type -> format
id -> kvscience/tasks
def -> the KVMACHINES telemetry build and the Claude Science page tests (group claude-science: one task per docs page, one MLflow run per page in MLflow experiment 2, kvmachines/claude-science, written straight to the sqlite store at /Users/zhouk/workspace/claudia/var/mlflow/mlflow.db -- not through a server, which a Claude Science sandbox cannot reach), one record per task, rendered by site/index.html at https://kvscience.com; state is done | running | waiting | building | open | blocked; `needs` names who or what the task waits on; `evidence` is the command, file or page that proves the state; deployed on push to main
keys -> type, id, then alphabetical
line -> key -> value; parse by splitting on the first " -> "
updated -> 2026-10-10T13:22:59Z

type -> task
id -> honeycomb-organization
evidence -> claude.ai admin settings, Claude Code, managed settings v12; data/admin-settings/server-managed-settings.json in the claudia repository
needs -> the operator pastes the Honeycomb ingest key (environment test, name mac-collector) into the managed settings editor
state -> done
title -> Honeycomb for every agent in the organization
what -> every Claude Code session of every member, on any host, exports traces, metrics and logs straight to api.honeycomb.io (team kvmachines.com, environment test, dataset claude-code) through the organization's server-managed settings

type -> task
id -> mlflow-managed
evidence -> enabledPlugins mlflow-tracing@mlflow-plugins and MLFLOW_CLAUDE_TRACING_ENABLED in managed settings v12; the Mac's tracking URI in its user and machine settings
needs -> nothing
state -> done
title -> MLflow tracing in managed settings
what -> the mlflow-tracing plugin is enabled for the whole organization; its Stop hook posts each session as a trace into experiment kvmachines/claude-code; only the tracking URI is per host

type -> task
id -> mlflow-cloud
evidence -> the tunnel is up and the REST API answers through it: GET /api/2.0/mlflow/experiments/get returns kvmachines/claude-science and runs/search returns all 4 runs, while the same call without the service token gets 302 to the Access login. Tunnel kv-mlflow (ca49064a-06c1-4bab-a5db-612c546337a6) -> http://127.0.0.1:5000, proxied CNAME mlflow.kv-keys.com, Access app kv-mlflow with an Operator policy and the non-identity token kv-mlflow agent
gotcha -> MLflow's own middleware rejected the proxied request with "Invalid Host header - possible DNS rebinding attack detected" because mlflow.kv-keys.com is not in the server's --allowed-hosts; fixed in the tunnel rather than the server by setting originRequest.httpHostHeader to 127.0.0.1:5000, the same rewrite kv-mirror uses
needs -> nothing for the Mac; a cloud VM still needs CF-Access-Client-Id and CF-Access-Client-Secret in its environment alongside MLFLOW_TRACKING_URI
state -> done
title -> MLflow tracing in cloud sessions
what -> a cloud VM gets the plugin and flags from the same managed settings; its setup script writes MLFLOW_TRACKING_URI https://mlflow.kv-keys.com into the VM's user settings

type -> task
id -> mac-experiment
evidence -> MLflow experiment 1 holds trace tr-934e895af1037aa7b0d94f19c46c99fa (2026-10-10T11:10:41Z, state OK, claude_code_version 2.1.296), found with POST /api/3.0/mlflow/traces/search
needs -> nothing
state -> done
title -> One traced session on the Mac
what -> one headless prompt on claudia's account ran with the plugin and the local tracking URI; MLflow recorded the trace with its cost

type -> task
id -> linux-headless
evidence -> claudia data/linux/Dockerfile and scripts/claudia linux-test build|run; build log var/log/linux-test-build.log
needs -> a long-lived token the operator creates with claude setup-token and stores in Proton Pass as kvmachines-claude-oauth
state -> building
title -> Linux headless test
what -> an Ubuntu 24.04 image with Claude Code, uv and MLflow builds under apple/container on the Mac; one headless prompt in it is traced the same way

