feat(memory): a coding turn from two weeks ago can still be found, after the conversation forgot it

CHANGELOG.md

@@ -8,6 +8,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/).
 
 ### Added
 
+- **A coding turn from two weeks ago can still be found, after the conversation itself forgot it: every finished turn joins an SQLite/FTS5 history index, and `recall_history` searches it.** Item 5 of the list audited on 2026-09-16. A code session keeps the model's message list and trims it at a user boundary — about thirty tool-using turns — so a question about the turn that fixed the login function a fortnight ago was a question the session file could no longer answer, and the memory store was the wrong place to look: a memory is a fact the agent chose to keep, not a record of what was asked. Now `chimera/memory/history.py` keeps one row per completed turn — the message, the answer, the files the turn read or edited (read off the same fold the replay endpoint shows), the tools, when, and whether the turn ran tainted — in `<home>/history.db`, FTS5 with the `LIKE` degradation the memory store already documents, through the same tokenizer and function-word list, scoped by project with the same key. Written by the turn's finishing code and never by the model, through `redact` (a pasted credential is not stored in clear), and never trimmed. The `recall_history` tool (read-only; in the set a step may run together) prints dated excerpts with the files each turn edited or read, scoped to the current project unless asked for every project, with a window in days; a turn that ran on untrusted content is labelled in-line the way a tainted memory is on recall. Deleting a conversation — one, or a whole project — deletes its rows, because the screen says the conversation is gone. An index that will not write logs and the turn goes on. Ten dictionaries for the Capabilities screen; the usage guide says which store is which in ten languages.
+
 - **The agent's browser draws on the Code screen: after every browser action, the viewport appears under the turn, with the page's address beside it.** Item 2 of the list audited on 2026-09-16 (the "cockpit"). The model reads a page as text — the tag tree `render_elements` prints — and until now the person read the same text, or watched a second Chromium window beside the app with `CHIMERA_BROWSER_HEADLESS` off; neither was watching the agent open the site and type the dates. Measured before wiring (2026-09-17, viewport 1280×720): a JPEG at quality 55 is 13 KB for example.com, 62 KB for a GitHub repo page, 96 KB for a Wikipedia article, 24–106 ms to capture — a PNG is 18/103/232 KB and up to 224 ms. So: one frame per browser action, never on a timer. `BrowserDriver.frame()` (the Playwright driver captures the viewport as JPEG 55), `BrowserTool.on_frame` (set by `assemble_registry` only when the request has a screen — a headless run never asks the driver for a picture, pinned by a test that counts), a `browser` frame on the turn's stream (`action`, `url`, `title`, `width`, `height`, base64 `jpeg`, `n`), and `BrowserView` under the exchange: the last frame — where the browser is, not where it was — with the URL as text, because a picture of a page is a claim about where the browser is and the address is the part a person can check. Frames are never written to the run log (replay is for words a dropped connection lost, not for redrawing a page that has moved on) and never replayed on reopen. A capture that fails is logged and skipped, never a word in the observation the model reads. Nothing in the panel is interactive: the person is watching, not driving. Ten dictionaries. Sabotage-verified on the wire: with the dispatcher's `browser` branch removed, the frame test fails.
 
 

