Repository navigation
docs(integrations): Cognee tracing (TH-8328) - #1017
Merged
Merged
Conversation
approved: Nikhil 2026-10-03 blanket
…8229) Rework the Cursor hook recipe into the coding-agent family foundation: - examples/coding-agents/hook-to-otlp: stdlib-only (Python 3.9+) hook command. One session = one trace (trace id derived from the session id), session -> turn -> tool/subagent/llm/event spans, start/end pairing via a locked per-session state file, vendor durations when no start hook, content off unless FI_HOOK_CAPTURE_CONTENT=1, always exits 0 with the response the host expects, network send in a detached child. - Pluggable adapters (adapters/<app>.py, contract in adapters/base.py); Cursor adapter mapped from https://cursor.com/docs/hooks, observation hooks only (no permission hooks). - replay.py for driving the emitter with recorded payloads; unit, CLI and Cursor contract tests. - Integrations > Coding Agents nav group with the shared page, the Cursor page, and one marker comment per remaining child. The cookbook page is replaced by the integration pages. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ssions A real Cursor 3.24.18 run showed sessionStart/sessionEnd firing for chats that were opened and closed without a message (sessionEnd with duration_ms 0, final_status "none"). Exporting the root at session start would create an empty trace per such chat. The root session span now rides along with every batch that exports a child span (same id and start time, so the latest export wins) and is sent at session end only when the session had activity. Adds the captured Cursor 3.24.18 sessionEnd payload as a contract fixture. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add an OpenCode plugin that exports one Future AGI trace per OpenCode session from documented plugin events and hooks (session.created/updated, message.updated, message.part.updated, session.idle/error/deleted, chat.message, tool.execute.before/after, dispose): opencode.session (AGENT) > opencode.turn (CHAIN) > chat <model> (LLM) > execute_tool <tool> (TOOL) Token usage, cache tokens, cost, finish reason, tool status and timings come from OpenCode's own message/part payloads. Prompt, answer, tool argument, tool output, title and error text are dropped unless FI_CAPTURE_CONTENT=true (then capped at 4 KiB with key patterns redacted). Zero npm dependencies; never blocks or rewrites tool calls; export failures never break a session. Includes contract tests against an event fixture recorded from a real OpenCode 1.19.0 run, run_e2e.sh, the docs page (incl. why OpenCode's built-in OpenTelemetry is not used), and the Coding Agents nav entry. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…288) Findings from the real E2E against a local Future AGI: - Retry each failed export once with the next flush (and twice-flush on dispose), so a slow or briefly unavailable collector does not drop a turn. Default export timeout 10 s, matching the OTel exporter default. - run_e2e.sh closes stdin: `opencode run` reads stdin until EOF when it is not a terminal, which hangs scripted runs before the session starts. Documented on the page with the exit-delay note. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ng (TH-8229) - Cursor page: Cursor reports model "default" when no model is selected; chats opened and closed without a message are not sent. - run_e2e.sh: space hooks 1s apart so the fixture's reported tool durations fit inside the turn, as in a real session. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…er session
Additive core changes requested by the Kiro/Antigravity/Devin Desktop
children; existing adapters (Cursor) are unaffected.
- A TURN START with an empty key opens turn-1, turn-2, ... per session;
a keyless TURN END/UPDATE applies to the open turn (ignored if none).
- A keyless END of a tool/subagent/llm/event span pairs with the oldest
open keyless START of the same scope and name (FIFO), for pre/post hooks
without call ids.
- A TURN START closes any other open turn of the session (and its
children) as incomplete; a late END for that turn re-sends it complete
with its original start.
