The cross-harness, self-improving memory layer for AI agents. Join our community: https://discord.com/invite/zdNChS2fBu
Go agent-ai-cliagent-harnessesagent-securitybrowser-extensionciclaude-codecloudcodexcursordetection-engineeringjevknowledge-basememorymobile-device-managementobservabilitysecuritysecurity-information-and-event-managementsecurity-toolstelemetrytraces
docs(changelog): entries for the PRs merged together on 2026-10-22 (#1004)
kyle-elliott-asymptotee92e1e2c5f9f
diff --git a/docs/changelog/cli.mdx b/docs/changelog/cli.mdxindex 4a877be1..f2c943e7 100644--- a/docs/changelog/cli.mdx+++ b/docs/changelog/cli.mdx@@ -11,6 +11,32 @@ Stay up to date with the latest changes to the Agent Beacon CLI.## Unreleased+- **Redacted browser capture no longer stalls on long tokens** · With retention set to redacted, a long unbroken run of letters and digits in a prompt, answer or tool call, such as a base64 blob, made the extension's email scrubber take about two seconds per 64 KB, growing with the square of the length. It now takes well under a millisecond.+- **`beacon login` opens your Windows browser under WSL** · Under WSL, `beacon login` said a browser was opening, showed no URL, and timed out when no Linux browser was installed. It now opens the Windows browser, as do `beacon endpoint connect` in user mode and the dashboard.+- **Reopened dashboard lenses show new events** · A lens closed and reopened on a live session page kept showing the data from when it was first opened; it now loads the session's current events.+- **Browser-extension answers show in traces, including earlier ones** · Answers the Claude.ai and ChatGPT extension recorded since July were stored in a shape traces could not read, so they still did not appear; they now do, earlier records included, and long answers no longer show as raw JSON.+- **Regenerating a candidate keeps its review** · Running `beacon memory candidates create` or `beacon memory evaluations run` again on the same trace reset an approved, rejected or superseded candidate to pending, so `beacon memory skills preview` refused it. Existing candidates now keep their review, and both commands print each candidate's stored state.+- **Cloud uploads no longer lose or resend events when hooks overlap** · When two cloud-agent hooks uploaded at the same time, a slower upload could replace the S3 or GCS object with an older copy, dropping the newest events, and Asymptote Managed ingest could receive the same events twice. Uploads now run one at a time, so the bucket object always ends with the latest events and each event reaches managed ingest once.+- **OpenCode no longer waits on a telemetry backlog** · When Beacon's hook command was slow, each OpenCode prompt, command and tool call waited for every earlier queued event to be recorded first, so a few slow events could hold a tool call for six seconds or more. Each call now waits at most two seconds; the queued events are still recorded in order.+- **Shared traces drop a host-only remote's query** · `beacon share` no longer includes a query string or fragment from a host-only git remote URL, such as `https://git.example.com?access_token=…`, in the shared trace's repository URL or project name.+- **Concurrent memory reviews no longer conflict** · Approving and rejecting the same memory candidate at the same time could let both commands report success, leaving a rejected candidate whose memory was still recalled. Each review now succeeds or fails as a whole, and the slower command reports that the candidate was already reviewed.+- **Pi-family traces show the assistant's answer** · The Pi, Oh My Pi, Prime Agent and Senpi extensions recorded each finished assistant message's reasoning and token usage but not its text, so a live-captured session showed the model's thinking without its reply. The reply is now recorded as `agent.message`.+- **Install rejects unusable OTLP ports** · `beacon endpoint install` accepted `--otlp-grpc-port` or `--otlp-http-port` values outside 1–65535, the same port for both, or a negative `--health-port`, and wrote a collector configuration and harness settings pointing at an address the collector can never listen on. It now stops before writing anything and says which flag to change.+- **Memory sees how long sessions end** · `beacon memory evaluations run` now samples the last events of sessions longer than 2,000 events instead of stopping at event 2,000, and `beacon memory candidates create` cites every event of such a session as evidence.