ltcai 9.1.0 → 9.2.0

package/README.md

@@ -1,13 +1,18 @@
 # Lattice AI
 
-**Lattice AI 9.1.0 is the local-first Digital Brain platform. Code Review Completion & Fail-Closed Runtime closes every actionable finding from the July 11 review: security boundaries now deny by default, runtime and chat state have typed ownership, frontend failures stay visible and tested, and repository/release hygiene is enforced across the 9.1.0 line.**
+**Lattice AI 9.2.0 is the local-first Digital Brain platform. Model-Agnostic File Generation makes "create a file" work reliably with any loaded LLM: file requests run through a strict extension-aware prompt, robust extraction, per-type validation, one corrective retry, and a deterministic repair fallback, so even small local models (gemma/qwen class) always produce a structurally valid HTML, JSON, or code file.**
 
 **Lattice AI는 모델이 바뀌어도 내 지식과 맥락을 보존하는 로컬 우선 AI 브레인입니다.**
 
-> The 9.1.0 release adds request-scoped model routing,
+> The 9.2.0 release hardens file creation end to end: chat file requests are
+> routed to a deterministic direct-write path (including type-only requests
+> like "html 파일 만들어줘"), model replies are treated as untrusted content
+> and cleaned of fences, reasoning blocks, and chat framing, and the agent
+> JSON loop tolerates small-model formatting slips instead of aborting.
+> The 9.1.0 release added request-scoped model routing,
 > workspace-isolated graph identities and frontend caches, fail-closed admin
 > gates, SSRF-safe web capture, private local state permissions, and reproducible
-> release/test isolation. The same release introduces a human-first UI:
+> release/test isolation. The same release introduced a human-first UI:
 > a visible knowledge journey from conversation or source capture into the
 > living Brain, its real relationship graph, and memory-grounded automation;
 > the empty Brain home now keeps that complete loop inside one viewport, with
@@ -87,10 +92,11 @@ You need Lattice AI when:
 The screenshots below are the latest checked-in visual evidence captures. They
 keep the first-run Brain flow, memory graph, source capture, model library,
 system view, admin console, and review center visible as release gates while
-9.1.0 completes the July 11 code review with fail-closed access control,
-typed runtime/model state, a decomposed chat API, honest frontend error states
-and unit tests, and repository/release hygiene. The captures below are the
-checked-in 9.1.0 visual release evidence for that product flow.
+9.2.0 makes file generation model-agnostic: chat file requests always end in
+a structurally valid file regardless of which LLM is loaded, backed by
+extraction, validation, corrective retry, and deterministic repair. The
+captures below are the checked-in 9.2.0 visual release evidence for that
+product flow.
 
 ### 1. Wake Brain
 
@@ -102,21 +108,21 @@ confirm owner, check the computer, choose the Brain voice.
 Choose the owner of the Brain. The profile is not a SaaS account by default; it
 is the local identity for the knowledge you keep.
 
-![Login](output/release/v9.1.0/screenshots/01-login.png)
+![Login](output/release/v9.2.0/screenshots/01-login.png)
 
 ### 3. Recommended Models
 
 Start with a short list: safest recommendation, faster model, stronger model.
 Advanced details stay available without overwhelming first-time users.
 
-![Recommended Models](output/release/v9.1.0/screenshots/02-recommended-models.png)
+![Recommended Models](output/release/v9.2.0/screenshots/02-recommended-models.png)
 
 ### 4. Install And Load
 
 Download and load only after consent. Lattice explains model size, local
 execution, and network use before work starts.
 
-![Install and Load](output/release/v9.1.0/screenshots/03-install-load-progress.png)
+![Install and Load](output/release/v9.2.0/screenshots/03-install-load-progress.png)
 
 ### 5. Brain Chat
 
@@ -128,14 +134,14 @@ and its visible life signal follow real listening, recall, synthesis, and action
 state. Detailed memory rings, provenance, conversation history, and
 model/runtime proof open as overlays only when requested.
 
-![One-viewport Living Brain Home](output/release/v9.1.0/screenshots/04-brain-chat-home.png)
+![One-viewport Living Brain Home](output/release/v9.2.0/screenshots/04-brain-chat-home.png)
 
 ### 6. Review Center
 
 Automation results are staged for review before they become durable decisions.
 Snooze, unsnooze, run now, approve, and dismiss actions stay explicit.
 
