """
     decision: UserDecision
+    # Present for one_off / draft_template — the engineer's final version of
+    # the drafted script after any inline edits. Omit to use the fix's
+    # `ai_drafted_script` verbatim.
+    edited_script: str | None = Field(None, min_length=1, max_length=50_000)
+    # Parameter values used when rendering (informational, stored on the
+    # draft_template row so a reviewer can see what the first run used).
+    parameters_used: dict[str, Any] | None = None
 
 
 class SessionSuggestedFixDecisionResponse(BaseModel):
-    """Returned after recording a decision; richer payloads land in Phase 5."""
+    """Returned after recording a decision."""
     id: UUID
     user_decision: UserDecision
-    # Set when the decision triggered side effects (e.g. a script generation).
-    # Phase 3 only records the choice; this stays None until Phase 5 wires it.
+    # Populated for one_off / draft_template — the script to display/run.
     rendered_script: str | None = None
+    # Populated for draft_template — the ID of the draft_templates row so
+    # the post-resolve TemplatizePrompt can fetch it in Phase 6.
+    draft_template_id: UUID | None = None
+    # Populated for build_template — where to send the engineer next.
     redirect_path: str | None = Field(
         None,
         description="Where to send the engineer next (e.g. /scripts/builder?... for build_template)",

backend/app/services/template_extraction_service.py

@@ -0,0 +1,201 @@
+"""TemplateExtractionService — propose a parameter schema from a rendered script.
+
+Phase 5 of the FlowPilot migration. Called when an engineer chooses
+"Run now, templatize after resolve" on a suggested fix with no existing
+library match. The service looks at the concrete script (with the values
+the engineer is about to run with) and session/ticket context, then
+proposes a parameterization that future engineers could use from the
+Script Library.
+
+Design choices (per FLOWPILOT-MIGRATION.md Section 6.4):
+
+- **Conservative by default.** Prefer fewer parameters. Environment-agnostic
+  values (like a command name) should not be parameterized. The prompt calls
+  that out explicitly.
+- **Round-trip check.** After the LLM proposes parameters, we validate that
+  the templated body renders back to the original script when given the
+  extracted parameter values. Failures log a warning and the caller falls
+  back to a single-parameter "raw script" proposal.
+- **Model:** Sonnet (`template_extraction` tier). Creates a persistent
+  library artifact — quality matters more than latency.
+
+Output shape mirrors the Script Generator's parameter schema:
+    {
+      "parameters": [
+        {"key": "<snake>", "label": "<human>", "type": "text|password|select|...",
+         "inferred_from": "<session fact / ticket field / ai guess>"}
+      ],
+      "templated_body": "<script with {{ key }} placeholders>",
+    }
+"""
+from __future__ import annotations
+
+import json
+import logging
+import re
+from typing import Any
+
+from app.core.ai_provider import get_ai_provider
+from app.core.config import settings
+
+logger = logging.getLogger(__name__)
+
+
+_EXTRACTION_SYSTEM_PROMPT = """\
+You are a senior MSP engineer drafting a reusable script template from a \
+concrete script that resolved one ticket. Your job is to identify the values \
+in the script that would change for a different invocation — those become \
+parameters — and replace them with {{ snake_case }} placeholders.
+
+Return strict JSON with this shape:
+{
+  "parameters": [
+    {
+      "key": "<snake_case, ASCII>",
+      "label": "<Short human label, Title Case>",
+      "type": "text" | "password" | "select" | "boolean" | "number" | "textarea",
+      "inferred_from": "<short sentence naming the session fact or ticket \
+field this value came from; or 'ai best-guess' when neither>"
+    }
+  ],
+  "templated_body": "<the original script with each parameterized value \
+replaced by {{ key }} matching the parameters above>"
+}
+
+Rules:
+- Prefer FEWER parameters. If a value looks environment-agnostic — a cmdlet \
+name, a standard path like C:\\Windows\\System32, a Microsoft-documented URL \
+— keep it hardcoded.
