Upstream kimi-code — Daily Summary (2026-07-30): plugin agents/system prompts, unified host identity, non-blocking TaskOutput

[zicochaos]

Upstream changes (since f8ec3d165, 2026-07-29)

Upstream advanced by 14 new commits (git fetch upstream main; new tip d8f455d69, 2026-07-30). The fork is now 31 commits behind upstream total. New work centers on plugins contributing agents + system prompts, unified host identity, and several kap-server/VS Code fixes.

🔌 agent-core-v2 — plugins contribute agents + system prompts (fa2c5ce18 MoonshotAI#2365, 02d77b20d MoonshotAI#2314) ⚠️

• Plugin-contributed custom agents (feat: support plugin-contributed custom agents MoonshotAI/kimi-code#2365): plugins can now contribute custom agent files; plugin agents awaited before the agent catalog and refreshed on v1 reload.
• Plugin system prompts (feat(agent-core-v2): let plugins contribute system prompt instructions via the manifest systemPrompt field MoonshotAI/kimi-code#2314): the plugin manifest gains a systemPrompt field (plus systemPromptPath to load from a file). A new Session-level convergence point reloads plugin skills into the session skill catalog before fanning out to every live agent prompt; restore no longer re-renders/re-persists prompts (a resumed agent keeps its replayed profile binding). Live refreshes after restart re-resolve the bound profile by name.

🪪 Unified host identity (40172c7ca, MoonshotAI#2382) ⚠️

X-Msh-Platform is now an explicit KimiHostIdentity.platform field (every host declares its own; kimi_code_cli stays the CLI fallback). userAgentProduct → productName. The bootstrap snapshot replaces the flat clientVersion with a required clientIdentity. This is the same hostIdentity domain (from a77ee0382) being wired end-to-end across OAuth, telemetry, and kap-server.

🐛 agent-core-v2 fault-injection removal + cache cold-miss (d36f4c58f MoonshotAI#2399, d10b1c130 MoonshotAI#2395)

• Fault-injection experimental feature fully deleted (IFaultInjectionService, flag, domain, tests, state keys).
• Cache entries missing required fields now treated as cold misses (rebuilt from disk, self-healing poisoned entries).

⏱️ agent-core — non-blocking TaskOutput (691ec4679, MoonshotAI#2379)

Removes the block/timeout parameters from the TaskOutput tool in v1+v2 (kept in model-facing parity) — the model could stall a turn up to 3600s. retrieval_status simplified to success/not_ready; stale callers passing block are silently treated as non-blocking.

📡 kap-server / OAuth / VS Code

• ea81c9a3c (feat(kap-server): expose the managed-account profile at GET /oauth/userinfo MoonshotAI/kimi-code#2363): exposes the managed-account profile at GET /oauth/userinfo (camelCase domain type end-to-end).
• 479403e70/0f3b106c4/6d0a04648: VS Code 0.6.6 release + sign-in message reword + keep sign-in reachable from the no-models screen.
• bc28e9d80/d8f455d69: CI release pass + changelog sync of 0.31.0 (notes secondary model as experimental).

🔁 Reverted

• 1896d1a13 reverts dbb69a267 (kosong "Unsupported image." rejection matching). Net no-op — the two commits cancel.

Impact on this fork

Conflict count rose from 19 to 24 (+5 new). The 5 new conflicts come from the plugin-agent/system-prompt and TaskOutput work landing in test files the fork has also modified:

Severity	New conflict file	Fork's change	Upstream commit	Nature
🟡	agent-core-v2/test/agent/profile/profileOps.test.ts	profile/AGENTS.md test work	02d77b20d (plugin system prompts)	fork's profile test vs. upstream's new prompt-binding assertions
🟡	agent-core-v2/test/agent/loop/loop.test.ts	(fork loop test mods)	691ec4679 (non-blocking TaskOutput)	loop test rewritten for removed block/timeout
🟡	agent-core/test/harness/runtime.test.ts	subagent/swarm harness test	fa2c5ce18 (plugin agents)	harness test updated for plugin-agent catalog loading
🟡	docs/en|zh/{agents,tools}.md	fork docs	fa2c5ce18 + earlier	doc overlap on agent/tool docs

The strategic collision from prior days is unchanged (tracked in #8/#10): the 16-conflict efac96c8a (MoonshotAI#2232 — upstream ported the fork's secondary-model + custom-agentfile feature) is still the dominant blocker. The new 5 sit on top.

Clean auto-merges (no fork overlap): unified host identity (40172c7ca) — note it touches profileService.ts which did conflict, but via MoonshotAI#2144 not today; fault-injection removal (d36f4c58f); cache cold-miss (d10b1c130); oauth/userinfo (ea81c9a3c); all VS Code work; the reverted kosong pair (net no-op).

Recommended follow-ups

• Conflict count now 24 (5 new, all test/doc). The strategic decision still stands and is the gating item: how to handle efac96c8a (feat(agent-core): custom agent files and secondary model on the v1 engine MoonshotAI/kimi-code#2232) feature collision — recommend adopt upstream's secondary-model/custom-agentfile, retire fork's 353a1c573/fb6d57cd2. See Upstream kimi-code — Daily Summary (2026-07-27 #2): secondary_model config API, hostIdentity domain, cloud telemetry #10.
• New 🟡 — profileOps.test.ts: after adopting the sync, re-apply fork's profile/AGENTS.md assertions onto upstream's new plugin-system-prompt test shape.
• New 🟡 — loop.test.ts: fork's loop test mods need porting onto the non-blocking TaskOutput assertions (no more block/timeout).
• New 🟡 — runtime.test.ts: reconcile the fork's subagent/swarm harness test with upstream's plugin-agent catalog-loading changes.
• New feature to evaluate: upstream now supports plugin-contributed agents + plugin system prompts (fa2c5ce18, 02d77b20d). If the fork's plugin story matters, this expands the agent-catalog surface — worth a deliberate check that fork's custom-agent work composes with it.
• Net no-op noted: dbb69a267/1896d1a13 cancel out; no kosong image-handling change to track.
• Carry-forward: full git merge upstream/main on a fresh branch remains the right sync strategy (31 commits, entangled refactors — don't cherry-pick). Run pnpm -r build + agent-core, agent-core-v2, kap-server, node-sdk, kimi-code TUI suites after.

---

Upstream tip: d8f455d69 (2026-07-30). Fork is 16 commits ahead, 31 behind. Conflict set 19→24 (+5 test/doc). Strategic MoonshotAI#2232 decision still unaddressed and blocking. Generated by daily upstream-summary automation.

[zicochaos]

Superseded by #23. The secondary-model/custom-agentfile decision described here was already resolved in #12, and #23 contains refreshed commit counts and the current five-file conflict set.