-
Notifications
You must be signed in to change notification settings - Fork 2
plat 019
github-actions[bot] edited this page Sep 20, 2026
·
1 revision
| Coordination | Value |
|---|---|
| Assigned agent | Codex |
| Ticket state | runtime_reverify |
| Last synchronized | 2026-08-04 |
Claim this ticket in this file before implementation. During active work, update this fragment rather than the shared index; synchronize the index once at handoff, review, or completion.
- Priority: P1
- Owner: Pulse agent-metrics accounting and pricing projection
-
Source run: RTS Latency
schedule-cron--42eca39a_1785810615371091000 -
Problem:
pulse_agent_metricscontains complete usage for Strategy Auditor, Workflow Review, and the consolidated Fixer, including roles, durations, call counts, model IDs, prompt/completion/cache tokens, andusage_status='captured'. All three rows nevertheless storetotal_cost_usd=0, and each model payload marks every call unpriced. The phase and authoritative execution ledgers price the sameclaude-opus-5andclaude-sonnet-5identities with versioned non-zero rate cards. - Impact: the Pulse UI can show time and tokens but reports zero dollars for the most expensive agents, while another ledger reports a real cost. Review economics cannot be compared reliably and cost-reduction decisions are made from contradictory projections.
-
Distinction from PLAT-008: PLAT-008 fixed phase-ledger arithmetic and model
rate-card selection; that arithmetic is working here. PLAT-019 is the failure
to apply the same canonical pricing identity when persisting
pulse_agent_metrics. -
Implementation (2026-08-04): cost aggregates retain the provider and the
token slice belonging specifically to unpriced calls. Pulse metrics apply
the canonical versioned model rate card only to that slice, preserve existing
priced cost, and expose component costs/pricing identity. Unknown models are
recorded as
captured_unpricedwith a non-empty explanation. -
Verification: a focused SQLite-ledger test captures an unpriced
claude-sonnet-5Pulse call and proves a non-zero versioned metric result. The next real Pulse pass must reconcile the displayed dollars to the phase ledger. -
Regression test:
TestPulseAgentMetricsPricesCapturedClaudeUsageWithCanonicalRateCard; the full costledger and step-based workflow packages pass on 2026-08-04. -
Acceptance: for a priced model, each Pulse metric row reconciles to the
same versioned component costs as the phase/execution ledger. Truly unpriced
models remain explicit with a non-empty reason;
usage_status='captured'must not silently imply$0.
Auto-synced from docs/ on main. Edit there, not here.