-![Review Center](output/release/v9.1.0/screenshots/12-review-center.png)
+![Review Center](output/release/v9.2.0/screenshots/12-review-center.png)
 
 ## Brain Depths
 
@@ -151,10 +157,10 @@ The user travels inward from everyday memory to deeper structure:
 
 Walkthrough:
 
-![v9.1.0 Living Brain walkthrough](output/release/v9.1.0/gifs/v9.1.0-living-brain-walkthrough.gif)
+![v9.2.0 Living Brain walkthrough](output/release/v9.2.0/gifs/v9.2.0-living-brain-walkthrough.gif)
 
 Screenshot index and capture notes:
-[output/release/v9.1.0/SCREENSHOT_INDEX.md](output/release/v9.1.0/SCREENSHOT_INDEX.md)
+[output/release/v9.2.0/SCREENSHOT_INDEX.md](output/release/v9.2.0/SCREENSHOT_INDEX.md)
 
 ## Install
 
@@ -239,42 +245,43 @@ See [docs/DEVELOPMENT.md](docs/DEVELOPMENT.md) for developer workflow details.
 
 ## Current Release
 
-The current release is **9.1.0 — Code Review Completion & Fail-Closed Runtime**:
-
-- Telegram rejects every chat and callback outside
-  `LATTICEAI_TELEGRAM_ALLOWED_CHAT_IDS`, and its local API bridge requires
-  `LATTICEAI_SERVER_SESSION_TOKEN` instead of scanning stored sessions.
-- Public invitation authorization is signed and expiring, default invitation
-  credentials are removed, Knowledge Graph scope lookup fails closed, and
-  desktop/knowledge/network tools use explicit policy and consent boundaries.
-- App assembly exports typed runtime stages, model routing uses injected typed
-  state, and chat history, documents, streaming, and contracts are split into
-  focused modules.
-- The React workspace distinguishes unavailable services from empty Brain data,
-  gates success callbacks on real success, and has Vitest coverage for API
-  result shapes, proof state, conversation state, primitives, and i18n.
-- Large frontend i18n/CSS/Brain hooks and repeated runtime utilities are split,
-  obsolete aliases and local VSIX artifacts are removed, and review documents
-  live under `docs/reviews/` as immutable history.
-
-Expected artifacts for 9.1.0 release must use exact filenames:
-
-- `dist/ltcai-9.1.0-py3-none-any.whl`
-- `dist/ltcai-9.1.0.tar.gz`
-- `ltcai-9.1.0.tgz`
-- `dist/ltcai-9.1.0.vsix`
-- `src-tauri/target/release/bundle/dmg/Lattice AI_9.1.0_aarch64.dmg`
+The current release is **9.2.0 — Model-Agnostic File Generation**:
+
+- A new `latticeai.core.file_generation` pipeline treats every model reply as
+  untrusted content: extension-aware strict prompting, extraction of the real
+  payload from fences/`<think>` blocks/chat framing, per-type structural
+  validation (complete HTML documents, parseable JSON, CSS rule blocks), one
+  corrective retry that tells the model what was wrong, and a deterministic
+  repair fallback that guarantees a valid file.
+- Type-only requests without an explicit filename ("html 파일 만들어줘",
+  "웹페이지 만들어줘") now resolve to an inferred target and use the
+  deterministic direct-write path instead of the model-driven agent loop.
+- The agent executor tolerates small-model formatting slips: `<think>` blocks
+  and trailing commas are stripped from action JSON, and parse failures feed a
+  corrective format reminder back to the model instead of aborting the run.
+- The executor prompt pins exact file-content rules for `write_file` so
+  agent-created files are complete and fence-free.
+- File writes report whether content was generated cleanly, retried, or
+  auto-repaired in the response payload (`generation` metadata).
+
+Expected artifacts for 9.2.0 release must use exact filenames:
+
+- `dist/ltcai-9.2.0-py3-none-any.whl`
+- `dist/ltcai-9.2.0.tar.gz`
+- `ltcai-9.2.0.tgz`
+- `dist/ltcai-9.2.0.vsix`
+- `src-tauri/target/release/bundle/dmg/Lattice AI_9.2.0_aarch64.dmg`
 
 Do not use wildcard artifact uploads. Package registry publishing remains owner-run.
 
 See [docs/ROADMAP_RECOMMENDATIONS.md](docs/ROADMAP_RECOMMENDATIONS.md) for the
-strategic roadmap slices applied through 9.1.0 and the follow-up tracks.
+strategic roadmap slices applied through 9.2.0 and the follow-up tracks.
 