- respond(): an adapter returning "" prints nothing (Kiro and Claude Code
add hook stdout to the model context); None still means "{}".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e (TH-8229) A real IDE agent session on the free Hobby plan showed payload details the hooks reference does not document: - afterAgentResponse and stop carry input/output/cache-read/cache-write tokens for the turn: recorded once on the "Agent response" LLM span as llm.token_count.*, and on the turn as cursor.usage.*. - afterAgentThought fires twice per thinking block, once with a generation_id suffixed "-<n>-<4 chars>": the suffix is stripped (no phantom turns) and the pair maps to one span id. - sessionStart does not fire for IDE chats; sessionEnd fails on window close. The Cursor page now says so, reports that agent hooks run on the Hobby plan, and covers the free-plan "named model" error turn. Adds the sanitized real turn as a contract fixture. Core: the default FI_HOOK_TIMEOUT is now 10 s (the send runs in a detached child, so the host never waits; 3 s timed out against a loaded collector). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
OpenClaw's official diagnostics-otel plugin can target any OTLP/HTTP
endpoint with custom headers, so N-04's unsupported-page branch does not
apply: this is a configuration-only recipe. A config patch points
diagnostics.otel at Future AGI (exact tracesEndpoint, X-Api-Key /
X-Secret-Key via ${VAR}, traces only, captureContent false), and
~/.openclaw/.env carries the keys, OTEL_RESOURCE_ATTRIBUTES for
project_name and OTEL_NODE_RESOURCE_DETECTORS=env so the command line
(which holds the prompt for `openclaw agent --local -m`) is not exported.
run_e2e.sh runs one turn with OPENCLAW_HOME in a temp dir, via --local or a
Gateway (GATEWAY=1). Tested end to end against a local Future AGI.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… id (TH-8229) Real Cursor 3.24.18 postToolUse payloads send tool_use_id as "call-<uuid>-<n>\nfc_<id>". The first line becomes the span key and gen_ai.tool.call.id; the provider function-call id is kept as cursor.tool.provider_call_id. Covered by the real-turn fixture test. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Adds a zero-dependency Pi extension (examples/coding-agents/pi) that turns Pi's documented extension events into one OTLP trace per Pi session: pi.session > pi.prompt > pi.turn > chat <model> / execute_tool <tool>, with model, provider, token counts, stop reasons, tool call ids and error status. Prompt, answer, tool argument and tool result text are exported only with FI_CAPTURE_CONTENT=true. The extension only observes notification events; it never blocks or rewrites tools or provider requests and never reads Pi's session files. Contract tests replay an event stream recorded from a real Pi 1.2.0 run. Page: src/pages/docs/integrations/coding-agents/pi.mdx, nav entry after the coding-agents:pi marker. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…/th-8288-opencode Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…8227) Claude Code's beta trace export (CLAUDE_CODE_ENHANCED_TELEMETRY_BETA=1) sends one claude_code.interaction trace per prompt. collector.yaml (otelcol-contrib transform) labels the spans for Future AGI (agent / llm / tool / chain), maps token counts onto gen_ai.usage.*, copies opted-in prompt and tool content to input/output, drops user.email, and sets compression: none because the Future AGI collector rejects gzip bodies. - page: setup through the collector, direct export as the minimal path, what is and is not captured, privacy flags, troubleshooting - examples/coding-agents/claude-code: collector.yaml, settings.json, run_e2e.sh, contract tests against real Claude Code 2.1.300 exports Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Adds a project-local omp hook that turns Oh My Pi's documented hook events (session_start/switch/shutdown, agent_start/end, turn_start/end, message_end, tool_execution_start/end) into one OTLP trace per omp session: omp.session > omp.prompt > omp.turn > chat <model> + execute_tool <tool>. Only the omp.sh fork is covered (binary omp, @oh-my-pi/pi-coding-agent, can1357/oh-my-pi). The hook only observes: it never returns from a handler and does not subscribe to tool_call, tool_result, context, input or before_provider_request. Prompt, answer, tool argument/result, intent and error text are exported only with FI_CAPTURE_CONTENT=true (4 KiB cap, key-pattern redaction). Exports are batched per turn and flushed within omp's 2 s session_shutdown budget. Tested end to end with omp 18.11.0 (print mode and the interactive TUI) against a local Future AGI; 12 contract tests replay a fixture recorded from a real 18.11.0 run. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…(TH-8288) `opencode.directory` holds an absolute local path, which usually contains the OS user name. It is now exported only with FI_CAPTURE_CONTENT=true, like other content. The contract tests assert that no path reaches the default export and that the opt-in still carries it. The page lists the path under what is not exported by default and under the opt-in. Re-verified with a real default session: every stored column in the project has 0 hits for '/Users/', '/var/folders', '/tmp/' and the OS user name. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`pi.cwd` holds an absolute local path, which usually contains the OS user name. It is now exported only with FI_CAPTURE_CONTENT=true, like other content. The contract tests assert that no path reaches the default export and that the opt-in still carries it. The page lists the path under what is not exported by default and under the opt-in. Re-verified with a real default session: every stored column in the project has 0 hits for '/Users/', '/var/folders', '/tmp/' and the OS user name. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…(TH-8290) `omp.cwd` holds an absolute local path, which usually contains the OS user name. It is now exported only with FI_CAPTURE_CONTENT=true, like other content. The contract tests assert that no path reaches the default export and that the opt-in still carries it. The page lists the path under what is not exported by default and under the opt-in. Re-verified with a real default session: every stored column in the project has 0 hits for '/Users/', '/var/folders', '/tmp/' and the OS user name. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…227) The collector now renames claude_code.tool spans to the value of tool_name (Bash, Read, mcp__server__tool), so the trace tree shows which tool ran. tool_name is exported with every OTEL_LOG_* content flag off (checked against the default 2.1.300 fixture). The rename is the last span statement because the earlier ones match on the original name; span.type keeps "tool". Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… identities (TH-8227) - MCP server and tool names: Claude Code 2.1.300 exports them on claude_code.tool (tool_name, gen_ai.tool.name, anthropic.mcp.server.name) with every content flag off. The collector now replaces them with tool_name_safe (mcp_other) unless FI_KEEP_MCP_NAMES=1; new fixture from a real default-flags run against a stub MCP server, plus tests for both modes. - Cost: also write gen_ai.usage.cache_read_tokens / cache_write_tokens; the page now warns that Future AGI's estimate prices cached tokens at the full input rate and points to claude_code.cost.usage for spend. - Input tokens fall back to input_tokens when a cache count is missing. - Delete user.groups (gateway sign-in); page lists the account identifiers that arrive and what the collector does with each. - Self-hosted port guidance (listener + OTEL_EXPORTER_OTLP_ENDPOINT), address-in-use troubleshooting, run_e2e.sh FI_COLLECTOR_PORT note. - Page states there is no fallback without beta traces (N-01 deviation). - Fixtures: local username and user.id hash replaced. - REQUIRE_OTELCOL=1 makes the tests fail instead of skip without the binary. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…renting, root start (TH-8229)
Hooks never block the host:
- Cursor hooks.json / hooks.user.json use the canonical guarded command
`python3 -B <emit.py> cursor <Event> 2>/dev/null || true`; new
test_cursor_hooks_config runs every shipped command through sh/zsh/bash
with the emitter missing, python3 missing, and installed (exit 0, empty
or expected stdout, no __pycache__).
- A misregistered permission hook now prints nothing (Cursor fails open on
no output) instead of answering {"permission": "allow"}.
Privacy and doc truth:
- Opted-in content keeps paths but replaces the user name in home paths with
<user>; pages now say paths inside prompts/tool input/output are exported
with content on, never as metadata. Test: error text with a path and user
name is absent by default and scrubbed with content on.
- State file keeps content only for open spans and turns closed as
incomplete; a duplicate end of a completed turn is ignored.
- Windows: inline send uses a 3 s default timeout; pages say background send
and state locking are macOS/Linux only.
- Cursor page: cache tokens are captured (removed from "not available").
Tree:
- Tool/subagent/llm/event spans without a turn id go under the open turn and
get coding_agent.turn.inferred=true (pages corrected; PRD R-02 deviation).
- Children never start before the root once its start is fixed.
- Subagent span key includes the turn and duration (no cross-turn overwrite).
- Root model comes from session/turn/llm events only.
- transcript_path test uses an unreadable canary and spies on os.open too.
