This guide takes a new user from installation to a verified live SkillRun. Skill Runtime observes supported Agents beside their normal workflow; it does not proxy model requests, replace the Agent UI, or require a cloud account.
- macOS or Linux;
- Python 3.9 or newer;
- at least one supported Agent: Codex, Claude Code, Qoder, or OpenCode;
- a browser for the local UI.
The standalone installer does not require sudo, a GitHub account, GitHub CLI,
or a source checkout.
curl -LsSf https://raw.githubusercontent.com/hellogxp/skill-runtime-intelligence/main/scripts/install.sh | sh -s -- --startThe installer:
- downloads the release payload for the current platform;
- verifies the published SHA-256 checksum;
- detects supported Agents and scans standard Skill locations;
- shows every path it intends to read or manage;
- asks once before enabling observation-only, fail-open integrations;
- creates
~/.skill-runtime/config.jsonand~/.skill-runtime/data/panorama.db; - starts the local Collector and UI.
All stored runtime evidence stays under ~/.skill-runtime by default.
Accepting Hook setup during installation enables every detected integration. To inspect the plan without changing any Agent configuration:
skill-runtime setupTo enable only one integration later:
skill-runtime setup --enable-codex-hooks
skill-runtime setup --enable-claude-hooks
skill-runtime setup --enable-qoder-hooks
skill-runtime setup --enable-opencode-pluginEvery managed configuration entry has a Skill Runtime marker. Existing entries are preserved and a backup is created before a supported configuration file is changed.
- Run
skill-runtime setup --enable-codex-hooks. - Open
/hooksin Codex. - Review and trust the commands managed by Skill Runtime.
- Start a new Codex task; already-open tasks do not hot-load new Hooks.
Enable the managed Hooks, restart Claude Code if it was already running, and begin a new session.
Enable the managed Hooks and restart Qoder. Qoder loads its Hook configuration at process startup.
Enable the managed observation-only plugin and restart OpenCode if its current process predates installation.
Use a Skill normally in a new Agent task. Then run:
skill-runtime doctorThe integration states have strict meanings:
| State | Meaning |
|---|---|
Live / Verified |
The local database received a real runtime event from this integration |
Pending |
Configuration exists, but no real event has arrived yet |
Fallback |
Evidence is being reconstructed from compatible local records |
Unsupported |
This Agent/version does not expose the relevant signal |
Disabled |
The integration was not enabled |
Writing a Hook configuration is never enough to claim live collection.
Open http://127.0.0.1:4317. The Runtime Overview identifies runs worth investigating; selecting a run opens its ordered Panorama, first observable boundary, evidence timeline, and Inspector.
Every node and relationship carries an evidence grade:
- Observed — directly present in a source event or scanned file;
- Derived — deterministically connected from observed evidence;
- Inferred — an uncertain explanation that cannot rewrite facts;
- Experimental — an effect measured through controlled evaluation.
Not observed is not a synonym for failed. The adapter capability panel
shows whether a source could have emitted the missing signal.
skill-runtime status
skill-runtime doctor
skill-runtime restart
skill-runtime stop
skill-runtime uninstall --keep-datauninstall removes only entries and files owned by Skill Runtime. Without
--keep-data, deletion of ~/.skill-runtime requires explicit confirmation.
Agent sessions, projects, and Skill sources are never removed.
- The server binds to
127.0.0.1by default. - Prompts, raw tool payloads, patch bodies, credentials, and Skill resource contents are not copied into the normalized index.
- Common secret patterns are redacted before persistence.
- Network export is off until an endpoint is explicitly configured.
- Hook delivery is bounded and fail-open: a collection failure does not deny, modify, or delay an Agent decision beyond the small configured timeout.
Review the exact data boundaries in Architecture and the signal limitations in the adapter capability matrix.
Skill Runtime exports normalized Skill-specific evidence over OTLP/HTTP:
skill-runtime start \
--otlp-endpoint https://collector.example/v1/tracesFor authenticated background export, use the standard
OTEL_EXPORTER_OTLP_HEADERS environment variable before starting the runtime.
Secrets are not written into Skill Runtime configuration or command history by
the product.
See Observability platform setup for Grafana/OpenTelemetry Collector, Datadog, and other OTLP-compatible backends, plus supported historical import profiles.
- confirm the Agent was restarted after setup;
- create a new task/session rather than reusing an existing one;
- for Codex, confirm the managed commands are trusted in
/hooks; - use a Skill so the Agent actually emits a Skill-relevant event;
- run
skill-runtime status, thenskill-runtime doctor.
No verified primary event has arrived for that run. The transcript adapter is
serving compatibility evidence and the UI labels it accordingly. Check
skill-runtime doctor and the
adapter capability matrix.
skill-runtime stop
skill-runtime start --port 4318Then open http://127.0.0.1:4318.
skill-runtime install --no-hooksThe product remains usable for Skill inventory, supported imports, and labeled fallback reconstruction, but it will not present those records as primary live Hook evidence.