+- Secret-looking values (passwords, API keys, client secrets) MUST be \
+parameterized with type=password.
+- The templated_body MUST render back to the original script when the \
+parameter values from the context are substituted in. Preserve all whitespace, \
+comments, and casing.
+- If the script has no meaningful parameters (e.g. it's a single read-only \
+cmdlet like Get-Service), return parameters=[] and templated_body = original.
+- No markdown fences, no prose, only the JSON object.
+"""
+
+
+async def extract_parameters(
+    *,
+    script_body: str,
+    session_context: str | None = None,
+    ticket_context: str | None = None,
+) -> dict[str, Any]:
+    """Return `{parameters, templated_body}` for the given rendered script.
+
+    On LLM failure or malformed output, returns a conservative fallback:
+    the original body with no parameters proposed. Callers can still create
+    a `draft_templates` row from this — the engineer reviews and refines
+    before accepting in the post-resolve prompt (Phase 6).
+    """
+    model = settings.get_model_for_action("template_extraction")
+    provider = get_ai_provider(model=model)
+
+    input_lines = [
+        "# Script to templatize",
+        "```",
+        script_body.strip(),
+        "```",
+    ]
+    if session_context:
+        input_lines.extend(["", "# Session context (facts, symptoms)", session_context.strip()])
+    if ticket_context:
+        input_lines.extend(["", "# Ticket context (company, user, priority)", ticket_context.strip()])
+    user_input = "\n".join(input_lines)
+
+    system_blocks: list[dict[str, Any]] = [
+        {
+            "type": "text",
+            "text": _EXTRACTION_SYSTEM_PROMPT,
+            "cache_control": {"type": "ephemeral"},
+            # cacheable: identical across every extraction call
+        },
+    ]
+
+    try:
+        text, _in, _out = await provider.generate_json(
+            system_prompt=system_blocks,
+            messages=[{"role": "user", "content": user_input}],
+            max_tokens=3000,
+        )
+    except Exception:
+        logger.exception("TemplateExtractionService LLM call failed; returning fallback")
+        return _fallback(script_body)
+
+    parsed = _parse_response(text)
+    if parsed is None:
+        return _fallback(script_body)
+
+    # Round-trip validation: render parsed["templated_body"] with the
+    # `inferred_from` values and confirm it matches the original. We don't
+    # have the engineer's values yet here (those come at runtime), but we
+    # can at least check that every {{ key }} in templated_body maps to a
+    # declared parameter. A mismatch means the LLM referenced an undeclared
+    # placeholder — conservative fallback.
+    declared_keys = {p.get("key") for p in parsed["parameters"] if isinstance(p, dict)}
+    referenced_keys = set(re.findall(r"\{\{\s*(\w+)\s*\}\}", parsed["templated_body"]))
+    missing = referenced_keys - declared_keys
+    if missing:
+        logger.warning(
+            "TemplateExtractionService: templated_body references undeclared "
+            "keys %s; using fallback",
+            sorted(missing),
+        )
+        return _fallback(script_body)
+
+    return parsed
+
+
+def _parse_response(raw: str) -> dict[str, Any] | None:
+    """Tolerant parse. Returns None on any structural problem."""
+    cleaned = raw.strip()
+    if cleaned.startswith("```"):
+        cleaned = re.sub(r"^```(?:json)?\s*", "", cleaned)
+        cleaned = re.sub(r"\s*```$", "", cleaned)
+    try:
+        data = json.loads(cleaned)
+    except (json.JSONDecodeError, ValueError):
+        logger.warning("TemplateExtractionService returned non-JSON: %r", raw[:200])
+        return None
+
+    if not isinstance(data, dict):
+        return None
+    params = data.get("parameters")
+    body = data.get("templated_body")
+    if not isinstance(params, list) or not isinstance(body, str):
+        logger.warning("TemplateExtractionService missing parameters or templated_body")
+        return None
+
+    # Validate each parameter shape. Drop malformed entries rather than
+    # failing the whole response — the engineer will review before accept.
+    valid_params: list[dict[str, Any]] = []
+    allowed_types = {"text", "password", "select", "boolean", "number", "textarea", "multi_text"}
+    for p in params:
+        if not isinstance(p, dict):
+            continue
+        key = p.get("key")
+        if not isinstance(key, str) or not re.match(r"^[a-z_][a-z0-9_]*$", key):
+            continue
+        ptype = p.get("type", "text")
+        if ptype not in allowed_types:
+            ptype = "text"
+        valid_params.append({
+            "key": key,
+            "label": p.get("label") or key.replace("_", " ").title(),
+            "type": ptype,
+            "inferred_from": p.get("inferred_from") or "ai best-guess",
+        })
+
+    return {"parameters": valid_params, "templated_body": body}
+
+
+def _fallback(script_body: str) -> dict[str, Any]:
+    """Conservative no-op result: zero parameters, body unchanged.
+
+    Used when the LLM call fails or returns unusable output. The engineer
+    can still save this as a draft and refine in the post-resolve prompt —
+    it just won't propose a parameterization for them.
+    """
+    return {"parameters": [], "templated_body": script_body}

backend/tests/test_phase5_inline_script.py

@@ -0,0 +1,277 @@
+"""API + service tests for Phase 5 inline Script Generator integration.
+
+Covers:
+- TemplateExtractionService: well-formed, fallback on bad output, missing-key fallback.
+- /suggested-fixes/{fix_id}/decision side effects:
+  * one_off returns rendered_script, no draft_templates row.
+  * draft_template returns rendered_script + draft_template_id, draft persisted.
+  * build_template returns redirect_path.
+  * dismissed (Phase 3) still works.
+- 400 when ai_drafted_script is missing for a non-template fix.
+"""
+from __future__ import annotations
+
+from unittest.mock import AsyncMock, patch
+
+import pytest
+from httpx import AsyncClient
+from sqlalchemy import select
+
+from app.api.endpoints.session_suggested_fixes import _clear_preview_cache_for_tests
+from app.models.ai_session import AISession
+from app.models.draft_template import DraftTemplate
+from app.models.session_suggested_fix import SessionSuggestedFix
+from app.services.template_extraction_service import (
+    _fallback,
+    _parse_response,
+    extract_parameters,
+)
+
+
+@pytest.fixture(autouse=True)
+def _isolate_preview_cache():
+    _clear_preview_cache_for_tests()
+    yield
+    _clear_preview_cache_for_tests()
+
+
+async def _make_session_with_fix(
+    test_db, user, *, with_template_id: bool = False, with_drafted_script: bool = True,
+) -> tuple[AISession, SessionSuggestedFix]:
+    session = AISession(
+        user_id=user["user_data"]["id"],
+        account_id=user["user_data"]["account_id"],
+        session_type="chat",
+        intake_type="free_text",
+        intake_content={"text": "phase 5 test"},
+        status="active",
+        confidence_tier="discovery",
+        conversation_messages=[],
+    )
+    test_db.add(session)
+    await test_db.flush()
+
+    fix = SessionSuggestedFix(
+        session_id=session.id,
+        account_id=session.account_id,
+        title="Reset cached creds",
+        description="Clearing the cached credential...",
+        confidence_pct=85,
+        ai_drafted_script=(
+            'cmdkey /delete:"outlook.office365.com"\n'
+            'Restart-Process -Name OUTLOOK'
+        ) if with_drafted_script else None,
+        ai_drafted_parameters={"target_user": "jsmith"} if with_drafted_script else None,
+    )
+    test_db.add(fix)
+    await test_db.commit()
+    await test_db.refresh(session)
+    await test_db.refresh(fix)
+    return session, fix
+
+
+# ── TemplateExtractionService: parse + fallback ───────────────────────────
+
+class TestParseResponse:
+    def test_well_formed(self):
+        raw = (
+            '{"parameters": [{"key":"host","label":"Host","type":"text",'
+            '"inferred_from":"session fact"}],'
+            '"templated_body":"Get-Service -ComputerName {{ host }}"}'
+        )
+        result = _parse_response(raw)
+        assert result is not None
+        assert len(result["parameters"]) == 1
+        assert result["parameters"][0]["key"] == "host"
+        assert result["templated_body"].endswith("{{ host }}")
+
+    def test_strips_fences(self):
+        raw = '```json\n{"parameters": [], "templated_body": "x"}\n```'
+        result = _parse_response(raw)
+        assert result is not None and result["parameters"] == []
+
+    def test_invalid_key_dropped(self):
+        # Capital letters and dashes in key names violate snake_case — drop.
