Summary

gstack-analytics (the local-usage dashboard) silently displayed (avg 0s) next to any /skill whose completion events emit "duration_s" as a quoted JSON string. The completion-status preamble that every skill runs (scripts/resolvers/preamble/generate-completion-status.ts:70) and bin/gstack-codex-probe:84 both emit "duration_s":"<N>", so the dominant in-the-wild case rendered as 0s and most per-skill averages in the dashboard have been wrong since this code path was added. bin/gstack-telemetry-log emits the unquoted numeric form, which is why the bug wasn't visible for every skill.

The companion Total time: line at the bottom of the dashboard was always correct — it uses a different parser, so the bug was confined to the per-/skill (avg X) column.

Current behavior on upstream main

Repro on origin/main (cf50443b):

mkdir -p /tmp/repro/.gstack/analytics
cat > /tmp/repro/.gstack/analytics/skill-usage.jsonl <<'JSONL'
{"skill":"review","duration_s":"120","outcome":"success","ts":"2026-05-25T10:00:00Z"}
{"skill":"review","duration_s":"180","outcome":"success","ts":"2026-05-25T11:00:00Z"}
{"skill":"review","duration_s":"60","outcome":"success","ts":"2026-05-25T12:00:00Z"}
{"skill":"qa","duration_s":300,"outcome":"success","ts":"2026-05-25T13:00:00Z"}
{"skill":"qa","duration_s":420,"outcome":"success","ts":"2026-05-25T14:00:00Z"}
JSONL
GSTACK_STATE_DIR=/tmp/repro/.gstack bin/gstack-analytics all

Before:

gstack usage (all time)
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  /review               ████████████████████  3 runs  (avg 0s)   <-- WRONG, should be 2m
  /qa                   █████████████  2 runs  (avg 6m)
Total time: 18m

Total time: 18m matches the real sum (120 + 180 + 60 + 300 + 420 = 1080s = 18m), confirming the total-time parser is already correct. Only the per-skill (avg ...) column is broken.

Root cause

The AVG_DUR block in bin/gstack-analytics:165-177 splits each line on the literal "duration_s": prefix and then strips non-numeric characters from parts[2]:

n = split($0, parts, "\"duration_s\":")
if (n >= 2) {
  gsub(/[^0-9.].*/, "", parts[2])      # greedy .*
  if (parts[2]+0 > 0) { total += parts[2]; count++ }
}

• Unquoted form 600,"ts":... → gsub matches at the comma → parts[2] becomes 600 → accumulates.
• Quoted form "600","ts":... → gsub matches at the leading " (position 0), and the trailing .* consumes the rest of the line → parts[2] becomes the empty string → "" + 0 > 0 is false → never accumulates → count stays 0 → average renders as 0.

TOTAL_DURATION (bin/gstack-analytics:114-125) uses -F'[:,]' field splitting and a non-greedy gsub(/[^0-9.]/, "", val), which isolates the value before stripping. That parser handles both forms and isn't affected.

Fix

Strip an optional leading double-quote with sub(/^"/, "", parts[2]) before the existing gsub so quoted and unquoted forms take the same path through the existing digit-extraction:

sub(/^"/, "", parts[2])                # NEW
gsub(/[^0-9.].*/, "", parts[2])

Three lines in bin/gstack-analytics (one new statement + two-line comment update). No producer-side changes — the parser fix accommodates both emitter formats that are already in the wild, so existing JSONL files on users' disks render correctly without migration.

Edge cases verified by hand:

Validation

After the fix, same repro renders correctly:

  /review               ████████████████████  3 runs  (avg 2m)   <-- (120+180+60)/3 = 120s = 2m
  /qa                   █████████████  2 runs  (avg 6m)
Total time: 18m

New regression test in test/gstack-analytics-avg-duration.test.ts exercises four cases (quoted-only, unquoted-only, mixed quoted+unquoted for the same skill, and total-time aggregation). I confirmed it fails on the pre-fix tree (expected "2m", got "0s") by stashing the patch and re-running, then passes after popping the stash.

$ bun test test/gstack-analytics-avg-duration.test.ts
 4 pass
 0 fail
 4 expect() calls

$ bun test test/telemetry.test.ts test/gstack-analytics-avg-duration.test.ts
 39 pass
 0 fail
 87 expect() calls

$ bun test
exit 0

Scope

bin/gstack-analytics (3 lines) + 1 new test file. No SKILL.md/template changes, no VERSION/CHANGELOG bumps (per the wave-process note in CONTRIBUTING.md). The producer-side emitters (generate-completion-status.ts, gstack-codex-probe, gstack-telemetry-log) are untouched — fixing the parser keeps existing JSONL on users' machines readable without a migration.