Skip to content

feat(server-utils): Add instrumentation for pi-durable - #24993

Draft
JPeer264 wants to merge 1 commit into
jp/bun-conversation-id-integrationfrom
jp/pi-durable-instrumentation
Draft

JPeer264 wants to merge 1 commit into
jp/bun-conversation-id-integrationfrom
jp/pi-durable-instrumentation

Conversation

@JPeer264

@JPeer264 JPeer264 commented Oct 2, 2026

Copy link
Copy Markdown
Member

Adds piDurableIntegration for @earendil-works/pi-durable, Earendil's new durable agent harness. It is on by default in Node. Each run becomes its own invoke_agent trace with chat and execute_tool children, and a subagent run nests under the tool call that started it.

pi-durable has no telemetry hooks (pi-telemetry exists, but no pi package emits spans through it yet), so the integration hooks Harness.open() and wraps the models and registry it receives.

Decisions worth a look:

  • Runs are tracked per conversation. pi-durable hands run control to a new pi.generation task after every tool round, so no single task spans a run. The run span ends when the commit that removes pi.live.run goes through.
  • Each run starts a new trace in clean scopes. The scheduler runs task phases in whatever async context last woke it, often an unrelated request. Forking that context attached runs to the wrong request and the wrong conversation.
  • gen_ai.conversation.id is <harness id>:<conversation id>. pi-durable numbers conversations per storage from 1, so the bare id would merge every root conversation into one. The cost is a new id after a restart, until pi-durable has a stable storage id.
  • Tool spans record the result the model receives, read from the committed result entry. That covers output streamed through api.output() (all of bash) and results an afterTool hook redacted.
  • Errors: failures pi-durable only passes to onReport (throwing hooks or sections) and faults of any task are captured. Throws from the built-in coding-tools extension are not, because bash throws for every non-zero exit, a normal result for a coding agent. The span keeps its error status.

Messages go through the pi-ai mapper from the Flue fix, and the system prompt and tools are replayed from pi-durable's positional system messages.

Not covered yet: deferred model responses, spans for tool calls pi-durable answers without running execute(), runs interrupted by harness.close(), spans for custom tasks, agent names, and Cloudflare Durable Objects.

🤖 Generated with Claude Code

@JPeer264
JPeer264 added this pull request to stack #25002 October 2, 2026 13:14
@github-actions

github-actions Bot commented Oct 2, 2026 •

Copy link
Copy Markdown
Contributor

size-limit report 📦

⚠️ Warning: Base artifact is not the latest one, because the latest workflow run is not done yet. This may lead to incorrect results. Try to re-run all tests to get up to date results.