type -> task
id -> honeycomb-research
evidence -> issues 38 to 43 in kvkeys/kvkeys-in-chrome, label telemetry, board github.com/orgs/kvkeys/projects/5; issue 38 runs on the finance seat
needs -> five more cloud sessions on Premium seats (design, legal, engineering)
state -> running
title -> Honeycomb research, six issues
what -> each issue is one research task a cloud session works from its own issue: send-data and limits, the honeycombio GitHub org, the MCP server and agent skill, resources as code, the Claude Code signal mapping, the collector

type -> task
id -> kvscience-site
evidence -> https://kvscience.com returns 200 and serves this file; worker kvscience in account 634ff28065e23b4f4a0b6200d62bc97b, custom domains kvscience.com and www.kvscience.com attached to zone bb7e8bef7641bf01f0e6eb48fd204311. GitHub Actions had never run (total_count 0) because deploy.yml triggers on push to main while the only branch was master; master is now renamed to main and set as the default
how -> the first deploy was made through the Cloudflare connector (an authenticated MCP session on the account), not wrangler, because the sandbox has no Cloudflare API token. It is a bootstrap worker that inlines site/index.html and site/tasks.kv rather than the wrangler [assets] build; CI replaces it with the real assets build on the first successful run
needs -> the repository secret CLOUDFLARE_API_TOKEN, which does not exist yet (actions/secrets total 0); until it is added in Settings -> Secrets and variables -> Actions the deploy job fails at the wrangler step, and every publish has to go through a Claude Science session
state -> done
title -> kvscience.com progress site
what -> this public page, behind Cloudflare, renders tasks.kv so anyone can follow the build without a login

type -> task
id -> claude-science
evidence -> Claude Science 0.1.60 serves on the Mac (port 8000), signed in as KVMACHINES from the scientist seat; MLflow experiment 2 kvmachines/claude-science holds the page tests, one run per page
needs -> the next page test, run by a Claude Science session on the scientist seat, never by claudia (the operator, 2026-10-10: "please use not claudia usage", "use scientist@kvmachines.com")
state -> running
title -> Claude Science experiment
what -> a Claude Science session runs one experiment that tests the build: reads its telemetry settings, sends one request through them, reports whether Honeycomb and MLflow received it

type -> task
id -> motion-videos
evidence -> none yet
needs -> renders in 16:9 and 9:16, named yyyymmdd-<unix ns>-<topic>
state -> open
title -> Motion videos, one per MLflow implementation
what -> three thirty-second Motion videos: managed settings, cloud sessions, the Mac and Linux experiments