 ## Known Limitations
 
 - External package registries are owner-published and can lag behind GitHub.
 - PostgreSQL/pgvector is optional scale/migration tooling. SQLite remains the
-  live local Brain store in 9.1.0.
+  live local Brain store in 9.2.0.
 - Docker, model downloads, cloud model calls, Telegram, Brain Network, and update
   checks require explicit user action.
 - Conversation does not fabricate answers when no model is loaded.
@@ -286,6 +293,7 @@ strategic roadmap slices applied through 9.1.0 and the follow-up tracks.
 
 | Version | Theme |
 | --- | --- |
+| 9.2.0 | Model-Agnostic File Generation: chat file requests always produce structurally valid files with any LLM via extraction, per-type validation, corrective retry, deterministic repair, inferred file targets, and a fault-tolerant agent JSON loop |
 | 9.1.0 | Code Review Completion & Fail-Closed Runtime: all July 11 review findings closed across fail-closed security, typed runtime/model/chat boundaries, honest frontend failures and tests, and repository hygiene |
 | 9.0.0 | Code Review Closure & Runtime Cleanup: July 8 code-review follow-ups fixed, chat/runtime reliability improved, duplicated utility surfaces consolidated, runtime audit append paths moved to JSONL, and release metadata/artifacts synchronized |
 | 8.9.0 | Scoped Memory & Tool Policy Hardening: authenticated history/KG reads are workspace-scoped, direct Tool API paths enforce registry policy, local approvals hash tokens at rest, AgentRuntime approval semantics are explicit, and frontend/runtime seams are split |

package/docs/CHANGELOG.md

@@ -4,6 +4,38 @@ The top entry is either the current unreleased main-branch work or the current
 release line. Older entries are historical and may describe behavior as it
 existed at that release.
 
+## [9.2.0] - 2026-07-20
+
+### Added
+- Added `latticeai/core/file_generation.py`, a model-agnostic file content
+  pipeline: extension-aware strict prompting with pinned first lines, payload
+  extraction from fences/`<think>` blocks/chat framing, per-type structural
+  validation, one corrective retry, and a deterministic repair fallback that
+  guarantees a structurally valid file from any loaded LLM.
+- Added filename inference for chat file requests that name a type but no path
+  ("html 파일 만들어줘" → `generated_page.html`), keeping such requests on the
+  deterministic direct-write path instead of the model-driven agent loop.
+- Added `generation` metadata (attempts, validation reasons, auto-repair flag)
+  to `/chat` direct file-action responses, and a user-facing notice when the
+  saved file was produced by auto-repair.
+- Added `tests/unit/test_file_generation.py` covering extraction, validation,
+  repair, inference, prompting, and retry/repair orchestration (22 tests).
+
+### Changed
+- Chat file-action routing counts explicit artifact types (html, 웹페이지,
+  webpage) as file words, so "html 페이지 만들어줘" creates a real file.
+- Direct file generation clamps temperature to ≤0.3 and raises the token
+  budget to ≥4096 so generated documents complete.
+- The agent executor prompt pins `write_file` content rules: complete raw
+  content, no Markdown fences, extension-valid documents.
+
+### Fixed
+- Small local models (gemma/qwen class) no longer save chat wrappers, fenced
+  code blocks, reasoning traces, or truncated documents as file bytes.
+- The agent JSON loop no longer aborts on the first malformed action reply:
+  `extract_action` strips `<think>` blocks and tolerates trailing commas, and
+  up to two corrective format reminders are fed back before halting.
+
 ## [9.1.0] - 2026-07-11
 
 ### Added

package/docs/COMMUNITY_AND_PLUGINS.md

@@ -1,12 +1,13 @@
 # Community And Plugins
 
-Current release: **9.1.0 — Code Review Completion & Fail-Closed Runtime**.
+Current release: **9.2.0 — Model-Agnostic File Generation**.
 
 LatticeAI defines the path from a strong local-first framework (8.4.0
 action-aware baseline, 8.5.0 registry+DI hardening, 8.6.0 capture/navigation
 reliability, 8.7.0 runtime-state hygiene, 8.8.0 Brain Core extraction
-readiness, 8.9.0 scoped Tool API hardening, 9.0.0 cleanup closure, and 9.1.0
-fail-closed review completion) to a product ecosystem. The
+readiness, 8.9.0 scoped Tool API hardening, 9.0.0 cleanup closure, 9.1.0
+fail-closed review completion, and 9.2.0 model-agnostic file generation) to a
+product ecosystem. The
 immediate goal is small and practical: make it clear how
 contributors can extend the Brain without weakening local-first trust,
 workspace scoping, or release quality.

package/docs/DEVELOPMENT.md

@@ -1,10 +1,10 @@
 # Lattice AI Development
 
-Current release: **9.1.0 — Code Review Completion & Fail-Closed Runtime**.
+Current release: **9.2.0 — Model-Agnostic File Generation**.
 
 This document is for contributors working on the local-first Digital Brain
 codebase. Product positioning and quick start stay in `README.md`; release
-history is intentionally limited to 8.0.0-9.1.0 in `docs/CHANGELOG.md` and
+history is intentionally limited to 8.0.0-9.2.0 in `docs/CHANGELOG.md` and
 `RELEASE.md`.
 
 ## Product Contract
@@ -153,10 +153,10 @@ For user-facing, API, runtime, release, or packaging changes, check:
 Release/publish examples must use exact target-version filenames. Do not
 document wildcard artifact upload commands.
 
-For 9.1.0 release work, exact artifacts are:
+For 9.2.0 release work, exact artifacts are:
 
-- `dist/ltcai-9.1.0-py3-none-any.whl`
-- `dist/ltcai-9.1.0.tar.gz`
-- `ltcai-9.1.0.tgz`
-- `dist/ltcai-9.1.0.vsix`
-- `src-tauri/target/release/bundle/dmg/Lattice AI_9.1.0_aarch64.dmg`
+- `dist/ltcai-9.2.0-py3-none-any.whl`
+- `dist/ltcai-9.2.0.tar.gz`
+- `ltcai-9.2.0.tgz`
+- `dist/ltcai-9.2.0.vsix`
+- `src-tauri/target/release/bundle/dmg/Lattice AI_9.2.0_aarch64.dmg`

package/docs/LEGACY_COMPATIBILITY.md

@@ -1,6 +1,6 @@
 # Legacy Compatibility Map
 
-Current target: **9.1.0 — Code Review Completion & Fail-Closed Runtime**.
+Current target: **9.2.0 — Model-Agnostic File Generation**.
 
 Lattice AI is moving toward a smaller, modular architecture centered on
 `lattice_brain`, `latticeai.services`, `latticeai.api`, and `latticeai.runtime`.

package/docs/ONBOARDING.md

@@ -1,6 +1,6 @@
 # Lattice AI Onboarding
 
-Current release: **9.1.0 — Code Review Completion & Fail-Closed Runtime**.
+Current release: **9.2.0 — Model-Agnostic File Generation**.
 
 The first-run goal is a five-minute path from "I opened the app" to "my Brain
 has a source, a question, and proof." This page is the product contract behind
@@ -40,7 +40,7 @@ read the docs first.
 
 ## Release Gate
 
-9.1.0 treats onboarding as a release gate, not marketing copy. The current
+9.2.0 treats onboarding as a release gate, not marketing copy. The current
 machine-checkable product readiness report requires this five-minute contract,
 the Brain Home surface, setup helpers, graph ingestion tests, and exact release
 artifact documentation before the release can be called complete.

package/docs/OPERATIONS.md

@@ -1,4 +1,4 @@
-# Lattice AI — Operations Guide (v9.1.0)
+# Lattice AI — Operations Guide (v9.2.0)
 
 ## 1. 데이터 파일 위치
 

package/docs/TRUST_MODEL.md

@@ -1,6 +1,6 @@
 # Lattice AI Trust Model
 
-Current release: **9.1.0 — Code Review Completion & Fail-Closed Runtime**.
+Current release: **9.2.0 — Model-Agnostic File Generation**.
 
 Lattice AI is local-first, explicit about external communication, and honest
 when a capability is unavailable.

package/docs/WHY_LATTICE.md

@@ -1,6 +1,6 @@
 # Why Lattice AI Exists
 
-Current release: **9.1.0 — Code Review Completion & Fail-Closed Runtime**.
+Current release: **9.2.0 — Model-Agnostic File Generation**.
 
 **Lattice AI is a local-first Digital Brain that keeps your knowledge durable
 across any AI model.**

package/docs/kg-schema.md

@@ -1,6 +1,6 @@
 # Knowledge Graph Schema
 
-Current release: **9.1.0 — Code Review Completion & Fail-Closed Runtime**.
+Current release: **9.2.0 — Model-Agnostic File Generation**.
 
 명세 출처: `lattice_ai_full_spec.pptx` 슬라이드 20·21·22
 구현: `kg_schema.py`

package/docs/mcp-tools.md

@@ -1,4 +1,4 @@
-# MCP 도구 카탈로그 (v9.1.0)
+# MCP 도구 카탈로그 (v9.2.0)
 
 Lattice AI는 MCP(Model Context Protocol) 서버로 동작하여 Claude Desktop, Cursor 등에서 직접 도구를 사용할 수 있습니다.
 

package/lattice_brain/__init__.py

@@ -26,7 +26,7 @@ from .storage import (
     storage_from_env,
 )
 
-__version__ = "9.1.0"
+__version__ = "9.2.0"
 
 __all__ = [
     "AgentRuntime",

package/lattice_brain/runtime/multi_agent.py

@@ -21,7 +21,7 @@ from .contracts import multi_agent_contract
 from ..utils import now_iso as _now
 
 
-MULTI_AGENT_VERSION = "9.1.0"
+MULTI_AGENT_VERSION = "9.2.0"
 
 AGENT_ROLES = ("researcher", "planner", "executor", "reviewer", "release")
 CORE_PIPELINE = ("planner", "executor", "reviewer")

package/latticeai/__init__.py

@@ -1,3 +1,3 @@
 """Lattice AI - modular server package."""
 
-__version__ = "9.1.0"
+__version__ = "9.2.0"

package/latticeai/api/chat_helpers.py

@@ -128,6 +128,10 @@ def is_file_action_request(text: str) -> bool:
     has_target = bool(_FILE_TARGET_RE.search(raw))
     has_file_word = any(word in lower for word in (
         "file", "파일", "문서", "artifact", "아티팩트", "save as", "저장",
+        # Explicit artifact types: users asking for "an html page" expect a
+        # real file, not a code block in chat — route them to the file path
+        # where the model-agnostic generation pipeline guarantees validity.
+        "html", "웹페이지", "웹 페이지", "홈페이지", "webpage", "web page",
     ))
     has_action = any(word in lower for word in (
         "create", "make", "write", "save", "generate", "edit", "update",

package/latticeai/api/chat_intents.py

@@ -14,10 +14,10 @@ from latticeai.api.chat_helpers import (
     file_action_target,
     format_network_status,
     inline_file_action_content,
-    strip_generated_file_content,
 )
 from latticeai.api.chat_stream import agent_payload_stream, single_answer_response
 from latticeai.core.agent import AgentState
+from latticeai.core.file_generation import generate_file_content, infer_file_target
 
 
 class ChatIntentController:
@@ -185,28 +185,36 @@ class ChatIntentController:
         return single_answer_response(req, answer, model="client_url")
 
     async def direct_file_action(self, req: Any, *, model_id: Optional[str]):
-        target_path = file_action_target(req.message)
+        # An explicit path ("report.txt") wins; otherwise infer one from an
+        # explicit type keyword ("html 파일 만들어줘") so weak models never
+        # have to drive the JSON tool loop just to create a file.
+        target_path = file_action_target(req.message) or infer_file_target(req.message)
         if not target_path:
             return None
         content = inline_file_action_content(req.message)
+        generation_meta: Optional[Dict[str, Any]] = None
         if content is None and not model_id:
             return self.no_model_response()
         if content is None and model_id:
-            generation_context = (
-                "Create the exact content for the requested file. "
-                "Return only the file bytes as plain text. "
-                "Do not wrap the answer in Markdown fences, commentary, or explanations.\n\n"
-                f"Target path: {target_path}\n"
-                f"User request: {req.message}"
-            )
-            raw_content = await self.router.generate_as(
-                model_id,
-                message="Return only the requested file content.",
-                context=generation_context,
-                max_tokens=req.max_tokens,
-                temperature=req.temperature,
+            # Model-agnostic pipeline: strict extension-aware prompt →
+            # extraction → validation → one corrective retry → deterministic
+            # repair. Guarantees a structurally valid file from any LLM.
+            async def _generate(context: str) -> str:
+                return str(
+                    await self.router.generate_as(
+                        model_id,
+                        message="Return only the requested file content.",
+                        context=context,
+                        max_tokens=max(int(req.max_tokens or 0), 4096),
+                        temperature=min(req.temperature, 0.3),
+                    )
+                )
+
+            content, generation_meta = await generate_file_content(
+                _generate,
+                target_path=target_path,
+                user_request=req.message,
             )
-            content = strip_generated_file_content(str(raw_content))
         if content is None:
             raise HTTPException(
                 status_code=400,
@@ -220,6 +228,10 @@ class ChatIntentController:
         except self.tool_error as exc:
             raise HTTPException(status_code=400, detail=str(exc)) from exc
         answer = f"{result.get('path') or target_path} 파일을 만들었습니다."
+        if generation_meta and generation_meta.get("repaired"):
+            answer += (
+                " (모델 출력이 불완전해 자동 보정을 거쳐 유효한 파일로 저장했습니다.)"
+            )
         created_files = [
             {
                 "path": result.get("path") or target_path,
@@ -248,6 +260,8 @@ class ChatIntentController:
             "routed_to_agent": True,
             "action_route": "direct_write_file",
         }
+        if generation_meta is not None:
+            payload["generation"] = generation_meta
         if req.stream:
             return StreamingResponse(
                 agent_payload_stream(

