Sessions resume across runs (ADR-0017). Without auto-compact the SDK jsonl
grows unboundedly — a 7-day / 26-run session hit 31 MB / 1995 lines, making
every API call resend the full history and inflating a trivial "change a
title" task to 22 minutes. Enable autoCompactEnabled in the SDK settings so
the SDK compacts automatically when the context window fills.
Agent tool downloads and trigger attachment staging both used the SDK
messageResource path, which fails closed for multi-MB teacher files and
did not share the bot-identity transport contract. Route every download
through Hub-owned createFeishuBotCli (secret via stdin, disposable HOME,
HUB_FEISHU_CLI_BIN), keep workspace containment on write, and inject the
adapter in trigger tests.
SubmitDocParserJobAdvance uploads PDFs to Aliyun OSS via tea/httpx, which
defaults readTimeout to 3000ms when RuntimeOptions is empty. Multi-MB
teacher PDFs on para silo failed with ReadTimeout(3000) before the job
could start. Set connectTimeout=15s and readTimeout=5m.
createReadStream emits async ENOENT without a listener, which became an
unhandled 'error' event and exited the silo process. Teachers then saw the
startup "process restart" notice. Wait for stream open and convert missing
files into DocmindClientError instead.
The claude_code preset exposed Agent/SendMessage/Task. Background agents
abort with reason "background", which the SDK maps to Bash
toolDenialKind "cancelled" ("user doesn't want this action") and freezes
command execution mid-run.
Headless Claud agents expose TaskCreate/TaskUpdate rather than TodoWrite.
Fold those tool events into the live card checklist, pass real tool names
and inputs through tool-result, and hide Task* noise once the panel is up.
Native Claude TodoWrite is not registered in headless agent mode even with
--tools default. Add mcp__cph_hub__todo_write (always enabled), mirror the
TodoWrite schema, instruct multi-step runs to use it, and keep the Feishu
progress panel parsing both native and hub tool names.
Unrestricted roles were still passed an explicit --tools name list. The
native Claude binary only reliably registers bundled tools like TodoWrite
on --tools default. Treat role tools JSON null as unrestricted, use the
claude_code preset (→ default) in that case, and keep TodoWrite on allowedTools.
Always expose Claude Agent SDK TodoWrite (todoFeatureEnabled) so multi-step
runs can plan in the open. Parse TodoWrite payloads into a progress panel on
the streaming Feishu card (completed/in_progress/pending) and filter raw
TodoWrite noise out of the tool-use list.
pbank materialize previously shelled out to `unzip` and soft-failed when
the binary was missing, so agents only saw titles. Read zip entries with
Node zlib (store/deflate) and write under workspace .pbank-sources.
Register pbank as an ADR-0027 external capability with org-scoped
username/password envelopes, readiness via /login, and in-process
cph_hub MCP tools (search/get/get_many) that materialize sources under
the run workspace. Extend the capability secret payload for docmind vs
pbank kinds, admin capabilities UI, role tool umbrella `pbank`, and the
pbank-problem-report skill. Credentials never reach the Agent process.
After removing the Typing reaction, add CheckMark on success or CrossMark
on failure so teachers can see completion on the source message without
opening the card.
Teachers convert many PDFs in one tool call with bounded Docmind concurrency.
Each item keeps its own output_dir/document.md and UsageFact; failures are
per-file. Mirror role skills to .cph/runtime-skills and CPH_RUNTIME_SKILLS_DIR
so agents can Read SKILL.md instead of dead .claude/sandbox stubs.
Host egress requires the local forward proxy; sandbox env previously
omitted PROXY vars so Bash/curl timed out on public image URLs. Pass
HTTP(S)/ALL/NO_PROXY (+ lowercase) and NODE_USE_ENV_PROXY from the
trusted service environment into the agent subprocess.
Defaults and silo env go to 150 turns / 1800s wall clock. Run completion
appends a clear Feishu notice for max-turns, timeout, and other failures
(partial answer kept). Startup process-restart kills notify the bound chat.
Release v0.0.38.
Feishu interactive markdown rejects  without image_key
(error 230099 empty/missing imagekey). Always mask residual image md in
card builders; skip img tags with empty keys; skip inline-code examples;
fetch remote images with a browser UA and without env HTTP_PROXY.
Steer the agent: use  for 图文, send_file only for
downloadable attachments.
Release v0.0.37.
Drop markdown_to_pdf MCP surface, implementation, tests, and md-to-pdf dependency.
Roles that still list markdown_to_pdf must be cleaned before startup.
Co-authored-by: Hong Jiarong <me@jrhim.com>
Co-committed-by: Hong Jiarong <me@jrhim.com>
Teachers need ad-hoc Markdown → PDF. Ship an MCP tool powered by
md-to-pdf (Marked + headless Chrome) so remote images/CSS work, with
workspace-scoped basedir, front-matter stripped so untrusted markdown
cannot override dest/basedir/launch options, and MathJax for $/$ math.
Also include WebFetch and WebSearch in the unrestricted role tool
surface by default. Deploy skips Puppeteer's browser download and
expects a host Chrome/Chromium (PUPPETEER_EXECUTABLE_PATH / CHROME_PATH).
The admin role model picker was hardcoded to the env-default model registry
(createDefaultModelRegistry), which only ever returned a single Sonnet model.
Roles could not select any other model regardless of what the org's provider
connection supported.
Replace the env-only model list with a ProviderModelCatalog that:
- Resolves the org's ACTIVE provider connection credential (BYOK or
platform-managed, encrypted via ADR-0024 envelope)
- Calls OpenRouter GET /v1/models?supported_parameters=tools to list
tool-capable models available to that org
- Caches results in-memory with a 5-minute TTL per organization
- Falls back to the env-default registry when no ACTIVE provider exists
The runtime modelRegistry no longer validates role.defaultModel against the
env model list — the admin already validated by selection from the provider
catalog. The env list remains as the fallback for roles with null defaultModel.
The admin roles page loads models independently (non-blocking) so roles
remain editable even if the provider API is slow or unreachable.
Add full skill lifecycle to the org-admin web surface: create, read,
edit, disable. Skills are directories (SKILL.md manifest + supporting
files), content-addressed by SHA-256 in an immutable store.
Backend:
- skillStore: extract commitSkillContent (shared populate→inspect→
dedup→atomic rename); add importSkillFromFiles (in-memory file list
ingestion) and readSkillFiles (read stored version back as UTF-8)
- configuration: add installSkillFromFiles, readSkillFiles, disableSkill
(soft-delete + archive bound role sessions), updateSkillDescription
(label-only, no archival); refactor installSkill to share
commitInstalledSkill
- agentConfigRoutes: wire skillStoreRoot; add GET
/agent-skills/:name/files, PUT /agent-skills/:name (create/replace),
PATCH /agent-skills/:name (description/disable)
- orgRoutes: pass readSkillStoreRoot() to agent config routes
Frontend:
- api.ts: agentSkillFiles, installAgentSkill, patchAgentSkill methods
- SkillEditor.svelte: file tree + text editor + version/description form
- skills/+page.svelte: skill list, create form (generates SKILL.md
template), per-skill editor
- layout: add 技能 nav item
ADR-0018: update Decision to reflect web surface joining host-console
CLI in the shared content-addressed ingestion pipeline.
Spec (AgentRole.lean): unchanged — storage mechanism is OPEN, web
installation is one implementation of it.
白
hongjr03