E2E tooling: replay --replace rewrites substrings (suffixed generation ids),
run_e2e.sh gives any fixture fresh UUID ids.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…by default (TH-8227) - Claude Code 2.1.300 copies the error message of a failed model request (provider, gateway or proxy text) into the `error` attribute and the span status message with every content flag off. The collector now sets the status message to Claude Code's `error_class` (or "error") and deletes `error`, unless FI_KEEP_ERROR_TEXT=1. - New fixture from a real default-flags run whose Bash and Read calls failed: Claude Code reports failed tools by class only (ShellError, TelemetrySafeError). - Tests: error text with a local path and user name is absent by default, fixed message without a class, real tool errors keep ERROR + class, and an opt-in class covering FI_KEEP_ERROR_TEXT=1 and FI_KEEP_MCP_NAMES=1 (replaces the subclass that re-ran every default test). - Page: privacy row for error text, FI_KEEP_ERROR_TEXT, real error class examples, identity.source and OTEL_METRICS_INCLUDE_ACCOUNT_UUID, Contrib 0.144.0 minimum for the otlp_http exporter name, where to get spend. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…cument retries and error text (TH-8283) Review fixes for docs#923: - futureagi.patch.json5 and the page no longer set plugins.allow/entries. plugins.allow is an exclusive allowlist and disabled every bundled plugin; `openclaw plugins install` already enables diagnostics-otel. - The patch sets OTEL_NODE_RESOURCE_DETECTORS=env through OpenClaw's config env block (verified: only service.name and project_name reach the resource). The page says what it hides and that a process/.env value wins. - Page: one trace per agent run; model retries are separate traces and roots can stay outcome=completed, find failures by ERROR on model.call / tool.execution. Privacy notes openclaw.error on failed runs and paths in opted-in tool input. model.usage is double counted by Future AGI. harness.run kind cell blank. Tested scope (CLI --local and Gateway; cron root). - run_e2e.sh: EXIT trap stops the Gateway and removes the temp dir holding keys (KEEP_WORK=1 keeps it), umask in a subshell, guarded kill, fails when the Gateway never becomes ready, no detectors line in .env. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ilures (TH-8229)
Cursor's bundled headless agent host passes the hook payload as a trailing
heredoc (`<command> <<'CURSOR_HOOK_EOF'`, commandHookPayloadTransport
argv_heredoc). With `python3 ... 2>/dev/null || true` the heredoc binds to
`true` and the emitter reads nothing. Every shipped command, the shared page,
the Cursor page and the README now use the subshell form
`(python3 -B <path>/emit.py cursor <Event> 2>/dev/null || true)`, which works
for stdin and heredoc and still exits 0 with no output when the emitter or
python3 is missing.
- tests/test_cursor_hooks_config.py runs every shipped command (project and
user hooks) through sh, zsh, bash and dash with the payload on stdin and as
a trailing heredoc: emitter missing, python3 missing, and installed (exit 0,
expected stdout, payload received verbatim, no __pycache__). A regression
test shows the unparenthesized form loses a heredoc payload.
- emit.py prints nothing (instead of "{}") when the adapter cannot be loaded,
respond() raises, or run() fails, so a misregistered permission hook never
gets schema-invalid "{}".
- Home-path scrub handles user names with spaces (when a separator follows),
~name paths, and this machine's own home directory anywhere in the text.
- replay.py --hooks-json/--cwd/--transport and run_e2e.sh HOOK_TRANSPORT run
the installed hooks.json commands the way Cursor does.
