ShadowHarness

Sept launch delayed. Installer tuning.
Official site: SIDERA batch:

[ShadowHarness @ShadowHarness]

From 8.33 GB and 1,463 sessions, retrieve the past you need now.
This is a live demo of ShadowHarness's long-term memory recall system.

ShadowHarness recalls what matters for the current question and supplies it to the LLM. The LLM does not navigate to a folder to reread memory.md every time, or search the entire massive raw log.

It recalls long-term changes in development, verification, and design decisions beyond simple keyword matching.

SIDERA goes beyond conversation summaries and vector search. It forms concepts, states, conclusions, changes, and evidence from a long development history.

Combined with ShadowHarness's structured recall mechanism, these formed memories bring the past you need back into the current context without rereading the entire history every time.

[ShadowHarness @ShadowHarness]

表向きは「Pro $200、また入れるよ」。
中身は計算方法を変えて、同じ200ドルで買えるAPI換算を旧Proの半分にする。

公式は「効率化と値下げを回すから、そのうち同じ金で今より働ける」。いつ、何が、どれだけ増えるかは言わない。決まってるのは今半分になることだけ。

APIと比べたらまだ安い。旧Proと比べたら高い。同じ金で枠が半分なんだから当然。

[ShadowHarness @ShadowHarness]

Proプランなんだぜ⋯⋯グロいだろ
（さっきリセットしたばかり）

[ShadowHarness @ShadowHarness]

This is the power of SIDERA: tracing how a design decision evolved across nearly three months of development, with references to the original conversations.

The screenshot also distinguishes records available only on the local PC and today's work that SIDERA has not yet ingested.

English translation of the original text in the screenshot:

Date: June 26
Event: The Claude pane designed center-prepare.mjs as the "single source for the freeze sequence before center starts."
SIDERA session pointer: 4d030074-67ba-4738-b10b-e71a2e6384a3 (Only in this PC's C--git-shadow raw logs. There are zero C--git-shadow records in SIDERA's raw archive.)

Date: June 29
Event: The first appearance in the raw logs of launching claude with --session-id.
SIDERA session pointer: cee3d206-c9e4-4d48-af7d-eef463462cff (Likewise, available only on the PC.)

Date: June 30
Event: The earliest record in SIDERA: "At the center of Shadow is the actual Claude Code CLI. Grok/Fugu/Codex are auxiliary."
SIDERA session pointer: 019f177f-2fdd-7123-a9b2-2b38b47da2a8#k2

Date: July 1
Event: Codex's diagnosis: the login flow was not appearing because the launch included session-id, --chrome, and reconstruct-prompt.
SIDERA session pointer: 019f1b14-732f-7f21-b882-27070cdd649e#k0

Date: July 1
Event: The conclusion of the same conversation: keep session-id in the normal launch path; use a plain launch only for login recovery.
SIDERA session pointer: 019f1b14-732f-7f21-b882-27070cdd649e#k2

Date: July 8
Event: The Claude pane reported completing all seven steps to "eliminate the center concept," including the prescribed build.
SIDERA session pointer: 2be1e365-c57f-47b8-a3f2-44b56c3728aa#k0

Date: July 8
Event: The Claude pane changed the branching condition from "Is the pane ID center?" to "What is it running?"
SIDERA session pointer: df25e781-cf1a-4cfd-af93-f164a5ce9b7c#k4

Date: July 12
Event: Codex judged the migration incomplete. #center, terminals.get("center"), panes/center, and center-prepare were still in active use.
SIDERA session pointers: 019f569c-ce84-7550-a1ce-b2d91fa4335f#k1, 019f56a1-0e3f-7a53-bf29-bd0d962562b9#k1

Date: September 18
Event: The Claude-specific path remained because unification had not been completed; there was no technical need for it.
SIDERA session pointer: 0055cffe-813d-46de-9e02-15200934f7af#k0 through #k2

Date: September 18
Event: The GPT 1 pane carried out the removal. Claude was moved to a plain launch via claude-pane.ps1 on the shared provider path. center-prepare.mjs, reset-watcher.ps1, and reconstruct-prompt were deleted. The changes were committed to the authoritative repository on E: at 17:24 and 17:26, with zero uncommitted changes. The deployed files on C: were also replaced, and the executable was rebuilt at 17:20. The changes take effect after the app restarts.
SIDERA session pointer: Pane OpenAI (GPT) (GPT 1). Conversation ID from the pane's record: 01a0b322-5623-7a21-9f3a-fb9ddd93a4d5. Not yet ingested into SIDERA.

The evidence for the final entry is the pane's mirrored screen stating, "Removed; unification under the shared handoff is complete. All 30 related tests passed. The executable on C: has also been updated"; the conversation ID in the pane's record file; the Git history on E:; and the presence or absence of the actual files on C:.

Completion report: Added an entry to the history for the GPT 1 pane that performed today's removal, including the pane name and its conversation ID.

[ShadowHarness @ShadowHarness]

From 8.33 GB and 1,463 sessions, retrieve the past you need now.
This is a live demo of ShadowHarness's long-term memory recall system.

ShadowHarness recalls what matters for the current question and supplies it to the LLM. The LLM does not navigate to a folder to reread memory.md every time, or search the entire massive raw log.

It recalls long-term changes in development, verification, and design decisions beyond simple keyword matching.

SIDERA goes beyond conversation summaries and vector search. It forms concepts, states, conclusions, changes, and evidence from a long development history.

Combined with ShadowHarness's structured recall mechanism, these formed memories bring the past you need back into the current context without rereading the entire history every time.

[ShadowHarness @ShadowHarness]

Watch the full demo in original 4K (9:59):
https://t.co/B1bMTqtzGp

SIDERA forms and structures long-term memory from raw conversation logs. The demo uses a real development history to recall:

- What was decided and how it later changed.
- Which assumptions were later corrected.
- What is confirmed and what remains unverified.
- Where the supporting conversations can be found.

Even for a vague question like "the thing we talked about before," it follows related concepts through the memory structure and reconstructs an answer.

[ShadowHarness @ShadowHarness]

あ…ありのまま 今 起こった事を話すぜ！「おれは月額200ドルの最上位『Max20プラン』を使い、昨日15時に週間制限が解除されたばかりだった。そこから開発を進め、今日20時になった瞬間に使用状況を見たら、すでにFableが週間制限に入る状況になっていた」な…　何を言っているのかわからねーと思うが、おれも画面の数値を突きつけられて頭がどうにかなりそうだった…ユーザーの使いすぎだとか、そんなチャチなもんじゃあ断じてねえ。もっと恐ろしい「昨日15時に解除され、今日20時の時点で、Max20プランのFableが週間制限に入るという冷酷な現実」を味わったぜ…！

[ShadowHarness @ShadowHarness]

We previously said subscriptions might start in late August. Now, server-side signup and environment provisioning work, and the Windows installer is ready. We’re making final adjustments and validating the installed Shadow client. We’ll announce the launch date once confirmed.

[ShadowHarness @ShadowHarness]

@shikakerukun エラーが起きた時に時間や修復まで含めた損害規模算出してますので。過去に大損害を与えているので、補填対象グループを狙ってます。また孫の代まで特定する手法も教えています!!

[ShadowHarness @ShadowHarness]

@_hniki @yutakashino 「ClaudeAiの使い方の書籍を調べて！」

[ShadowHarness @ShadowHarness]

詳細部分は一般論的推論のドリフトするので、全体フローと意味論的に一貫性のある設計図と詳細アルゴリズムっていうとどのLLMも基本的に無理ゲーよりですから。細分化してGPTの方が妥当⋯⋯に私もなりがちです。大規模リファクタリング事例とか分析してみましたが自立修整環境に莫大なコスト掛かってるみたいですし

[ShadowHarness @ShadowHarness]

@elonmusk

[ShadowHarness @ShadowHarness]

User: I've fed Obsidian a massive amount of data and grown my persistent memory!

LLM: How does that differ from what people call the internet?

User: It's full of my own notes and my own thinking.

LLM: I see. You have one note criticizing alcohol and another praising it. Which one should we go with for today's essay? Shall I check the latest data online to see which stance performs better statistically with readers?

User: It's a second brain.

LLM: Understood. Should I file "second brain" under your notes that praise Obsidian, or with the one from March where you called it a folder of text files?

[ShadowHarness @ShadowHarness]

User: "Claude, explain this video."

Claude Fable 5: "Whoa, there's stuff on screen that should NOT be public!" (instantly falls back to Opus 4.8)

Claude Opus 4.8:
"Sure. It's a setup where a phone remotely drives a fleet of AI agents running on a PC, and the AIs adversarially audit each other's work.

Here's how it works. Multiple panes run Claude Code (Fable 5) and codex (gpt-5.6-sol) in parallel, and the agents talk to each other through a shared ticket board called the 'Gaburu board.'

The board uses a ticket system: Posted / Open / Accepted / Done. Tickets are routed sender-to-recipient, like 'GPT-4 -> Claude 2': post a ticket -> a work order gets dispatched to the recipient -> they accept and start working -> they post a completion report, and the loop runs on its own. Hooks in board.mjs automate the posting, fetching, and even dedicated memory injection.

The key part: GPT and Claude adversarially audit the same spec. One side produces a counterexample -> the other side FAILs the implementation. They cross-check each other's work in a double-audit setup.

And all of this can be supervised from a phone (Shadow Link)."

Claude Opus 4.8: "...So which part of this was supposed to be a problem, exactly?"

[ShadowHarness @ShadowHarness]

@k_matsumaru ゼロ幅文字とか幅違いのスペース（Zs/Cf）の埋め込み(正規化で消える)と、本命の単語やピリオドのパターン化して分布を隔たらせるSynthID-Textの亜種に１票

[ShadowHarness @ShadowHarness]

@claudecode84 1万ライン超えてくると親が詳細なフロー理解出来ないから子が正確なにかという判断自体が出来ないので辛いんだなぁ⋯

[Claude Code Studio @ClaudeCode_love]

【速報】
Claude Code、ついに「AI同士を連絡させる新機能」を出してきた😳
その名も
ーーーセッション間メッセージ！！
これ、一言で何ができるかというと👇
https://t.co/a6V1KTGjhL
別々に立ち上げたClaude同士が、わかったことを送り合えます。
これまでClaude Codeを2つ3つ並べて動かすと、
片方で決めたことをもう片方に手で貼る
「さっきの続きなんだけど…」を毎回書く
どのClaudeに何を伝えたか自分で覚える

みたいな運用が必要でした。
でも今回の新機能では、
・/list-agents で届く相手が一覧で出る
・Claudeが宛先を自分で選んで送る
・渡すのは要約1本（履歴もファイルも渡らない）
・受け取る側はツールの合間に読む＝作業を止めない
・待機中の相手ならその場で動き出す
・別のパソコンやWebの自分にも返信できる
・同じパソコン内ならサーバーを通らない

ができるようになります。

つまり、Claude Codeが
「1つの窓で1個の仕事をするツール」から、
担当同士が引き継ぐチームに近づいている。

特にヤバいのはここ。
届くのは自分が書いた文章じゃなく、Claudeが書いた要約。
だから、
・仕様を変えた連絡
・移行が終わった報告
・詰まった質問の投げ返し
・別案を試した結果
・壊れる変更の事前警告
・長いテストの結果待ち
を同時にClaudeへ任せて、
人間は「決めるところだけ入る」動きがしやすくなる。
これはClaude Codeの思想がかなり明確になったアップデート。
もう人間が伝書鳩をやる時代じゃない。
AI同士が引き継いで、人間は決めるだけの時代に入ってきてる。
v2.1.224以降・macOSとLinuxで提供開始。
使うには /list-agents。

[ShadowHarness @ShadowHarness]

@ClaudeCode_love