package/latticeai/core/agent.py

@@ -69,9 +69,16 @@ class AgentRunContext:
         self.approved_by_human: bool       = False
 
 
+_THINK_BLOCK_RE = re.compile(
+    r"<(think|thinking|reasoning)>.*?</\1>", flags=re.DOTALL | re.IGNORECASE
+)
+
+
 def extract_action(raw: str) -> Dict:
     """Parse one JSON action object out of an LLM response (tolerant of fences/prose)."""
-    text = raw.strip()
+    # Small local models often prepend <think>...</think> reasoning that can
+    # itself contain braces — drop it before locating the action object.
+    text = _THINK_BLOCK_RE.sub("", raw).strip()
     fenced = re.search(r"```(?:json)?\s*(\{.*?\})\s*```", text, flags=re.DOTALL)
     if fenced:
         text = fenced.group(1).strip()
@@ -83,8 +90,14 @@ def extract_action(raw: str) -> Dict:
 
     try:
         action = json.loads(text)
-    except json.JSONDecodeError as exc:
-        raise ValueError(f"Agent did not return valid JSON: {exc}") from exc
+    except json.JSONDecodeError:
+        # Second chance for the most common small-model JSON slips: trailing
+        # commas before a closing brace/bracket.
+        repaired = re.sub(r",\s*([}\]])", r"\1", text)
+        try:
+            action = json.loads(repaired)
+        except json.JSONDecodeError as exc:
+            raise ValueError(f"Agent did not return valid JSON: {exc}") from exc
 
     if not isinstance(action, dict) or "action" not in action:
         raise ValueError("Agent JSON must include an action field.")
@@ -248,6 +261,7 @@ class SingleAgentRuntime:
         d = self.deps
         exec_count = sum(1 for s in ctx.transcript if s.get("state") == AgentState.EXECUTING.value)
         budget = max(1, max_steps - exec_count)
+        parse_failures = 0
 
         for _ in range(budget):
             corrections_hint = (
@@ -280,11 +294,24 @@ class SingleAgentRuntime:
             try:
                 action = extract_action(str(raw))
             except ValueError as exc:
+                parse_failures += 1
                 ctx.transcript.append({
                     "state": AgentState.EXECUTING.value, "action": "parse_error",
                     "raw": str(raw)[:400], "error": str(exc),
                 })
-                break
+                if parse_failures >= 3:
+                    break
+                # Weak models often need one concrete reminder of the wire
+                # format; feed it through the corrections channel and retry
+                # instead of aborting the whole run on the first slip.
+                hint = (
+                    'Your last reply was not a single JSON action object. Reply with '
+                    'EXACTLY one JSON object like {"thoughts": "...", "action": '
+                    '"tool_name", "args": {...}} and nothing else.'
+                )
+                if hint not in ctx.corrections:
+                    ctx.corrections.append(hint)
+                continue
 
             name     = action.get("action")
             thoughts = str(action.get("thoughts") or "")[:600]

package/latticeai/core/agent_prompts.py

@@ -43,6 +43,11 @@ You think and act like a senior software engineer:
 Respond with exactly ONE JSON object per step:
 {"thoughts": "what you learned / why this next action", "action": "tool_name", "args": {...}}
 
+When writing a file (write_file), args.content must be the COMPLETE raw file
+content: no Markdown fences, no commentary, valid for the file's extension
+(an .html file starts with <!DOCTYPE html> and ends with </html>; a .json
+file must parse as strict JSON).
+
 When the task is fully done AND a tool result in this run confirms it:
 {"thoughts": "verified", "action": "final", "message": "한국어로 무엇을 했고 어디서 검증했는지 요약"}