+- **DeepSeek Harness session sync reads session format v4** · `beacon endpoint dsh sync` and the install backfill read the session files DeepSeek Harness writes (`session.v4.jsonl.zstd`, current through 0.2.1-rc.2) with the shapes of an older format. Every tool call came out as `tool.invoked` with empty arguments, every result was dropped (no commands, exit codes, file reads, edits, MCP calls or failures), each session lost its working directory and parent, injected context such as agent instructions and time snapshots was recorded as prompts, and a forked session recorded its parent's history again as its own. Sync now decodes the arguments, joins each result to its call, records the tools a `run_code` program dispatches, counts only the prompts a person sent, records each call's result once when compaction rewrites it, skips a fork's inherited history, and records the tokens of compaction's model call. Sessions already collected are not read again; to re-read them, remove `~/.beacon/endpoint/state/dsh.json` and run `beacon endpoint dsh sync`, which writes their events again under the same ids. See [Reading the session store](/runtimes/deepseek-harness#reading-the-session-store).+- **OpenCode 2.x sessions sync** · `beacon endpoint opencode sync` and the install-time backfill read only OpenCode 1.x's session tables, so every session started on OpenCode 2.x was missed. They now read 2.x's `session_v2` and `session_message`: prompts, assistant text and reasoning, tool calls and `!` shell commands, and token usage and cost, including compaction. Sessions migrated from 1.x are not written twice. A long first sync also saves its progress every few seconds, so an interrupted run resumes where it stopped.+- **Hermes sync reads older state databases** · `beacon endpoint hermes sync` and the install-time backfill failed with `no such column: cwd` on a Hermes `state.db` from before Hermes added that column, `archived` or `messages.active`. Missing columns now read as empty or as their defaults.+- **Browser extension permission banner shows current access** · When site access changed while the extension's popup or options page was open, a slower earlier check could overwrite the newer one, so revoking access could leave the warning that capture is unavailable hidden. The banner now shows only the latest check.+- **Dashboard inventory says when it can't load** · The Agent Inventory page lost its status line in a redesign, so a failed load stayed on "Loading inventory…" and a failed refresh gave no warning. The page again shows the error, and the last scan time once it loads.+- **Browser chat events say when content was cut** · The browser extension trims long prompts, answers and tool arguments or results to fit, but tool records never said so, and the collector kept the extension's truncation flag only in raw attributes, which metadata-only forwarding drops. Browser chat events now carry `field_truncated` whenever any of their content was cut.+- **Browser chat text no longer ends in `�` when cut** · When the browser extension trimmed a long field and the cut fell inside an emoji, the kept text ended in a replacement character that wasn't in the original. The cut now drops the whole character.+- **An MCP tool description that changes back is recorded** · When a Claude Code MCP tool's description changed and then changed back within a few seconds, the runtime log and forwarding destinations kept the changed description as the latest listing, even though the tool was back to its original. The revert is now recorded too.+- **Expanded findings stay on the finding you opened** · The dashboard's Findings page refreshes every 15 seconds and remembered expanded findings by row position, so when a refresh added or removed findings above one you had open, a different finding's evidence appeared in its place. Expansion now follows the finding itself.+- **OpenCode hooks follow `XDG_CONFIG_HOME`** · OpenCode reads its plugins from `$XDG_CONFIG_HOME/opencode` when that is set, but `beacon endpoint hooks install --harness opencode` always wrote the Beacon plugin under `~/.config/opencode`, so OpenCode never loaded it and nothing was captured. Install, discovery and inventory now use the directory OpenCode reads, as `beacon mcp connect` already did.+- **Qwen Code settings follow `QWEN_HOME`** · `beacon endpoint hooks install --harness qwen`, `beacon endpoint discover` and inventory now use `$QWEN_HOME/settings.json` when `QWEN_HOME` is set, as Qwen Code does, including a leading `~` and a relative path. They always used `~/.qwen/settings.json` before, so with `QWEN_HOME` set the hooks went where Qwen never reads them and nothing was captured. A per-user install or repair also writes the resolved `QWEN_HOME` into the scheduled inventory job.+- **curl passwords and cookies are redacted** · Commands like `curl -u admin:password`, `curl -b 'sessionid=…'` and `curl -H 'Cookie: …'` reached the runtime log and forwarding destinations with the password or cookie readable. Redaction now replaces curl's `-u`/`--user`/`--proxy-user` password (keeping the user name), cookie strings passed with `-b`/`--cookie`, and `Cookie` and `Set-Cookie` header values. A cookie file passed to `-b` and prose that mentions "cookie:" are left as written.