meshbay.git/CLAUDE.md, branch 0.13

docs: remove the documents MESHBAY_DESIGN.md replaces

2026-09-10T22:19:06Z

Twenty-four files, about 17 000 lines: the two architecture drafts, the three security reviews, eleven design notes, the roadmap, the decisions file, the v1–v4 archive, the deprecated user guide and the stale quickstart. Their content is in MESHBAY_DESIGN.md, and git history holds the originals. The reason to delete rather than keep bannered: a document that is superseded but present still gets read, and a reader cannot always tell which of two accounts of one mechanism is the live one. That was the argument for retiring the user guide rather than repairing it, and it applies to the whole set. What made this safe is the concordance. Roughly 290 comments and docstrings cite these files by section — `musicbay.md §6`, `mediacenter.md §5.5`, `draft-v6 §2.11` — and section 16 maps every one onto its replacement, so not a single comment needs editing to stay followable. It now says plainly that the files are gone and where to recover them, and it gained rows for the three reviews (their findings are section 13), and for the two guides. Four kept documents pointed into the set and were repointed first: `playlists.md` (nine references — it is a live proposal and must not dangle), `WINDOWS-PORT.md`, and CLAUDE.md's example. No dangling reference remains outside section 16. Two files were dropped from the list after checking what they hold. `HTTPS.md` is an operational runbook — Caddy, certificate renewal, DNS, troubleshooting — and MESHBAY_DESIGN.md deliberately covers no operations, so nothing would replace it; the versioned Caddyfile is the config, not the procedure. `cast-smart-tv.md` is the plan for the unbuilt DLNA phase of a feature whose first two phases ship, and section 11.4 summarises it in four lines rather than carrying the SSDP/UPnP work. There is no user guide now, and section 0.1 says so rather than leaving a reader to discover it. Suites green: 2258 passed, 4 skipped. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01YVoHVCcfBqud6ZjG4db3y7

docs(claude): cut the architecture, keep the lessons

2026-09-10T16:45:15Z

CLAUDE.md was half a second specification. It carried summaries of three security reviews, the invite redesign, per-node identity, the desktop client, a protocol-version history and a sixty-row module table — all of it now in MESHBAY_DESIGN.md, and some of it wrong: an errata list asserted a keystore parameter that had been raised months earlier, and the module table pointed at two implementations that no longer exist. An errata list beside a specification is a second specification, and the older one wins by being read first. Those sections become a pointer table naming which part of the design document answers which question. What stays is what has no other home: the conventions, and the engineering lessons — the ones that are not deducible from the design because they are what the code and the platforms actually do. They keep every word. The module table stays as locators, stripped of the design prose it duplicated, with a note kept only where it is a rule about editing the code. Every path in it was verified to exist; two were wrong, and the harness directory was wrong throughout. Two references to files outside the repository are gone from the reference table — a document about this repository should not send a reader somewhere they cannot follow. The rules about that directory stay, because they are rules about what must never be committed. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01YVoHVCcfBqud6ZjG4db3y7