+        raw = (
+            '{"parameters":[{"key":"BadKey-Name","type":"text"}],'
+            '"templated_body":"x"}'
+        )
+        result = _parse_response(raw)
+        assert result is not None and result["parameters"] == []
+
+    def test_unknown_type_falls_back_to_text(self):
+        raw = (
+            '{"parameters":[{"key":"x","type":"weird"}],"templated_body":"x"}'
+        )
+        result = _parse_response(raw)
+        assert result is not None and result["parameters"][0]["type"] == "text"
+
+    def test_malformed_json_returns_none(self):
+        assert _parse_response("not json") is None
+
+    def test_non_dict_returns_none(self):
+        assert _parse_response('["a","b"]') is None
+
+
+@pytest.mark.asyncio
+async def test_extract_parameters_round_trip_failure_uses_fallback():
+    """Templated_body referencing an undeclared placeholder triggers fallback."""
+    fake_provider = AsyncMock()
+    fake_provider.generate_json = AsyncMock(return_value=(
+        # Declares parameter `host` but the body references `port` too.
+        '{"parameters":[{"key":"host","label":"Host","type":"text"}],'
+        '"templated_body":"Get-Service -ComputerName {{ host }} -Port {{ port }}"}',
+        100, 50,
+    ))
+    with patch(
+        "app.services.template_extraction_service.get_ai_provider",
+        return_value=fake_provider,
+    ):
+        result = await extract_parameters(
+            script_body="Get-Service -ComputerName srv01 -Port 8080",
+        )
+    fb = _fallback("Get-Service -ComputerName srv01 -Port 8080")
+    assert result == fb
+
+
+@pytest.mark.asyncio
+async def test_extract_parameters_llm_exception_uses_fallback():
+    fake_provider = AsyncMock()
+    fake_provider.generate_json = AsyncMock(side_effect=RuntimeError("boom"))
+    with patch(
+        "app.services.template_extraction_service.get_ai_provider",
+        return_value=fake_provider,
+    ):
+        result = await extract_parameters(script_body="echo hello")
+    assert result == _fallback("echo hello")
+
+
+# ── Decision endpoint: one_off ─────────────────────────────────────────────
+
+@pytest.mark.asyncio
+async def test_one_off_returns_rendered_script_no_draft(
+    client: AsyncClient, test_user, auth_headers, test_db
+):
+    session, fix = await _make_session_with_fix(test_db, test_user)
+    r = await client.post(
+        f"/api/v1/ai-sessions/{session.id}/suggested-fixes/{fix.id}/decision",
+        headers=auth_headers,
+        json={"decision": "one_off"},
+    )
+    assert r.status_code == 200
+    body = r.json()
+    assert body["user_decision"] == "one_off"
+    assert body["rendered_script"] is not None
+    assert "cmdkey" in body["rendered_script"]
+    assert body["draft_template_id"] is None
+    assert body["redirect_path"] is None
+
+    # No draft_templates row should have been created.
