The Google wire carries the model in the path, and both parsers assumed that id
contains no "/": the native interceptor's MODEL_API_SUFFIX and the proxy's
googleModelFromPath. A provider prefix relayed verbatim (omp's
`google-generative-ai` provider: /v1beta/models/antigravity/gemini-3.8-flash:
streamGenerateContent) or a gateway route (models/group/hard:generateContent)
therefore fell through to `if (!isModelApiUrl(url)) return send(input, init)` --
the request went straight to the upstream, so the session never reached the
proxy: no window resolution, no compression, no conversation registration
(acp_status answered 404 "unknown plugin conversation") and the proxy-side
model-keyed decisions had no model at all.
Widen the id's character class to [^:?] in both parsers. The method set, the
v1internal control plane and the non-model Google paths (discovery list,
:predict) keep their existing behaviour, so nothing else starts routing.
The judge's I1/I2 only reconcile entry↔exit continuity; the folded region was
a free zone, so a fold that ate ONE half of a tool pair earned a clean fold
verdict right before the upstream answered 400 (tool_use ids without tool_result
blocks). I3 checks the exit body's own protocol structure: every tool-call
answered by a later same-id tool-result, every tool-result preceded by a
same-id tool-call — kernel-parity directional walk (prune.ts #3046),
fold-split agnostic, all four modes (legacy/render-lane/main/loop).
Attribution stays 'whose fault': host-sent half-pairs skipped per-id, a
volume guard arbitrates id re-mints (#2821), compress calls stay exempt
(hideConsumedCompressCalls parity). Pure detection — zero outbound bytes
changed, never blocks, diagnostics.invariantJudge knob as before.
Tests: 12 rows (NEG half-pair folds/render prefix/re-mint laundering;
POS host-broken echo/compress in-flight/whole-pair fold/same-id precedence;
wired main+loop lanes). Existing judge suites (46) stay green.
The embedded bili settings UI drives its iframe by hash route; PAGE_IDS
never listed "blog", so the master-pulled post feed had no entry point
in dsh's own panel even though the framed page (#/blog) already exists.
Owner follow-up: expose it as a sixth tab right after 版本/Versions.
- src/agent/dsh-native-client.ts: append "blog" to PAGE_IDS (after
"updates") + PAGE_KEYS/i18n entries (博客 / Blog). Clicking only flips
the iframe hash — no new data channel. Older embedded frames without
the #/blog route fall back to their default page inside the frame.
- tests/dsh-settings-ui.test.ts: button counts 6→7, zh label list gains
博客, action-button indices shift by one.
Owner review flagged the access-driven-only cadence: with no traffic a
served process would show stale content indefinitely. Every process now
runs a lazy hourly chain — the first touch of the module (the /__bili/blog
route, the status-path warm, or any round) arms a repeating force-kick at
ROUND_TTL_MS intervals:
- content served by any long-lived process ages out to ~1h even with zero
further visits;
- the handle is unref'd so short-lived CLIs exit normally and tests never
inherit a live timer (_resetBlogRefreshForTest clears it);
- force lets a failed round retry on the hour instead of slipping a full
window (a failing round costs ~1 anonymous contents-API request);
- disabled modules never arm.
tests: 2 new cases under mocked clock (~1h/~2h forced rounds without
traffic) + disabled-no-arm; devlog REQ/DESIGN updated.
Access-driven overlay refresh for the /__bili/blog feed (#4587 serves
content frozen at release time): a background round downloads the live
master tree (website/blog/posts/*.md + articles.json entries + referenced
images) from the public GitHub APIs and registers it as the overlay the
feed/asset routes prefer over the bundled bytes; on failure the last good
overlay (or bundled floor) stays up with an "offline" marker.
- src/web/blog-refresh.ts (new): fail-soft round engine — anonymous public
API only, no credentials/config/env surface; single-flight coalescing,
1h TTL between rounds (anonymous rate limit ~60 req/h/IP), per-file
timeouts, size/file-count budgets, atomic registration.
- src/web/blogfeed.ts: remote-overlay seams — pure resolveRepoRel /
collectImageTargets helpers, buildRemoteBlogPosts, registerRemoteBlogFeed,
getRemoteBlogAsset; getBlogFeed merges remote-wins per id over the
bundled floor, asset route checks the in-memory map before disk.
- src/server/admin.ts: GET /__bili/blog fires a fire-and-forget kick after
answering (response always instant).
- src/plugin.ts: /acp status path warms once while COLD (no master snapshot
in-process yet) — same cadence class as warmChangelogOnce; warm processes
keep the status path network-free.
tests: web-blog-refresh.test.ts (12 cases: merge priority, asset split,
image-404 churn, offline degradation, TTL/single-flight, warmer cold/warm
semantics); devlog REQ+DESIGN under devlog/2026-10-27_web-blog-github-refresh.
sanitizeResponsesInputIds rewrote ANY over-64-char input item id to
msg-fix-<hashId> regardless of item type, renaming provider-issued opaque
ids (reasoning rs_*, function_call fc_*, custom_tool_call, item_reference,
compaction) that upstreams with id-shape validation reject with HTTP 400 on
every turn. The length rewrite now applies only to msg-proxy-* ids — the
exact namespace the #242 healing was written for; everything else crosses
ingress byte-identical.
Owner review refuted the injected-private-fetch root cause with unpacked evidence across all published versions (options?.fetch is undefined end-to-end; the OpenAI SDK falls back to the patched globalThis fetch). The one-shot NO MODEL REQUESTS detection stands regardless of the true cause, so its wording no longer names any single confirmed cause. The dsh takeover gate now logs each distinct refused endpoint once per process with the attribution state at refusal time, making the pending runtime repro one-pass diagnosable.
OC 2.0.x (@opencode/cli) rejects bare file paths in the config plugin array (WARN 'configured plugin path must be a directory'; a directory entry loads its index.js), so bili opencode silently degraded every 2.x launch to proxy mode. New opencodeMajorVersion --version probe (cached, failure -> 1) selects pluginDirMode in prepareOpencodeHttpRewrite, which injects a temp wrapper dir re-exporting dist/agent/opencode.js; 1.x keeps the bare file path. Verified end-to-end on @opencode/cli 2.0.3: [plugin] tool acp_status executed via plugin, zero [acp-loop]. Docs: README + agent comment block + devlog now carry the confirmed 2.0.x facts (both hooks fire, ctx.tool.reload exists, editor.add works with Tab+Enter TUI invocation, run mode dispatches no slash commands).
External report (@jensenojs) showed all four next-17444 pre-release observations invert on 2.0.1 stable (model.request fires, ctx.tool.reload exists, context mutation reaches the wire). Independent probes of two adjacent npm dev builds confirmed the 2.x plugin surface changes between builds (one loads V1-only, the other exposes setup() but no ctx.session/ctx.tool). Docs now state facts per-build, the thin-shell rationale is re-based on single-authority + kernel-drift elimination instead of 'plugins cannot mutate context', and a new test pins setup({}) inert-safety on seam-less hosts. No code behavior change.
The V2 kill switch only gated header stamping in httpRequestHook; native tool execute() still forwarded to the proxy via proxyBaseFromEnv() (the launcher always sets BILLION_CONTEXT_PROXY) and session.compaction.ended reporting ignored it too — a half-disabled state inconsistent with pi/omp detectProxyBase semantics (fully inert). Gate all three sites with one pluginDisabled() helper; strengthen the kill-switch test to set the proxy env and assert zero tool forwarding.
omp 17.x mounts extension tools that omit loadMode under xd:// devices
(invisible to the main turn); declare loadMode:"essential" so the four
ACP tools reach the model natively. omp emits no before_provider_headers,
so bind the conversation via the launcher identity register (#162 path)
once tools are ready — bound sessions run in plugin mode with wire
injection suppressed. before_provider_request doubles as the retry
driver. Full native roundtrip e2e-verified against omp 17.3.8.
bili dsh now splices a --patch overlay (~/.dsh-bili/.bili-acp.patch.yml)
into dsh's argv, inserting dist/agent/dsh-acp.js (a cordis command plugin,
same shape as dsh-command-compact) into whatever profile boots: a real
/acp slash command on the web/tui surfaces.
- src/agent/dsh-acp.ts: /acp → status panel (fallback=latest), else
armed-but-idle info (#235 wording), else unreachable error
- src/agent/shared.ts: +fetchStatusLatest (conversationId=dsh placeholder;
server rejects the empty string)
- src/launcher.ts: +writeDshAcpPatch (independent of the settings rewrite,
exists with zero custom providers) +dshArgsWithPatch (prepend / after
'web' / skip for 'plugin' & --dump-default-config)
- tests: launcher 85 (incl. argv assertions in runLaunch dsh e2e),
tests/dsh-acp.test.ts 4 outcomes; 610/610
Sanitize over-long input item ids before the projection reads them so
conversations stuck on 400 recover on the next request. Complements the
root-cause fix (hashId remapping) in this branch.
- setupMitm: new allowRemoteClients flag; remote CONNECT accepted only for
whitelisted model hosts (blind tunnels stay loopback-only — no open relay)
- server: derive the flag from --host; honest 0.0.0.0 display + [security]
warning on non-loopback binds
- tests: #77 case extended to the full remote policy matrix
- docs: README/CONFIGURATION remote-agents section (en+zh)
A 404 from the status endpoint on a live proxy means the conversation
has sent no model request yet (e.g. /acp right after bili pi starts) —
not an error. Probe the manifest for liveness+version and show an info
notice; keep the old warning only when the proxy is unreachable.
launcher pi branch prepends pi -e <dist/agent/pi.js> only when the user
has NOT run plugin install pi (mutually exclusive: pi dedups extensions
by resolved path, packages entry vs -e file would double-register).
Uninstalled users now get the native tool panel and /acp out of the box,
matching bili omp / bili opencode behavior.