apps/desktop/src/lib/i18n.tsx

@@ -96,6 +96,8 @@ const en: Dict = {
     "What a background job started by run_shell(background=true) is doing: running, finished (with its exit code), cancelled or lost — and the tail of its output. Without a job_id, lists every job.",
   "tools.desc.job_cancel":
     "Stop a background job started by run_shell(background=true): kills the command and everything it started. A job that already ended is reported as it is.",
+  "tools.desc.recall_history":
+    "Search what was asked and answered in this project's earlier coding conversations — turns the current conversation no longer holds. Returns dated excerpts with the files each turn edited or read. This is the conversation history, not the memory store.",
   "tools.desc.http_get":
     "Fetch a URL with an HTTP GET and return status + body text.",
   "tools.desc.execute_code":
@@ -1442,6 +1444,8 @@ const pt: Dict = {
     "O que um job em segundo plano iniciado por run_shell(background=true) está fazendo: rodando, terminado (com o código de saída), cancelado ou perdido — e o fim da saída dele. Sem job_id, lista todos os jobs.",
   "tools.desc.job_cancel":
     "Para um job em segundo plano iniciado por run_shell(background=true): mata o comando e tudo que ele iniciou. Um job que já terminou é reportado como está.",
+  "tools.desc.recall_history":
+    "Busca o que foi perguntado e respondido nas conversas de código anteriores deste projeto — turnos que a conversa atual já não guarda. Devolve trechos datados com os arquivos que cada turno editou ou leu. É o histórico da conversa, não o armazenamento de memória.",
   "tools.desc.http_get":
     "Busca uma URL com um HTTP GET e devolve o status + o texto do corpo.",
   "tools.desc.execute_code":
@@ -2837,6 +2841,8 @@ const es: Dict = {
     "Qué hace un trabajo en segundo plano iniciado por run_shell(background=true): en ejecución, terminado (con su código de salida), cancelado o perdido — y el final de su salida. Sin job_id, lista todos los trabajos.",
   "tools.desc.job_cancel":
     "Detiene un trabajo en segundo plano iniciado por run_shell(background=true): mata el comando y todo lo que inició. Un trabajo que ya terminó se informa tal como está.",
+  "tools.desc.recall_history":
+    "Busca lo que se preguntó y respondió en las conversaciones de código anteriores de este proyecto — turnos que la conversación actual ya no conserva. Devuelve extractos fechados con los archivos que cada turno editó o leyó. Es el historial de conversación, no el almacén de memoria.",
   "tools.desc.http_get":
     "Pide una URL con un HTTP GET y devuelve el estado + el texto del cuerpo.",
   "tools.desc.execute_code":
@@ -4206,6 +4212,8 @@ const fr: Dict = {
     "Ce que fait une tâche de fond lancée par run_shell(background=true) : en cours, terminée (avec son code de sortie), annulée ou perdue — et la fin de sa sortie. Sans job_id, liste toutes les tâches.",
   "tools.desc.job_cancel":
     "Arrête une tâche de fond lancée par run_shell(background=true) : tue la commande et tout ce qu'elle a lancé. Une tâche déjà terminée est rapportée telle quelle.",
+  "tools.desc.recall_history":
+    "Recherche ce qui a été demandé et répondu dans les conversations de code précédentes de ce projet — des tours que la conversation actuelle ne garde plus. Renvoie des extraits datés avec les fichiers que chaque tour a modifiés ou lus. C'est l'historique de conversation, pas la mémoire.",
   "tools.desc.http_get":
     "Récupère une URL par un HTTP GET et renvoie le statut + le texte du corps.",
   "tools.desc.execute_code":
@@ -5585,6 +5593,8 @@ const de: Dict = {
     "Was ein mit run_shell(background=true) gestarteter Hintergrundjob tut: läuft, beendet (mit Exit-Code), abgebrochen oder verloren — und das Ende seiner Ausgabe. Ohne job_id werden alle Jobs aufgelistet.",
   "tools.desc.job_cancel":
     "Stoppt einen mit run_shell(background=true) gestarteten Hintergrundjob: beendet den Befehl und alles, was er gestartet hat. Ein bereits beendeter Job wird so gemeldet, wie er ist.",
+  "tools.desc.recall_history":
+    "Durchsucht, was in früheren Code-Gesprächen dieses Projekts gefragt und beantwortet wurde — Züge, die das aktuelle Gespräch nicht mehr enthält. Liefert datierte Auszüge mit den Dateien, die jeder Zug bearbeitet oder gelesen hat. Das ist der Gesprächsverlauf, nicht der Gedächtnisspeicher.",
   "tools.desc.http_get":
     "Ruft eine URL per HTTP GET ab und gibt Status + Body-Text zurück.",
   "tools.desc.execute_code":
@@ -6955,6 +6965,8 @@ const zh: Dict = {
     "由 run_shell(background=true) 启动的后台任务在做什么：运行中、已完成（含退出码）、已取消或已丢失——以及其输出的末尾。不带 job_id 时列出所有任务。",
   "tools.desc.job_cancel":
     "停止由 run_shell(background=true) 启动的后台任务：终止该命令及其启动的一切。已结束的任务按原样报告。",
+  "tools.desc.recall_history":
+    "搜索本项目早先编码对话中问过和答过的内容——当前对话已不再保留的轮次。返回带日期的摘录，以及每轮编辑或读取的文件。这是对话历史，不是记忆库。",
   "tools.desc.http_get": "用 HTTP GET 抓取一个 URL，返回状态码和正文文本。",
   "tools.desc.execute_code":
     "运行一段 Python 3 代码并返回它的 stdout/stderr。它在所配置的沙箱里运行，而默认的沙箱就是这台机器本身 —— 不是隔离环境。",
@@ -8263,6 +8275,8 @@ const ja: Dict = {
     "run_shell(background=true) で開始したバックグラウンドジョブの状態：実行中、終了（終了コード付き）、キャンセル済み、または喪失 — と出力の末尾。job_id を省くと全ジョブを一覧します。",
   "tools.desc.job_cancel":
     "run_shell(background=true) で開始したバックグラウンドジョブを停止します：コマンドとそれが起動したすべてを終了します。すでに終わったジョブはそのまま報告します。",
+  "tools.desc.recall_history":
+    "このプロジェクトの以前のコーディング会話で尋ねられ答えられた内容を検索します — 現在の会話がもう保持していないターンです。各ターンが編集または読み取ったファイルとともに、日付付きの抜粋を返します。これは会話履歴であり、記憶ストアではありません。",
   "tools.desc.http_get":
     "URL を HTTP GET で取得し、ステータスと本文テキストを返します。",
   "tools.desc.execute_code":
@@ -9581,6 +9595,8 @@ const it: Dict = {
     "Cosa sta facendo un job in background avviato da run_shell(background=true): in esecuzione, terminato (con il codice di uscita), annullato o perso — e la coda del suo output. Senza job_id, elenca tutti i job.",
   "tools.desc.job_cancel":
     "Ferma un job in background avviato da run_shell(background=true): uccide il comando e tutto ciò che ha avviato. Un job già terminato viene riportato così com'è.",
+  "tools.desc.recall_history":
+    "Cerca cosa è stato chiesto e risposto nelle conversazioni di codice precedenti di questo progetto — turni che la conversazione attuale non conserva più. Restituisce estratti datati con i file che ogni turno ha modificato o letto. È la cronologia della conversazione, non l'archivio della memoria.",
   "tools.desc.http_get":
     "Richiede una URL con un HTTP GET e restituisce lo stato + il testo del corpo.",
   "tools.desc.execute_code":
@@ -10952,6 +10968,8 @@ const pl: Dict = {
     "Co robi zadanie w tle uruchomione przez run_shell(background=true): działa, zakończone (z kodem wyjścia), anulowane lub utracone — oraz końcówka jego wyjścia. Bez job_id wypisuje wszystkie zadania.",
   "tools.desc.job_cancel":
     "Zatrzymuje zadanie w tle uruchomione przez run_shell(background=true): zabija polecenie i wszystko, co uruchomiło. Zadanie już zakończone jest raportowane takie, jakie jest.",
+  "tools.desc.recall_history":
+    "Przeszukuje, o co pytano i co odpowiedziano we wcześniejszych rozmowach o kodzie w tym projekcie — tury, których bieżąca rozmowa już nie przechowuje. Zwraca datowane fragmenty z plikami, które każda tura edytowała lub czytała. To historia rozmowy, nie magazyn pamięci.",
   "tools.desc.http_get":
     "Pobiera URL przez HTTP GET i zwraca status + tekst treści.",
   "tools.desc.execute_code":
@@ -12316,6 +12334,8 @@ const ru: Dict = {
     "Что делает фоновая задача, запущенная через run_shell(background=true): выполняется, завершена (с кодом выхода), отменена или потеряна — и конец её вывода. Без job_id выводит все задачи.",
   "tools.desc.job_cancel":
     "Останавливает фоновую задачу, запущенную через run_shell(background=true): убивает команду и всё, что она запустила. Уже завершённая задача сообщается как есть.",
+  "tools.desc.recall_history":
+    "Ищет, о чём спрашивали и что отвечали в прежних беседах о коде этого проекта — ходы, которых текущая беседа уже не хранит. Возвращает датированные выдержки с файлами, которые каждый ход редактировал или читал. Это история беседы, а не хранилище памяти.",
   "tools.desc.http_get":
     "Запрашивает URL через HTTP GET и возвращает статус и текст тела ответа.",
   "tools.desc.execute_code":

chimera/api/code_api.py

@@ -1019,6 +1019,7 @@ def register_code_api(
     from chimera.core.jobs import jobs_for
     from chimera.core.redact import redact
     from chimera.interface.session import recall_facts
+    from chimera.memory.history import files_of_exchange, history_for
     from chimera.memory.models import project_key
 
     # What the injected `memory` IS, so a later turn can tell "the owner changed the backend" from
@@ -1033,6 +1034,11 @@ def register_code_api(
     live: Callable[[], Settings] = live_settings or (lambda: settings)
 
     store = CodeSessionStore(settings.home / "code_sessions")
+    # The index of finished turns (`chimera.memory.history`), one per home, shared with the
+    # `recall_history` tool every registry mounts. The session file is what a conversation is
+    # RESUMED from and trims itself accordingly; this is what a person's question about a turn
+    # from two weeks ago is answered from, and it never trims.
+    history = history_for(settings.home)
     # Beside the conversations, not inside them: a project you have added but not yet worked in
     # has no conversation to hang off, which is the whole reason the list cannot be derived.
     projects = CodeProjectRegistry(settings.home / "code_projects.json")
@@ -1423,6 +1429,30 @@ def _verify_and_finish(payload: dict[str, Any]) -> None:
                         store.save(session)
                     except OSError as exc:  # noqa: BLE001 — a failed record must not fail the turn
                         _log.debug("could not store the turn receipt: %s", exc)
+                    # The turn joins the conversation history index — the record that outlives the
+                    # session's own trimming, so "what did we do about the login page two weeks
+                    # ago?" has somewhere to look. Written by this code and not by the model, after
+                    # the transcript is saved, and read off the same fold the replay endpoint shows
+                    # (so the files the index says a turn touched are the files the screen shows
+                    # it touching). A record that will not write must not fail a turn that was
+                    # already paid for; the index logs and the turn goes on.
+                    try:
+                        from chimera.api.code_replay import exchanges_from_messages
+
+                        exchanges = exchanges_from_messages(session.to_dict()["messages"])
+                        history.record(
+                            turn_id=turn_id,
+                            session_id=session_id,
+                            project=project_key(ws),
+                            asked=message,
+                            answered=str(payload.get("answer") or ""),
+                            files=files_of_exchange(exchanges[-1]) if exchanges else [],
+                            edited=list(edited),
+                            tools=[str(t) for t in (payload.get("tool_names") or [])],
+                            tainted=bool(payload.get("tainted")),
+                        )
+                    except Exception as exc:  # noqa: BLE001 — a failed record must not fail the turn
+                        _log.debug("could not index the turn in the history: %s", exc)
                     emit("done", payload)
 
                 # Same swap the chat turn uses, under the same per-session lock: hand the agent the
@@ -1932,9 +1962,13 @@ def delete_code_session(session_id: str) -> dict[str, bool]:
         """Forget a conversation. An unknown id is ``{ok: false}`` with a 200, not a 404 — that is
         exactly the state a second click on Clear hits, and it is not an error."""
         try:
-            return {"ok": store.delete(session_id)}
+            gone = store.delete(session_id)
         except ValueError:
             return {"ok": False}
+        # Its rows in the history index go with it: the screen says the conversation is gone, and
+        # an index that still answered questions about it would make that a lie.
+        history.forget_session(session_id)
+        return {"ok": gone}
 
     @app.delete("/api/code/projects", dependencies=[guard], response_model=DeletedCountOut)
     def delete_code_project(workspace: str) -> dict[str, int]:
@@ -1945,7 +1979,12 @@ def delete_code_project(workspace: str) -> dict[str, int]:
         mean the transcripts — and the count comes back so the screen can say how many went rather
         than reporting a success with no size.
         """
-        return {"deleted": store.delete_project(workspace)}
+        # The ids first, then the files, then the index: the index is keyed by session id, and
+        # the list is the only place the workspace-to-id mapping exists.
+        ids = [str(m["id"]) for m in store.list_meta() if m["workspace"] == workspace]
+        deleted = store.delete_project(workspace)
+        history.forget_sessions(ids)
+        return {"deleted": deleted}
 
     # The registered projects live at `/workspaces`, NOT at `/projects`, and the distance is
     # deliberate. `DELETE /api/code/projects` above already means "delete every conversation filed

chimera/core/agent.py

@@ -43,6 +43,8 @@
         "scrape", "extract", "map",
         # Reads a job's record and log; `job_cancel` kills and is not here.
         "job_status",
+        # Reads the index of finished turns; writes nothing.
+        "recall_history",
     }
 )
 #: How many of a step's calls run at once. Four is a fetch batch, not a fan-out: a model asks for a