Subagent / vision subflow prompts (audience=agent) get a dedicated section:
report goes to the parent session, and thinking, step notes and reports stay
in Chinese even when the goal, tool results or web pages are in English.
Prompt regression test covers both the shared hard rule and the new section.
Child reports reaching the user are already humanized: the child session page
uses the same FIX-7 report-humanize path, and the parent's final report resolves
child run IDs to their goal titles; agent.spawn/await tool rows show no raw summary.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- snapshots only for sites with list; feed polling only for RSS kinds;
user data / health checks only for sites with login; search site set
only includes searchable sites; live search / snapshot skip with
reason unsupported instead of hitting the site
- refresh-login / credentials need login, manual snapshot needs list (400)
- POST /sites accepts kind and rssUrl (site + feed in one transaction,
key generated from host); RSS belongs to a site: feeds created without
siteId go to the same-host custom RSS source site, siteId cannot be
cleared
- fix: secret name clash when a site gets a second default-named feed
- command palette regrouped; agent site.list returns kind and
capabilities; system prompt notes RSS lives under sites
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- tool results carry a budget reminder near limits; text before the last tool
call is dropped as report when cut by budget/abort
- usage persisted when a run suspends, so waiting sessions show spent tokens
- task page shows budget_exceeded via business.outcome instead of completed
- budget-exceeded decisions merge by root cause (decision / schedule / goal
template); new summaries added to investigate allowlist
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- browser.screenshot returns the page as an image content block (≤1280px wide, JPEG quality
configurable) for vision-capable upstreams; non-vision upstreams get a redacted text snapshot.
Login pages / password fields are marked sensitive and withheld unless a
browser.sensitive_screenshot grant exists and allowSensitive is passed.
- browser.click_at / type_at / scroll / wait act in screenshot pixels; the server converts via
viewport/DPR, requires automation control and an allowlisted host, and writes a thumbnail
evidence file (DATA_DIR/tmp/evidence) referenced from agent_run_event.
- browser.solve_challenge runs a bounded child agent_run (vision caps only, maxSteps ≤ 12,
budget), judges success from page facts, stops immediately on 2FA / SMS / device confirmation,
and otherwise moves the session to waiting with evidence and a takeover link.
- PiAgentRuntime passes modelImage outputs as pi-ai image content when model.input has image;
the ledger never stores image data; coordinate actions use per-invocation idempotency keys.
- Tests: image codec, runtime image content, E14 e2e (login solved, 2FA stop, step limit,
sensitive withheld). Real camofox 1.17.0 check on example.com recorded in the runbook.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- GET /system/health splits processes into current instances (hosts with
a beat in the last 10 min, newest per host+kind, not stopped) and
retiredProcesses[]; 45s stale and level only consider current ones
- heartbeat writer marks stopped_at on graceful exit (migration 0043),
upsert clears it; pure api role now stops its writer on SIGTERM
- scheduled job system.heartbeat_cleanup (maintain, 1h) deletes rows
without a beat for 24h and samples older than 7 days
- status page: per-role alarms ("io worker 45 秒没有心跳"), collapsible
retired instances; glossary/runbook updated
- bump all members to 0.2.2, CHANGELOG [0.2.2]
Co-Authored-By: Claude Opus 5.5 <[email protected]>
- agent.plan emits plan_updated; system prompt asks for a plan first and updates while executing
- agent.propose_options registers a decision and suspends the run (waitForFact, no worker slot);
the answer or timeout (recommended/stop) resumes it as the tool result
- console-session runs get inline approvals for plan-bound capabilities: approve replays the call
with the approval, reject fails it with the reason and records it on the decision
- proposeFirst blocks execute capabilities until an option is chosen; E12 covers all paths
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Found in v0.2 visual QA (docs/plan/09-visual-qa.md):
- actor strings (admin_session:<uuid>, user:<uuid>, system) -> 管理员/系统/API Key
in policy revisions, subscriptions, inbox, trash, business tasks, organize detail
- cache volume without key shows the downloader name instead of instance:<uuid>
- download timeline/plan: task and organize statuses, plan purpose, priority and
status reason localized; audit gets labels for capability/files/site/feed actions
- organize record reason uses conflict detail (没有可入库的主视频) instead of
blocked:no_main_video; slot/candidate reasons drop inflight:<uuid>
- identity decision summary no longer shows no_viable_candidate
- upstream test: 'Connection error.' classified as network, with a Chinese hint
- new policy draft prefills key default (placeholder looked like a value)
- dialog/sheet close sr-only text in Chinese
Co-Authored-By: Claude Opus 5.5 <[email protected]>