type -> task
id -> claude-science-01-overview
evidence -> MLflow experiment 2 run 46bcaac6ea124e27b25a66fea40226ed, attempt 2, pass 1: all five claims of the page hold on this install — version (Settings > General > About reads 0.1.60), kernels (Python 3.11.16, R 4.5.3, bash 5.3.20), sandbox (bind EPERM, no DNS, proxy-only egress, private IPs hard-refused), artifacts (versioned, each with a latest_version_id and lineage), reviewer (host.findings() answers, count 0); attempt 1 run 33be6d00bdcd4f4d81b48f6d391796f0 scored pass 0 only because it read the version from a stale 0.1.55 bundle
group -> claude-science
corrected -> there is only ONE install; the earlier claim that /Applications/Claude Science.app was "a separate stale install at 0.1.55" was wrong. That bundle is the launcher shell and its CFBundleShortVersionString is the launcher's own version; the daemon that serves the app runs from ~/.claude-science/runtime/0.1.60-release and About reports that. The Mac session trashed the bundle on this agent's advice, found the running app (pid 16393) executing from ~/.Trash, and moved it back. Read the version from `ls ~/.claude-science/runtime` or the daemon log, never from the plist, and never delete that bundle
needs -> nothing
page -> https://claude.com/docs/claude-science/overview
run -> 46bcaac6ea124e27b25a66fea40226ed
state -> done
title -> Claude Science · Claude Science
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-02-get-started
evidence -> MLflow experiment 2 (kvmachines/claude-science) run 88a253340f2e4880b5a1a6f0ac52719a, attempt 1: installed (0.1.60) and serving on port 8000, but the app was signed into the retired AKWLABS account, so pass 0; attempt 2 follows the sign-in fix; attempt 2 run d20d535575ec426b9c0a5353dda97a65: signed in as KVMACHINES from the scientist seat, environments ready, pass 1
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/get-started
state -> done
title -> Claude Science · Get started
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-03-run-on-remote-linux-server
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/run-on-remote-linux-server
state -> open
title -> Claude Science · Run on a remote server
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-04-core-concepts
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/core-concepts
state -> open
title -> Claude Science · Core concepts
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-05-safeguards
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/safeguards
state -> open
title -> Claude Science · Safeguards
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-06-multiple-computers
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/multiple-computers
state -> open
title -> Claude Science · Multiple computers
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-07-artifacts
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/artifacts
state -> open
title -> Claude Science · Artifacts
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-08-comments
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/comments
state -> open
title -> Claude Science · Comments
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-09-the-reviewer
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/the-reviewer
state -> open
title -> Claude Science · The reviewer
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-10-tools-and-environments
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/tools-and-environments
state -> open
title -> Claude Science · Tools and environments
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-11-troubleshooting
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/troubleshooting
state -> open
title -> Claude Science · Troubleshooting
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-12-connectors-and-skills
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/connectors-and-skills
state -> open
title -> Claude Science · Connectors and skills
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-13-remote-compute-clusters
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/remote-compute-clusters
state -> open
title -> Claude Science · Remote compute clusters
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-14-compute-providers
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/compute-providers
state -> open
title -> Claude Science · Compute providers
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-15-cloud-storage
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/cloud-storage
state -> open
title -> Claude Science · Cloud storage
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-16-literature-access
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/literature-access
state -> open
title -> Claude Science · Literature access
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-17-custom-connectors
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/custom-connectors
state -> open
title -> Claude Science · Custom connectors
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-18-enable-claude-science
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/enable-claude-science
state -> open
title -> Claude Science · Enable Claude Science
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-19-how-claude-science-works-with-your-data
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/how-claude-science-works-with-your-data
state -> open
title -> Claude Science · How Claude Science works with your data
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-20-admin-controls
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/admin-controls
state -> open
title -> Claude Science · Admin controls
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-21-manage-on-devices
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/manage-on-devices
state -> open
title -> Claude Science · Manage on devices
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-22-corporate-networks
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/corporate-networks
state -> open
title -> Claude Science · Corporate networks
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-23-network-requirements
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/network-requirements
state -> open
title -> Claude Science · Network requirements
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-24-monitor-usage
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/monitor-usage
state -> open
title -> Claude Science · Monitor usage
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-25-glossary
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/glossary
state -> open
title -> Claude Science · Glossary
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-26-command-line-settings
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/command-line-settings
state -> open
title -> Claude Science · Command line settings
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-27-configuration-file-reference
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/configuration-file-reference
state -> open
title -> Claude Science · Configuration reference
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-28-changelog
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/changelog
state -> open
title -> Claude Science · Changelog
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page

type -> task
id -> claude-science-29-legal-and-compliance
evidence -> none yet
group -> claude-science
needs -> one MLflow run in experiment kvmachines/claude-science that records the test of this page (pass 1 or 0, the evidence) after the page was followed on the operator's Mac
page -> https://claude.com/docs/claude-science/legal-and-compliance
state -> open
title -> Claude Science · Legal and compliance
what -> follow the page as written, one page per run; record what worked and what did not as the run's params, metrics and tags; fix before moving to the next page