feat(hub): pin the app controls while a library scrolls

2026-09-10T08:29:12Z

Files, Videos, Music and Photos are read by scrolling, and everything that steers that reading left with the first screenful. Three bands now pin under the navigation bar, in a group and on the Search page alike: the tab bar (the search field, on Search), the application's own toolbar, and the file table's column heads. The group's name and description still scroll — they say nothing a reader needs while walking a directory, and the height they would cost is height the list does not get. A band's offset is the heights of the bands above it, and those are not constants: the toolbar wraps to three rows on a phone, grows a field while a folder is being named, and loses its filter on Search. So each band measures itself and publishes `--chrome-h` / `--toolbar-h` (static/sticky.js) and the stylesheet does the arithmetic in calc(), rather than a number written down twice — the fault CLAUDE.md already records against this layout twice over. A band publishes height *plus its own bottom margin*, and paints that margin as a ring of page colour, so the pinned layout is pixel-identical to the flow layout and nothing shifts at the moment a band pins. Three overflow faults came out of it, all of the same class and all of them what "the header does not stay" actually meant on Android — a document wider than the screen leaves everything pinned attached to a viewport the reader can no longer see, the navigation bar included: - a directory's name cell was a bare , so an unbreakable folder name (`Rage_Against_The_Machine_Discography_1992-2000_FLAC`) set the column's minimum: a 527px table in a 390px window - Search's group column did the same at 442px with an underscored group name. It also goes entirely below 768px, where there is no room for it and the breadcrumb already names the group - the shared-directories table has four columns of controls with a combined minimum near 440px, none of it compressible. On a phone the row stops being a row: the name and its eject/remove pair on one line, the two switches — each carrying the column head's own string as a label — on the next - and, found by measuring at 360px, the tab bar itself was 19px too wide `.file-table` moves to separated borders: a collapsed border belongs to the table rather than to the cell, so the column heads lost their rule the moment they pinned. Measured, not read. tests/harness/sticky_header_probe.py drives the shipped GroupPage and SearchPage against a stub node, walks to each application, scrolls to the end and reports every rectangle — 11 views x 4 widths x 2 engines. Its fixture says what real data says: the first version used `note-007.txt` and `un groupe`, which fit any screen, and found none of the above. A fixture narrower than real data tests the fixture. Also: `test_desktop_shell` no longer looks for the CSP after the first `-->`, which made it fail on correct markup as soon as a comment was added above it, and `search-page.js` joins test_hook_ordering's file list. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01Tx16FhyD2BUdpooGb5jcyN

docs: three lessons from the download freeze

2026-09-08T21:22:39Z

The service worker one is the expensive one: an idle worker is terminated and a streaming response does not count as activity, so `writable.write()` hung with no error anywhere while the node stayed healthy. Written down with what made it unfindable — a stress probe fast enough to hide it, three measurements that correctly cleared everything they touched, and an empty console that was the evidence rather than the absence of it. The other two are the framing headers (three of them, all visible in one `curl -I`, fixed one at a time over an afternoon) and `encodeURIComponent` leaving the apostrophe that RFC 5987 uses as its delimiter. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01HCGdheDLxGReuKHga3BtST

feat(chat): Tier 2 — a member verifies another member's device itself

2026-09-07T19:04:56Z

Chat messages have been signed by the sending device since MNP 2.0, but a reader had no way to know that the device belonged to the account the node named: the signature proved *a device*, and `sender_id` was still the node's word. This closes that for any account a client has already seen. **What was blocking it was not effort — the evidence was not being kept.** `_do_device_add` verified the countersignature that admits a second device and stored only `added_by_pk`: *which* key approved, never the proof. And `device_add_transcript` binds `nonce_node`, the approving connection's handshake nonce, so even a stored signature was unverifiable by anyone who had not been on that connection. `identities` gains `add_sig`, `add_nonce` and `add_ts`, added before the migration's early return — which fires on every roster widened since 2026-08-18, i.e. all of them, so putting them inside it would have meant they never arrived. `group_roster_req`/`resp` relays, sealed under a new groupbox purpose and answered to **any member of the group**, every live device of every active member with the evidence that admitted it. The node decides nothing: it hands over evidence and the client walks the chain from each account's root outwards (`_verifyRoster`). That is deliberate — the node is the party the property holds against, so it is not asked to assert trust. Two holes the tests caught while this was being built: - "no signature" was being treated as a trust root, so a node that writes the roster could put any key in an account's row and have it laundered straight into the verified set. A root is a device that names **no** countersigner. - pinning only the verified subset at first sight raised "key changed" on legitimate second devices whose countersignature predates this change. First sight pins everything the node says, because that is what trust-on-first-use means and an alarm that fires on normal events stops being read. The property, and it must not be rounded up: **once a client has seen an account, a node that later substitutes a key for it is detected. Nothing is gained at first sight**, where there is nothing to compare against — the same boundary `per-node-identity-v1.md` draws, unmoved. The cost, stated because it is real: the roster is member-visible, so every member learns how many devices the others hold and their public keys. It stays inside the group, the hub is not involved, and it is scoped per group. A member who cannot see the keys cannot check them. User-visible surface: one notice, "this account is using a key you have not seen before", in ten languages. Nothing else. 16 tests — 7 on the node (the evidence is stored, it verifies from the roster alone, a fabricated device carries none, another group's members are not disclosed), 9 running the shipped `_verifyRoster` under node against rosters built by the shipped Python: a chain of three in any order, a signature by the wrong key, one for another node, one for another account, and two fabricated devices signing each other admitting nothing. Tier 3 (operator-signed roster attestation) stays deferred, with nothing depending on it. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01TZZxYjz8YeWRz13xDi8LJr

Merge origin/main into the chat encryption work

2026-09-07T16:03:52Z

Both sides landed a breaking MNP change and both called it 2.0, which is right: the sealed upload, the removal of `stream_seg` and mandatory chat encryption share one flag day. They are recorded as one version in `__init__.py` rather than as a race between two. The resolutions that were decisions rather than mechanics: * **`MNP_MIN_SUPPORTED` moves to "2.0".** The sealed upload alone was a *confined* break — a 1.x peer could still connect, browse, download, stream and chat, with only its uploads refused by `upload_not_sealed` — so the floor deliberately stayed at "1.0". Mandatory chat encryption ends that confinement: a 1.x peer can neither produce a sealed chat message nor read one, so it would connect, look fine, and be unable to say anything. Refusing it at the handshake is the honest form. The per-message `upload_not_sealed` path is untouched and still right if the floor is ever lowered. * **`sendChat` throws on an `error` reply**, from origin, applied to the sealed send. It matters more after this change, not less: the node now refuses a stale epoch, a malformed envelope and a device claim that is not the connection's own, so there are three new ways for a message to be rejected and none of them may look like a message that was sent. * **`req_id` supersedes the per-type routing** this branch added for `chat_keys_resp` and `device_hello_ack`. Both blocks are kept beside the existing `chat_hist_resp` one, for the same stated reason — a node too old to stamp — and their comments no longer claim to be the mechanism that closes the class. `req_id` is. * **`chat_send_probe.py` is rebuilt on origin's structure**, not beside it: two scenarios, a stub that stamps `req_id`, `music_meta_req` as the older pending request. The encrypted path is layered on — a real Ed25519 device key generated in the page, and a `chat_keys_resp` sealed by the shipped Python, because a payload the page built itself would prove only that the page agrees with the page. * **`test_reply_correlation.py` now sends a sealed message.** Its subject is which of the two messages leaving that handler carries the id; plaintext chat was only the fixture, and the node refuses one now. * `groupbox` keeps both new purposes (`upload`, `chat_keys`); `protocol.py` keeps origin's removal of `STREAM_SEGMENT` and this branch's correction of the "Double Ratchet message" comment on `CHAT_MESSAGE`, which was wrong when it was written and is wrong differently now. Full suite on the merged tree: 1993 passed, 11 failed — the same 11 that fail on a pristine checkout (2 Windows service tests, 1 apps-enabled policy, 7 transcode tests that pass in isolation, and the WebRTC invite test that hangs on its own). Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01TZZxYjz8YeWRz13xDi8LJr

feat(chat): encrypt group chat under per-device epoch keys (MNP 2.0)

2026-09-07T15:50:28Z

Chat messages are sealed with AES-256-GCM under a key derived per group, per epoch, per *device*, and signed over the ciphertext with the device key the node pinned. The node relays and archives; it cannot read a message. There is no switch. MNP goes to 2.0 and MNP_MIN_SUPPORTED moves with it, so a 1.x peer is refused at the handshake with `version_too_old` rather than admitted and then unable to speak. An opt-in flag was designed and rejected: every node is a test node, so it would have bought nothing and left a plaintext branch reachable — C6's lesson one feature later. A test reads the source and refuses any code that consults a `chat_encrypted` setting. Not Sender Keys, and `senderkeys.py` is now documented as unused. With distribution under the group key and a node that serves history to devices which were not present, the node must retain each chain's earliest key, and a chain key at iteration i yields every message key from i on by pure HKDF — forward secrecy is zero either way. What the ratchet was left buying was stateful client code with silent failure modes, three of them reproduced: any member could sign as any other, a second device dropped the first's chain, and the skipped-key cache grew without bound. The reasoning is in docs/chat-sender-keys.md, which is the specification and the decision record. Epochs, not rotation: the epoch key is wrapped under the group key at delivery and never stored under it, so `gek_rotate` is a re-wrap. A group-key-derived archive key would have made every message ever sent unreadable on the first `member unpin`, which is the documented step after removing a member. A new epoch opens on member revoke/unpin, device revoke and `gek_rotate`; old epochs are kept and still delivered, so history stays readable to everyone who could already read it, and nothing anywhere deletes one. Three prerequisites this needed, each a live defect on its own: * The peer registry was keyed by user_id, so one account's second device evicted the first and the broadcast skipped recipients by account — a person's phone never saw what they typed on their laptop. * The handshake authenticated an account, never a device. `device_hello` (additive, signed, refused unless the key is a live device of this account in the node's own roster) is what lets the node refuse a member claiming somebody else's key. * `_admin_exec_file_delete` authorized against the exact uploading key, so device linking had already broken deleting your own file from your other device. It now authorizes against any non-revoked device of `uploader_id`. Found by driving the real panel over the real transport, not by reading source: `chat_keys_resp` was routed by arrival order and handed to an unanswered `media_meta_req` — the original frozen-tab defect in a message type that did not exist when that probe was written. And `_asText` had been deleted with an unrelated helper beside it; its only caller sits inside a promise the panel catches, so every conversation rendered empty with nothing in the console. Existing node data is migrated by QE/migration/migrate_chat_encryption.py (not versioned, per the QE rule), run with the node stopped. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01TZZxYjz8YeWRz13xDi8LJr

fix(mnp): give a reply an id, so it stops being routed by luck

2026-09-07T14:02:10Z

MNP carried no correlation id. A reply named its own type and nothing else, so a client with more than one request in flight worked out which one a message answered from the message itself — and for the replies that name nothing it could not. `_dispatch` fell through to matching by arrival order, which is a guess. `_sendAndWait` had the right value all along: it keys `_pending` by `this._seqId++` and never put it on the wire. The guess fails asymmetrically, which is why it hid. The victim is not the request that was answered wrongly — it is the unrelated one that now waits out its own 30s timeout for a reply already delivered elsewhere. Live on 2026-09-06: five `music_meta_req` sat pending for over 100 seconds behind a failing MusicBrainz, and a `device_list_result` was handed to one of them. The composer is disabled while a send is in flight, so a chat message whose reply went astray the same way left the Chat tab looking frozen for thirty seconds, then unfroze on its own. The `ack` half of this was fixed on 2026-08-30 by matching on request type. That closed the instance and left the class open: a refusal has no type to match on either, and `_dispatch_message`'s catch-all answers every unforeseen failure with `{"type": "error", "detail": "Request failed"}` — 238 of this module's 240 error sends name nothing at all. `req_id` now rides on the request and comes back on the reply. On the node it is published for the whole handler in a ContextVar and stamped by `_send`: a parameter would have meant threading an argument through all 240 send sites, and asyncio copies the context into a task, so a handler that `_spawn`s its real work still answers under the right id. It is never stamped on a broadcast — those answer nothing, and the owner check in `_send` is what keeps a chat broadcast or an index push from reaching another peer looking like a reply. On the client, `_dispatch` resolves on `req_id` first and the arrival-order fallback is gone the moment a node proves it stamps (`_correlates`, armed by the handshake's own reply). The fallback stays for an MNP 1.0 node, unchanged and no wider: there it is the only thing there is, and removing it would leave device_list_result, join_result and the handshake replies reaching nobody. Two things fall out. `sendChat` refuses an `error` reply like every other request in the file — it returned it as success, which did not matter while a refusal reached the wrong caller anyway and would now show a rejected message as sent. And `_group_ctx` uses `.get`: a reload pops a removed group while sessions connected to it are open, and every request they had left raised KeyError into that same catch-all. Sealed index messages are the one exception to the fast path. They cannot be handed over until they are opened, which is asynchronous while `_dispatch` is not — resolving on the id alone gave `fetchIndex` the envelope and skipped `onIndexSync` entirely. Caught by extending `index_seal_probe.mjs` to stamp a reply the way a current node does, after the hub suite passed over it: the probe built its own frames and had never seen one. Tests, all failing before and passing after: `test_chat_send.py` drives the real ChatPanel over the real transport for both shapes of reply with an older request pending (3 of its 6 are new, and the 3 for `ack` pass either way, so it discriminates); `test_reply_correlation.py` pins the node's half — the refusals that name nothing else, the broadcast that must not be stamped, and a late reply from a spawned task answering under its own id rather than the most recent request's. Full suite: 1897 passed, same 11 pre-existing failures as before. QUIC keeps its own dispatch and is not stamped. It is disabled by default and no browser request reaches it, but the asymmetry is real. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01Dn1xYx9uT69mCB6UDvyKAN

feat!: MNP 1.0 — seal index and handshake_ack under the group key

2026-09-03T14:16:55Z

`index_sync`, `index_delta` and the `handshake_ack` config payload now travel sealed under a GEK-derived subkey (`meshbay_common/groupbox.py`, mirrored by `sealGroup`/`openGroup` in `crypto.js`). Only `type`, `v`, `group_id` and the ack's `node_pk`/`proof`/`sig` stay in clear — a receiver must route and authenticate before it would trust a decryption. Verify, then decrypt. The ack line is integrity, not confidentiality: the signed handshake transcript names no ack field, so `is_node_admin`, `enabled_apps`, `video_root` and the rest were authenticated by the DTLS channel alone. The index line is defence in depth against a repeat of C1/C6 — a peer served before the handshake completes now gets ciphertext, not filenames. Nothing against an observer, the hub, or a member; that is the whole claim. `index_progress` stays clear (D3, counters only). Chat is out of scope. Failure is fatal: a payload that does not open ends the session naming the message type — never an empty index or an empty `enabled_apps`, both of which are legitimate states. Version negotiation ships here too (phase 15.6, brought forward): `v` + `v_min` on `handshake` and `handshake_challenge`, refused with `version_too_old` / `version_too_new` / `version_unreadable`. The flag day was already being paid for; the next breaking change now costs a refusal message. BREAKING CHANGE: breaks the WebRTC wire every deployed client speaks. Hub and every node must deploy together; the SPA is served by the hub, so a browser picks up the new client on reload. See MESHBAY_NODE_PROTOCOL.md §11.1a, §13.1. Co-Authored-By: Claude Opus 5 Claude-Session: https://claude.ai/code/session_01HkzbhmMmK8PqQBtGz5zCvY

docs: Windows port audit and sender key distribution decision

2026-09-03T13:34:33Z

Add docs/WINDOWS-PORT.md with the full portability audit (what is already portable, what blocks, implementation plan W1-W7). Reverse structural decision 20: sender keys are distributed GEK-wrapped, not pairwise to identity keys. The GEK is the group secret; files and chat share the same access boundary. Per-device chains (15.0b) remain required for correctness. Co-Authored-By: Claude Opus 4.6