- cursor.mdx: CLI section no longer claims a CLI session was traced (it was
not run); troubleshooting covers tools directly under the session (no
marker) separately from coding_agent.turn.inferred.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… error (TH-8289) Review fixes for the Pi extension: - Exports: at most one request in flight; flushes requested meanwhile coalesce into one follow-up. session_shutdown (quit, /new, /resume, /fork, /reload all await it) stops after min(FI_OTLP_TIMEOUT_MS, 3000) ms, aborts the request in flight and never retries after a timeout. - Drop the turn_end handler (an actionable boundary that makes Pi rebuild the session projection every turn). The turn outcome comes from the assistant message's stop reason, as Pi computes it; the turn closes on the next turn_start, agent_end or session_shutdown. - Cancelling a run (Escape, /new, /resume, /fork) is OK with pi.prompt.aborted=true; the cut-off model call and turn are pi.incomplete. Only provider errors fail the prompt and session. - Each export carries provisional copies of the open pi.prompt and pi.turn (same ids, pi.incomplete=true), so finished turns never show at the top level while a run is still going. - Provider error messages are content: opt-in only, like a failed tool's error text. Content is cut at 4096 characters on a code-point boundary; truncated tool arguments are labeled text/plain. - run_e2e.sh accepts a relative workdir; docs wording for /reload, queued follow-ups and the shutdown bound. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…e-only, error privacy (TH-8288) Review of docs#913 (head c92e438): - Queued follow-up message no longer force-closes the running model step: a turn closes only when none of its model steps is still open, so the step keeps its tokens, finish reason and end time (verified on a real `opencode serve` session). - The session span is re-sent, with its deterministic id and creation start, in every idle flush (cumulative counts, latest wins), and with a subagent's batch for the parent root too; dispose is a final refresh. Killed or long-running sessions keep their root. - session.error is not sticky: aborts (MessageAbortedError) and recovered ContextOverflowError are attributes, not ERROR; the session span takes the latest turn's status plus opencode.turn.error_count. - Observe-only: no chat.message or tool.execute.before hook. Prompt text comes from the user message's text parts, tool arguments from the tool part. dispose is cited from the @opencode-ai/plugin 1.19.0 Hooks type. - Error text is content: by default only a sanitised error type is exported (status message = type); text only with FI_CAPTURE_CONTENT. - Subtask commands: one task span (hooks only enrich part-backed tools), no 0-token model span for the synthetic assistant message. - Redaction matches JSON keys and names ending in api_key, secret_key, private_key, access_key, token, secret, password. - Content capped at 4096 UTF-8 bytes on a character boundary; state is pruned after each idle; the test seam is a Symbol key that opencode.json cannot set; default export timeout 3 s, dispose waits at most 2x. - Page: attribute table complete, status rules, resume and kill behaviour as measured, built-in OTel row corrected. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…8289)
A real Pi 1.2.0 run cancelled while bash was running showed that Pi's
next model request fails at setup with stopReason "error" ("This
operation was aborted"), which turned the prompt and session red. Classify
a stop reason "error" as aborted when the run's abort signal (ctx.signal)
is set, the same rule Pi's providers apply. Contract test reproduces the
recorded RPC sequence, with a control run that stays a provider error.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…essions, branch (TH-8290) - Page: new "omp's built-in OpenTelemetry export" section (evaluated against local Future AGI with omp 18.11.0: one invoke_agent trace per prompt, no session.id, intent and raw error text exported by default), a double-export warning in setup and troubleshooting, and how to turn it off (OTEL_TRACES_EXPORTER=none, /settings OTLP Telemetry Export). - Fork guard: handle events only when ctx.agent (omp's agent identity) is present, so the file is inert under upstream pi (PRD V-NEG / R-05). - A session with no prompt, turn, model or tool call exports nothing. - Register session_branch like session_switch; ignore events after session_shutdown until the next session_start. - Content opt-in: JSON-shaped api keys are redacted; cut-off tool JSON is labelled text/plain; cap documented as 4096 characters. - Troubleshooting: drop the non-existent `omp --log-level`, cite omp's "Failed to load extension" message and /extensions; correct the duplicate-copy cause (same-named post/ copies load once). - Doc details: json mode, tested-version wording, subagents are separate traces, FI_OTLP_DEBUG with -p, TUI /branch and /rewind keep the session id, fixture provider renamed to my-gateway. - Tests: 19 contract tests (branch, empty session, late events, fork guard, error text with a path and user name absent by default, JSON key redaction). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…recheck nits (TH-8283)
- The patch's env.vars OTEL_NODE_RESOURCE_DETECTORS=env is set in OpenClaw's
process environment, so commands the agent runs with exec inherit it.
OpenClaw 2026.10.3 has no narrower mechanism: plugins.entries.<id>.env is
rejected by the config schema ("Unrecognized key: env"), and the Gateway's
own environment reaches exec children the same way. The page, patch comment
and env.example now say what it affects (Node programs using
@opentelemetry/sdk-node default detectors lose host/process attributes) and
how to opt out (set the variable where that program is launched, or
`openclaw config unset env.vars.OTEL_NODE_RESOURCE_DETECTORS`), plus a
troubleshooting entry.
- Privacy: exec commands also inherit FI_API_KEY, FI_SECRET_KEY and
OTEL_RESOURCE_ATTRIBUTES from ~/.openclaw/.env (verified with a probe).