Path Size % Change Change
@sentry/browser 29.52 kB - -
@sentry/browser - with treeshaking flags 27.68 kB - -
@sentry/browser - with treeshaking flags tracing without tracing 27.57 kB - -
@sentry/browser (incl. Tracing) 51.45 kB -0.01% -1 B 🔽
@sentry/browser (incl. Tracing + Span Streaming) 51.46 kB -0.01% -1 B 🔽
@sentry/browser (incl. Tracing, Profiling) 54.46 kB - -
@sentry/browser (incl. Tracing, Replay) 91.04 kB - -
@sentry/browser (incl. Tracing, Replay) - with treeshaking flags 80.03 kB - -
@sentry/browser (incl. Tracing, Replay with Canvas) 95.74 kB - -
@sentry/browser (incl. Tracing, Replay, Feedback) 108.71 kB +0.01% +1 B 🔺
@sentry/browser (incl. Feedback) 47.04 kB - -
@sentry/browser (incl. sendFeedback) 34.57 kB - -
@sentry/browser (incl. FeedbackAsync) 39.68 kB - -
@sentry/browser (incl. Metrics) 30.54 kB - -
@sentry/browser (incl. Logs) 30.82 kB - -
@sentry/browser (incl. Metrics & Logs) 31.48 kB - -
@sentry/react 31.36 kB - -
@sentry/react (incl. Tracing) 53.8 kB - -
@sentry/vue 37.51 kB - -
@sentry/vue (incl. Tracing) 54.33 kB - -
@sentry/svelte 29.55 kB - -
@sentry/remix (Remix 3 client bundle) 55.77 kB -0.02% -9 B 🔽
CDN Bundle 31.22 kB - -
CDN Bundle (incl. Tracing) 51.98 kB - -
CDN Bundle (incl. Logs, Metrics) 33.46 kB - -
CDN Bundle (incl. Tracing, Logs, Metrics) 53.92 kB - -
CDN Bundle (incl. Replay, Logs, Metrics) 74.21 kB - -
CDN Bundle (incl. Tracing, Replay) 89.56 kB - -
CDN Bundle (incl. Tracing, Replay, Logs, Metrics) 91.53 kB - -
CDN Bundle (incl. Tracing, Replay, Feedback) 95.72 kB - -
CDN Bundle (incl. Tracing, Replay, Feedback, Logs, Metrics) 97.71 kB - -
CDN Bundle - uncompressed 92.14 kB - -
CDN Bundle (incl. Tracing) - uncompressed 154.49 kB - -
CDN Bundle (incl. Logs, Metrics) - uncompressed 98.71 kB - -
CDN Bundle (incl. Tracing, Logs, Metrics) - uncompressed 160.45 kB - -
CDN Bundle (incl. Replay, Logs, Metrics) - uncompressed 228.28 kB - -
CDN Bundle (incl. Tracing, Replay) - uncompressed 274.22 kB - -
CDN Bundle (incl. Tracing, Replay, Logs, Metrics) - uncompressed 280.16 kB - -
CDN Bundle (incl. Tracing, Replay, Feedback) - uncompressed 287.93 kB - -
CDN Bundle (incl. Tracing, Replay, Feedback, Logs, Metrics) - uncompressed 293.86 kB - -
@sentry/nextjs (client) 56.32 kB -0.01% -1 B 🔽
@sentry/sveltekit (client) 51.87 kB - -
@sentry/core/server 40.59 kB - -
@sentry/core/browser 13.63 kB -0.01% -1 B 🔽
@sentry/node 148.02 kB +2.26% +3.27 kB 🔺
@sentry/node/import (ESM hook with diagnostics-channel injection) 83.26 kB +0.05% +39 B 🔺
@sentry/node - without tracing 93.39 kB +0.08% +74 B 🔺
@sentry/node - without channel injection 126.25 kB +2.68% +3.29 kB 🔺
@sentry/aws-serverless 101.62 kB +0.04% +33 B 🔺
@sentry/cloudflare (withSentry) - minified 208.61 kB - -
@sentry/cloudflare (withSentry) 517.4 kB - -

View base workflow run

@JPeer264
JPeer264 removed this pull request from stack #25002 October 2, 2026 13:31
@JPeer264
JPeer264 changed the base branch from jp/flue-map-messages-conventions to jp/bun-conversation-id-integration October 2, 2026 13:39
@JPeer264
JPeer264 force-pushed the jp/pi-durable-instrumentation branch from 1dad424 to 91a0b35 Compare October 2, 2026 13:39
@JPeer264
JPeer264 added this pull request to stack #25004 October 2, 2026 13:57
pi-durable has no telemetry hooks, so `piDurableIntegration` wraps the
`models` and `registry` passed to `Harness.open()`. Each run becomes its
own `invoke_agent` trace with `chat` and `execute_tool` children, and a
subagent run nests under the tool call that started it.

Messages are mapped to the gen_ai conventions, and the system prompt and
tools are read from pi-durable's positional system messages. Tool spans
record the result the model receives. Failures pi-durable only reports,
and faults of any task, are captured. Failures of the built-in coding
tools are not, because they report expected results to the model.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@JPeer264
JPeer264 force-pushed the jp/pi-durable-instrumentation branch from 91a0b35 to 26d39c9 Compare October 2, 2026 16:59

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant