Repository navigation
Fixes found watching a live session: usage on the dashboard, MCP and skill names through the collector, and
skills started by slash command. Restart the collector and Grafana after upgrading (docs/local-stack.md).
Fixed
-
A skill started by slash command is recorded, and a loop skill started that way opens its scope
(#144). stdtel captured skills only through the Skill tool. A skill the user types as/name
never goes through that tool, so it was invisible. Verified:/probe-echoproduced a turn and no
skill activation. A loop skill declaringtelemetry.scopetherefore never opened its container
when started the usual way. The Stop hook now reads typed commands from the transcript:- A command is a skill when the harness loaded one for it (its "Base directory for this skill:"
entry), or when the catalogue knows the name. Built-ins such as/modeland/compactare
ignored. A typed skill is recorded with triggeruser-slash. /loop /epic-loop …: the harness activates onlyloop, andepic-loopis its argument.
By decision, a catalogued skill named first in/loop's argument is recorded as invoked with
triggerloop, so the inference is labelled and can be filtered out.- Both are
std.artefact.source=transcript. Only the command name and that first argument word
are read. Replayed against the real sightline transcript, it finds the session's five
/loop /epic-loopstarts, the first on 31 August, and ignores/model,/compactand
/advisor.
- A command is a skill when the harness loaded one for it (its "Base directory for this skill:"
-
MCP server, MCP tool and skill names survive the collector (#142). With the detailed view on,
Claude Code puts them only insidetool_parameters, beside shell commands. Every user-configured
MCP call'stool_nameis the literal"mcp_tool". The collector deletedtool_parameterswhole,
so the names a user opted in to never reached storage.- A new
transform/tool_nameslifts exactly four keys out first:mcp_server_name,
mcp_tool_name,skill_nameandsubagent_type. Only values shaped like a name are lifted
(short, no spaces); the content is then dropped as before. It runs on logs and traces, under both
overlays, and is tested on the real image with a content marker beside the names. - Verified live: the first MCP call after the collector restart arrived as
llm-council, and no
content key appeared in any of 66 real events. - The operational dashboard gains tables for MCP calls by server and tool, skills invoked, and
tool calls by tool and success. A hidden name reads(hidden: detailed view off). - Tables now show their value instead of a timestamp, and cache-read tokens sit on their own axis.
- A new
-
The operational dashboard shows usage (#140). During a live session that spent $15.43 over 167
model requests, it showed none of it. Three causes:- Nothing read Loki, where Claude Code's own per-request records land. A new top row shows
spend and requests in the range, spend per hour, tokens per hour by type, requests per hour by
model, and spend by skill, agent and repo. All of it is live from Loki, with no loader in between. - "Claude Code tokens by type" read a native metric the collector drops by design since #110,
so it could never show anything. It is removed. A test now fails any panel that queries a metric
prefix in the collector'sfilter/native_metrics. - Seven "/ h" panels plotted per-second rates. "0.003 turns/h" was about 11. They are scaled to
per hour, and a test holds every "/ h" rate to it.
Skill panels now say "No skill ran in this range" instead of "No data". Every Loki panel's query is
run against Loki in tests. - Nothing read Loki, where Claude Code's own per-request records land. A new top row shows
-
Grafana reads datasource provisioning only at startup, so a stack running since before 0.8.0
has no Loki datasource and every usage panel fails with "Data source not found".
docs/local-stack.mdgives the one-line fix: restart Grafana alone.