Repository navigation
0.9.0
A skill's cost is now the harness's own requests named for it, and stdtel's tail-rule estimate is
retired (ADR-014 decision 10). Breaking on the wire and in the warehouse, as listed below; a minor
bump, as with 0.4.0 and 0.7.0, because this is pre-1.0 and the changelog says plainly what breaks.
Upgrade both halves: the package and the plugin (stdtel-query 1.2.0 and the analyst 1.1.0 ship in
it).
Changed — breaking, on the wire and in the warehouse
- A skill's cost is the harness's own requests, and stdtel's estimate is retired (ADR-014
decision 10, #117). Native records were verified live in 0.8.0, which started the clock the ADR
set. A user will see three things:- Skill spans carry no tokens and no model.
std.skill.load_tokens,std.skill.tail_tokens,
std.skill.tail_tokens_first_only,std.skill.llm_requestsand thegen_ai.usage.*and
gen_ai.request.modelattributes are gone fromkind=skillactivations, and the allowlist
refuses them. Turn, sub-agent, compaction and session spans keep their counts: those are what the
transcript observed, not an attribution. - Nine
skill_invocationcolumns are dropped: the three estimates,llm_requests, the four
token columns andmodel. They defaulted to 0, which for every new row would have read as
"measured, and free". As at the ADR-013 cutover, the old values (estimates) are not migrated. - Per-skill cost is NULL for any session without Claude Code's own telemetry on: every
machine that has not runstdtel-install settings(or--native-onlywith the plugin) since
0.8.0. Such runs are counted asunmeasured, never as free.
- Skill spans carry no tokens and no model.
- New views:
skill_request_cost, one row per (session, prompt, skill) overllm_request_attributed;skill_activation_cost, the same with stdtel's version, plusabandoned, which is NULL when
the prompt has no native record.
- Rebuilt on the new views:
- the scorecard:
total_cost_usd,total_tokens,p50_tokens,requests_derived,
abandonment_rateandunmeasured_activationsreplacep50_load_tokens,p50_tail_tokens
and the oldtotal_tokens; - Q5 (
skill_tokens,n_prompts,n_unmeasured); - Q6, which counts each request in a scoped prompt once, plus external spend, and reports
prompts_without_usage; - the loop and token-by-kind dashboard panels;
- Q1, whose skill row now reports the harness's requests named for a skill in the session, and
counts unmeasured runs, rather than showing every skill as usage-free; - the demo fleet, which now seeds
llm_request. Every value it seeded before is unchanged: a full
dump ofgenerate(42)before and after, minus the retired fields, is byte-identical.
- the scorecard:
- The scorecard panel is now held verbatim to
scorecard.sqlby a test, and every Warehouse panel
is executed against the schema in tests. Before, a panel selecting a dropped column passed all of
them. stdtel-query1.2.0 and the scorecard analyst 1.1.0 read cost from the views.telemetry.emit: falsemeans less than it did. It still suppresses stdtel's named span. It
cannot stop Claude Code naming the skill on its own requests, which are now where a skill's cost is
read from.docs/for-developers.mdsays so.- Copilot loses nothing it had. Copilot's own hook dialect sends
transcriptPath, which the Stop
hook never read, so the tail rule never produced Copilot data. Copilot per-skill tokens arrive with
#115.
Privacy
- The transcript reader no longer reads a Skill tool result's content. chars/4 was the only reason
it ever did.
Fixed
stdtel-install settingsno longer redirects Claude Code's telemetry without saying so (#133).
It used to overwrite anOTEL_EXPORTER_OTLP_ENDPOINTalready in the file. Run from a shell without
STDTEL_OTLP_ENDPOINT, that pointed Claude Code at localhost. A different endpoint is now kept,
with its protocol (the two only work as a pair), and reported.--replace-endpointoverwrites both.- The content check and the detailed-view gate judge the collector Claude Code's events actually
go to (#133). AnOTEL_EXPORTER_OTLP_LOGS_ENDPOINTtakes precedence, as the harness resolves it:
each variable comes from the highest-precedence settings file that sets it, then from the
environment. - A session open across an upgrade no longer loses its skill windows. Saved state is loaded with
unknown keys skipped. Without that, the removedload_tokensfield would have made every 0.8
state file raise inside a hook that exits 0, so the loss would have been silent.