3 Commits

Include renames

Author	SHA1	Message	Date
Hermes@mac13	ef2e060c95	fix(subagent): emit an explicit mode=unknown so the aside open survives both gates ... The `subagentchat` aside open failed silently: the host's `parseSubagentChatAddress` (the tab type's `canOpen`) rejects a MISSING mode (`get("mode")` → `null`), so a mode we could not stand behind — omitted to dodge the controller's 401 — made `claim` refuse the address (`sidebarRight: tab type "subagentchat" refuses ...`), and the right column never revealed. Fix: the codec now normalizes an unknown/undefined mode to an EXPLICIT `mode=unknown` instead of dropping it. That value passes both gates: - client: `parseSubagentChatAddress` accepts `unknown` → `canOpen` true → claim accepts; - server: `validateAddress` skips the compare when `mode === "unknown"` → no 401. A real (one-shot/continuable) mode is still emitted when the host catalog is `ready` and has the child; only an honest `unknown` is sent when it is not. Drops the TEMP `diag.append` probe route + the client before/after diag fetches (their mission is done) and rewrites the two spec cases to assert the explicit `mode=unknown`. Verified with the host's real `parseSubagentChatAddress` pulled verbatim from the app bundle: the old no-mode address → canOpen false (reproduces the refuse), the new `mode=unknown` address → canOpen true. typecheck / lint / 29 subagent+sidechat specs green.	2026-11-06 14:07:33 +08:00
figmar	061964e3e8	fix(subagent): stop 401-ing the aside-open with a guessed child mode ... The aside button could hit `subagent/unauthorized: subagent mode does not match the supplied address`. The host's session-controller (dsh-api-session- controller validateAddress) compares the `mode` in the resource address against the child's durable descriptor mode and rejects any mismatch. The address `mode` came from the row, and on the degraded path the row's mode is a best-effort guess from the summary `running` flag (running => continuable) — so a child that is genuinely one-shot but still running was guessed as continuable and the very open it was meant to trigger 401'd. Two layers, so it holds on both the real-catalog and the degraded path: - subagentChatAddress: only pass a mode through when it is one of the two durable values (`one-shot` / `continuable`). Anything else is omitted, which the controller reads as `mode: "unknown"` and therefore skips the comparison — the safest form when the caller does not actually know. - openChildAside: prefer the host's real descriptor mode for the child (the `subagentsByParent` entry carries the durable mode) over the address mode, falling back to the address only when the real catalog never arrived. Gates: typecheck / lint / build / test 132 files 1392 passed.	2026-10-14 10:18:58 +08:00
figmar	6b4a29ba24	feat(subagent): open a child's chat in the native right sidebar ... The host's own catalog puts an "open beside" action on every non-current child row; this port had never wired it. Add it: the row grows a hover button that opens the child's chat as a `subagentchat` resource in the native right Sidebar, so a user can watch a child work without giving up the pane he is in (the in-pane jump stays the plain row click). - src/client/native/subagent-chat.ts: the resource-address codec, copied from @deepseek-ai/dsh-client-ui-subagent rather than imported (the package exports no helpers and the plugin must not reach into it). - native/surface.ts: carry `kind` / `preferNewPane` on an open. - SubagentView: `openChildAside` via `ctx.get?.('sidebarRight')` — an optional probe, since a fake ctx without `get` otherwise crashes the whole render (9 specs caught it) — threaded through all three CatalogRows call sites. - locales: `subagentOpenAside` in zh/en plus all 20 third-language dictionaries (tests/locales.spec.ts enforces key-set parity). - docs: the conventions page now documents the resource-open recipe. Gates: typecheck / lint / build / test 132 files 1390 passed / check:consumer-types, plus the repo's runtime-contract symbol check.	2026-10-14 09:32:16 +08:00