- webhook.error / message.processed error text: stated precisely (no 4 KB cap;
message.processed has only the status message).
- diagnostic.phase: count varies per start (0 or about 20), not every start.
- openclaw.exec row: fields verified with a local OTLP capture.
- Patch comment and page no longer point to env.example, which step 2 does
not download.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…uned tool parts With FI_CAPTURE_CONTENT=true, bounded() ran the key-value redaction regex on the whole value before the 4096-byte cap, and the key pattern started with `[A-Za-z0-9_.-]*`, which backtracks over a whole alphanumeric run at every position. A 50K-character run in a tool result took 1.7 s inside the awaited tool.execute.after hook; 1 MiB did not finish in 90 s. - The key pattern drops the prefix. It is unanchored, so OPENAI_API_KEY, x-api-key and client_secret still match at the keyword, with identical output; every pattern is now linear. - Redaction runs on an 8,192-code-unit prefix, doubled only while the redacted prefix is shorter than the cap plus 64 code units, so a secret cut by the prefix's end is never kept and a secret cut by the cap is still redacted before the cut. - Tool parts re-published by compaction pruning (state.time.compacted) are ignored, so a continued session in a new process does not get a phantom tool span. Tests: 1 MiB run (tool args and output), a long secret that makes the prefix grow to 1 MiB with a 512 KiB run inside it, secrets cut by the cap and by the prefix end, surrogate pair at the prefix end, and a pruned tool part in a continued session. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The previous round let a home-path user name run across spaces up to the
next slash, so opted-in Shell command text such as
"cd /Users/alice && ls src/foo" lost "&& ls src". The generic scrub now
replaces only the one directory component after /Users/, /home/ or
C:\Users\: it stops at whitespace, quotes and punctuation no account name
uses, and does not end with a dot.
Names with spaces are handled through this machine's own home directory:
the basename of os.path.expanduser("~") is matched literally after any
home prefix (raw, JSON-escaped or forward-slash Windows form), with a
boundary check so /Users/al does not match inside /Users/alice. This also
fixes "/Users/Jo Ann." and "/Users/Jo Ann," leaking "Ann". Another
account's spaced name keeps the part after the first space; README and
the shared page document that.
Docs: the shared page no longer says every app page uses the guarded
command (hook-based apps do; native-OTel apps have no hook command), and
both pages note that fish reads the subshell parentheses as command
substitution.
Tests: command-text regression cases, own spaced home (POSIX and
Windows forms, end of string, punctuation, al/alice), and the documented
limitation for other accounts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Keep unsupported status notices as reference links, not working-integration cards. Preserve current catalog and public-wording checks. Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Code <noreply@anthropic.com>
Keep unsupported status notices as reference links, not working-integration cards. Preserve current catalog and public-wording checks. Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Code <noreply@anthropic.com>
Keep unsupported status notices as reference links, not working-integration cards. Preserve current catalog and public-wording checks. Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Code <noreply@anthropic.com>
Keep unsupported status notices as reference links, not working-integration cards. Preserve current catalog and public-wording checks. Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Code <noreply@anthropic.com>
Resolve only adjacent catalog entries, keeping every route. Co-Authored-By: Claude Code <noreply@anthropic.com>
Keep every shared catalog entry and preserve all guide content. Co-Authored-By: Claude Code <noreply@anthropic.com>
Keep every shared catalog entry and preserve all guide content. Co-Authored-By: Claude Code <noreply@anthropic.com>
Keep every shared catalog entry and preserve all guide content. Co-Authored-By: Claude Code <noreply@anthropic.com>
Keep all existing module exports and every earlier/new guide entry. No catalog gate is weakened. Co-Authored-By: Claude Code <noreply@anthropic.com>
Keep all existing module exports and every earlier/new guide entry. No catalog gate is weakened. Co-Authored-By: Claude Code <noreply@anthropic.com>