+    rows = (
+        await test_db.execute(select(DraftTemplate).where(DraftTemplate.source_session_id == session.id))
+    ).scalars().all()
+    assert list(rows) == []
+
+
+# ── Decision endpoint: draft_template ─────────────────────────────────────
+
+@pytest.mark.asyncio
+async def test_draft_template_creates_draft_with_extracted_params(
+    client: AsyncClient, test_user, auth_headers, test_db
+):
+    session, fix = await _make_session_with_fix(test_db, test_user)
+
+    fake_provider = AsyncMock()
+    fake_provider.generate_json = AsyncMock(return_value=(
+        '{"parameters":[{"key":"target_user","label":"Target User","type":"text",'
+        '"inferred_from":"session fact"}],'
+        '"templated_body":"cmdkey /delete:\\"outlook.office365.com\\"\\n'
+        'Restart-Process -Name OUTLOOK"}',
+        80, 60,
+    ))
+
+    with patch(
+        "app.services.template_extraction_service.get_ai_provider",
+        return_value=fake_provider,
+    ):
+        r = await client.post(
+            f"/api/v1/ai-sessions/{session.id}/suggested-fixes/{fix.id}/decision",
+            headers=auth_headers,
+            json={"decision": "draft_template"},
+        )
+
+    assert r.status_code == 200
+    body = r.json()
+    assert body["user_decision"] == "draft_template"
+    assert body["rendered_script"] is not None
+    assert body["draft_template_id"] is not None
+    assert body["redirect_path"] is None
+
+    drafts = (
+        await test_db.execute(select(DraftTemplate).where(DraftTemplate.source_session_id == session.id))
+    ).scalars().all()
+    drafts = list(drafts)
+    assert len(drafts) == 1
+    draft = drafts[0]
+    assert draft.status == "pending"
+    assert draft.proposed_name == fix.title
+    proposed = draft.proposed_parameters.get("parameters") if isinstance(draft.proposed_parameters, dict) else None
+    assert isinstance(proposed, list) and len(proposed) == 1
+    assert proposed[0]["key"] == "target_user"
+
+
+# ── Decision endpoint: build_template ─────────────────────────────────────
+
+@pytest.mark.asyncio
+async def test_build_template_returns_redirect_path(
+    client: AsyncClient, test_user, auth_headers, test_db
+):
+    session, fix = await _make_session_with_fix(test_db, test_user)
+    r = await client.post(
+        f"/api/v1/ai-sessions/{session.id}/suggested-fixes/{fix.id}/decision",
+        headers=auth_headers,
+        json={"decision": "build_template"},
+    )
+    assert r.status_code == 200
+    body = r.json()
+    assert body["redirect_path"] is not None
+    assert str(session.id) in body["redirect_path"]
+    assert str(fix.id) in body["redirect_path"]
+
+
+# ── Decision endpoint: 400 when no drafted script ─────────────────────────
+
+@pytest.mark.asyncio
+async def test_one_off_without_drafted_script_returns_400(
+    client: AsyncClient, test_user, auth_headers, test_db
+):
+    """A template-matched fix takes the dedicated /scripts/generate path; trying
+    to one_off it via this endpoint without an ai_drafted_script must surface
+    a clear client-error, not silently render nothing."""
+    session, fix = await _make_session_with_fix(test_db, test_user, with_drafted_script=False)
+    r = await client.post(
+        f"/api/v1/ai-sessions/{session.id}/suggested-fixes/{fix.id}/decision",
+        headers=auth_headers,
+        json={"decision": "one_off"},
+    )
+    assert r.status_code == 400
+    assert "ai_drafted_script" in r.json()["detail"]
+
+
+# ── Decision endpoint: edited script overrides ai_drafted_script ──────────
+
+@pytest.mark.asyncio
+async def test_edited_script_overrides_ai_drafted(
+    client: AsyncClient, test_user, auth_headers, test_db
+):
+    session, fix = await _make_session_with_fix(test_db, test_user)
+    r = await client.post(
+        f"/api/v1/ai-sessions/{session.id}/suggested-fixes/{fix.id}/decision",
+        headers=auth_headers,
+        json={
+            "decision": "one_off",
+            "edited_script": "Get-Service -Name Dnscache",
+        },
+    )
+    assert r.status_code == 200
+    assert r.json()["rendered_script"] == "Get-Service -Name Dnscache"