ODF docs
GitHub ↗
Documentation menu

ODF linear-cycles baseline — 2026-10-01

2 min read

#Purpose

Record the available runtime evidence before changing ODF route execution, as required by issue #95. This is a partial baseline, not an end-to-end workflow latency measurement.

#Snapshot

  • Captured: 2026-10-01 UTC.
  • Window: 14 UTC date buckets, 2026-09-18 through 2026-10-01 inclusive.
  • Source: the installed ODF metrics directory, read in place with ODF_CONFIG_DIR=~/.config/opencode; no raw telemetry was copied into this repository.
  • Command: node scripts/odf-toolkit.js metrics --days 14 --json.
  • Raw records: 475. The dashboard's aggregate delegation cohort contains 142 samples after its lifecycle/span/join filtering and run de-duplication.

#Delegation-duration baseline

Durations are for individual delegated phase calls, not user-visible workflow wall-clock time.

Cohort Samples p50 p95
All eligible phase calls 142 250,784 ms (4m 11s) 900,174 ms (15m 00s)
Phase Samples p50 p95
ASSESS 8 417,083 ms 900,280 ms
DESIGN 24 428,303 ms 900,815 ms
EXPLORE 2 554,083 ms 595,483 ms
FIX 9 402,346 ms 665,154 ms
IMPLEMENT 57 147,325 ms 600,364 ms
PROPOSE 9 14,695 ms 211,933 ms
QA-PLAN 13 269,238 ms 883,214 ms
VERIFY 20 204,738 ms 749,402 ms

Outcome distribution for the 142-sample cohort:

Outcome Count Share
ok 43 30.3%
blocked 60 42.3%
error 15 10.6%
timeout 24 16.9%

#Coverage and limits

  • work_type is present for 77/142 samples (54.2%): 8 small-change, 69 feature; 65 are unknown. This is not enough to compare route-specific outcomes confidently.
  • A usable model identity is unavailable for all 142 samples.
  • Validation-evidence, receipt, and escalation fields are unreported for all 142 samples.
  • entry_to_final_gate has 0 samples and is reported as unavailable. The dataset does not establish request-to-VERIFY wall-clock time, user turns, or orchestration handoffs.
  • The historical workflow review found qualitative friction around repeated status/proof reconstruction during continuation and archive-state recovery after concurrent activity. Those observations are not timed and are not included as numeric samples here.

Therefore these figures establish phase-call latency only. They do not support a claim that a runtime route change improves whole-workflow latency by 50%, nor do they justify enabling the advisory triage binding as auto-routing.

#Measurement gate before runtime routing changes

  1. Capture a cohort linking entry, route/work type, phase calls, and verified final gate to the same change/run identity; include outcomes and handoff/turn counts without recording prompt or customer content.
  2. Verify that the cohort has usable entry_to_final_gate samples and adequate work_type coverage before comparing routes.
  3. Keep the existing advisory binding advisory until its execution consumer and persisted authorization contract have dedicated tests. Preserve the mandatory BUILD/VERIFY evidence, security, OCA, TDD, receipt, and storage compatibility gates.

Re-run the same command after a low-risk rollout and compare equivalent cohorts; do not interpret missing data as zero latency or a successful workflow.