Keep all existing module exports and every earlier/new guide entry. No catalog gate is weakened. Co-Authored-By: Claude Code <noreply@anthropic.com>
Keep all existing module exports and every earlier/new guide entry. No catalog gate is weakened. Co-Authored-By: Claude Code <noreply@anthropic.com>
Resolve only the old literal navigation versus shared catalog; retain every earlier entry. Keep recipe scope versioned and preserve privacy limitations. Co-Authored-By: Claude Code <noreply@anthropic.com>
Resolve only the old literal navigation versus shared catalog; retain every earlier entry. Keep recipe scope versioned and preserve privacy limitations. Co-Authored-By: Claude Code <noreply@anthropic.com>
Resolve only the old literal navigation versus shared catalog; retain every earlier entry. Keep recipe scope versioned and preserve privacy limitations. Co-Authored-By: Claude Code <noreply@anthropic.com>
Resolve only the old literal navigation versus shared catalog; retain every earlier entry. Keep recipe scope versioned and preserve privacy limitations. Co-Authored-By: Claude Code <noreply@anthropic.com>
Resolve only the old literal navigation versus shared catalog; retain every earlier entry. Keep recipe scope versioned and preserve privacy limitations. Co-Authored-By: Claude Code <noreply@anthropic.com>
The npm entrypoint requires Bun; the official --binary --ref path installs the checksummed tested binary without it. Verified omp/18.11.0 in an isolated install directory. Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Code <noreply@anthropic.com>
Exclude unsupported modes explicitly without implying new verification. Keep all privacy, exception-content and exporter limitations; register the guide in the existing shared catalog. Co-Authored-By: Claude Code <noreply@anthropic.com>
Co-Authored-By: Claude Code <noreply@anthropic.com>
Exclude unsupported modes explicitly without implying new verification. Keep all privacy, exception-content and exporter limitations; register the guide in the existing shared catalog. Co-Authored-By: Claude Code <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
New page
src/pages/docs/integrations/traceai/cognee.mdx: how to send Cognee's own OpenTelemetry spans (add, cognify, search, LLM and embedding calls, graph and vector queries) to Future AGI. There is notraceAI-cogneepackage; the page is the docs side of the recipe in future-agi/traceAI#226 (python/examples/cognee):register()before the first Cognee call, the re-add line, and the export filter that keeps prompts, documents, answers and questions out by default.Sections: install, environment variables, setup code, what is captured, privacy (default off, opt-in, stack traces never exported, what was checked), Cognee's own exporter as an alternative (content on;
:443rule; self-hosted gRPC needs lower-case header names), troubleshooting, limitations, and how it differs from traceAI instrumentors, EverOS and the Langfuse cookbook.Nav suggestion (not edited here; the TH-8103 nav PR adds it): group traceAI / Integrations, next to EverOS, title Cognee.
Versions
Evidence (2026-10-16, local Future AGI, real OpenAI
gpt-4o-mini+text-embedding-3-small)7befb18a-2497-4590-b5c8-f9a2c5c5bb19openai/gpt-4o-mini; 0 errorsce1159ea-c50f-4b15-8399-6aa7bbcc1e81memory.process, root); statuslitellm.exceptions.NotFoundError (detail removed: content capture off)17fb9d86-8dc8-4c93-9185-3b99962ced91c23c4360-9d41-4f8e-b781-41a8153ff34b, failure30da4db6-39fc-4922-bc56-7650b3ab6f17https://api.futureagi.com:443/tracer/v1/traces(Future AGI cloud)fi_verify (content off project):
traces=4 spans=262 ... PASSwith--expect-type llm --expect-type retriever --expect-type embedding.Privacy: with the defaults, a unique marker in the document and the question, the document text, the answer, local paths, the user name, the host name and all keys: 0 hits across every stored column (ClickHouse spans/traces/sessions, Postgres traces/spans/project) of 262 spans. With capture on, the marker is in the LLM and search spans' Input/Output; a capture-mode failing run sent 0
exception.stacktracekeys and 0 local paths on the wire (the previous filter sent 10 stack traces with 422 local-path occurrences).Build:
docs_build.shBUILD_EXIT=0, audit-links: 0 broken nav links, 0 broken content links (the new page is one of the expected orphans until the nav PR).Limitations
Alpha recipe. No tokens or cost (Cognee records none). Only search and LLM spans get a type.
cognee.recalland agent-memory entry points are not mapped. Cognee Cloud and non-default backends not tested. Cognee's own exporter exports content (documented as such).Linear: TH-8328 (parent TH-8103).
🤖 Generated with Claude Code