+- **Upgrades update the Beacon Cloud forwarder's service definition** · The forwarder's launchd or systemd unit was written only by `beacon endpoint connect`, so endpoints connected before v1.3.37 kept a forwarder that logged to `/tmp`, where in system mode any local user could plant a link that root then wrote through. Install, repair and package upgrades now rewrite a forwarder unit that differs from the one Beacon writes, which moves those logs to `~/Library/Logs/Beacon` or `/Library/Logs/Beacon`, and restart the forwarder if it was running. A stopped forwarder stays stopped, and the device key, enrollment and Vector's checkpoints are left alone.+- **Commit links no longer credit a session with files its shell command did not write** · When a shell command contained a word Beacon cannot resolve, such as `$DEST`, `~/backup/` or a quoted `"$TOOL"`, that word was dropped and the words after it moved up, so `cp a.txt b.txt ~/backup/` was read as writing `b.txt` and `sudo "$CMD" rm keep.txt` as deleting `keep.txt`. That could link the session to a commit that changed those files. Unknown words now keep their place and are never reported, and `cp "$SOURCE" output.txt` is now counted as writing `output.txt`.- **Hooks ignore settings a repository could supply** · The hooks no longer take the auto-share settings file from `BEACON_SHARE_SETTINGS_FILE`; they read only `~/.beacon/endpoint/share.json`, so an environment a repository supplies cannot turn auto-share on for its own sessions. `beacon share auto` still honors the variable and warns when the hooks read a different file. The cloud shuttle now reads and writes only under `/tmp/beacon`, so the same environment cannot have it upload a host's runtime log, and `beacon cloud … print-hooks` rejects a `--log-path` outside it rather than writing a log that is never uploaded.- **Devin CLI session history sync** · `beacon endpoint devin sync` reads Devin CLI's committed SQLite session store locally and read-only into the runtime log, adding what the live hooks cannot see: assistant replies, reasoning, per-message token usage, the per-turn model with model-change events, compaction summaries, and `approval.denied` for rejected tool calls. Sessions the hooks already cover get only those additions, so nothing is counted twice, and the install backfill includes Devin CLI. Verified on macOS; the Linux and Windows paths are best effort. See [Devin CLI](/runtimes/devin#session-database-sync).- **Claude.ai and ChatGPT answers show in traces** · Answers the browser extension records as `agent.response.completed` were classified as `other` in trace views, so `beacon endpoint traces show`, the dashboard and shared traces showed the prompt without the answer. They are now assistant messages carrying the answer, as are plans recorded as `agent.plan`. OpenCode's separate completion event, which carries only usage, stays out of the conversation. Local history is rebuilt on its next read, so earlier sessions show their answers too.diff --git a/docs/changelog/sdk.mdx b/docs/changelog/sdk.mdxindex 44d6a961..399a3d6e 100644--- a/docs/changelog/sdk.mdx+++ b/docs/changelog/sdk.mdx@@ -11,6 +11,8 @@ Stay up to date with the latest changes to the Beacon Observe SDK.## Unreleased+- **Early-cancelled streams keep the consumer's error** · Cancelling an observed stream before reading it kept the consumer's error and exported its span only if the source's `return()` rejected; a `return()` that throws synchronously now behaves the same.+- **Global tracers export after reinitialization** · After `Observe.shutdown()` and `Observe.initialize()`, spans from `trace.getTracer()` were dropped because the stopped provider stayed registered globally. New global tracers now use the new provider; a host's own tracer provider is left untouched.- **Observed streams read through their own API keep their span** · An observed function returning a stream that has its own reading API, such as the Anthropic SDK's `client.messages.stream()` read with `finalMessage()` or events, a `ReadableStream` read with `getReader()`, or a Node stream sent through `pipe()`, never exported its span: the SDK replaced the stream with a wrapper that only `for await` went through, which also made `instanceof` checks on it fail. Such streams are now returned unchanged and their span ends when the function returns. Async generators and the Claude Agent SDK's `query()` are still traced until they are read to the end.- **Observe endpoints keep their query string** · An endpoint with query parameters, such as a gateway URL carrying a tenant or routing parameter, had `/v1/observe` appended after its query, so spans went to the wrong route without any error. The path is now added to the URL's path, and the query and fragment are kept. `Observe.initialize()` now throws for an endpoint that is not an absolute URL, instead of exporting to an unusable address.- **Spans from an observed stream nest under it** · Work an observed async generator does as it is read, such as tool or model calls, was recorded under whoever consumed the stream, so those spans appeared beside the stream's span instead of inside it. They are now its children, including work in the stream's cleanup.