commit patterns analysis
what distinguishes threads that reach COMMITTED status?
overview
- 305 threads reached COMMITTED (6.6% of 4,656 total)
- avg turns: 57 | min: 2 | max: 506
- avg steering: 0.42 | avg approval: 1.79
structural patterns
thread length distribution
| bucket | count | % of committed |
|---|---|---|
| very_long (60+) | 101 | 33% |
| medium (11-30) | 88 | 29% |
| long (31-60) | 71 | 23% |
| short (1-10) | 45 | 15% |
longer threads have higher commit rates:
- long (31-60): 9.9% commit rate
- medium (11-30): 8.8%
- very_long (60+): 8.1%
- short (1-10): 2.7%
hunch: longer threads represent sustained, focused work rather than quick questions.
steering levels in committed threads
| steering | count | avg turns |
|---|---|---|
| no_steering | 224 | 39.0 |
| low_steering (1-2) | 70 | 91.6 |
| high_steering (3+) | 11 | 201.7 |
73% of commits happen with zero steering. but steered threads that DO commit tend to be substantially longer—users invest more effort to course-correct and still push through.
105 threads (34%) were 30+ turns with zero steering—sustained, smooth collaboration.
final message patterns
keyword frequency in final user messages:
| keyword | count |
|---|---|
| commit | 214 |
| push | 107 |
| ship | 53 |
| merge | 45 |
| pr | 19 |
| done | 19 |
| good | 10 |
| great | 7 |
| worktree | 4 |
| lgtm | 2 |
common phrasings:
- “commit and push” / “git commit push”
- “commit the files you changed/touched and push”
- “commit with bench numbers”
- “ship it”
- explicit instructions like “git add
&& git commit -m …” - spawn-style task instructions: “Migrate X to Y package…“
per-user commit rates
| user | commits |
|---|---|
| concise_commander | 137 (45%) |
| verbose_explorer | 82 (27%) |
| steady_navigator | 20 |
| swift_solver | 19 |
| feature_lead | 13 |
heavy concentration among 2 power users.
key takeaways
-
explicit directives dominate: users say “commit” or “push” explicitly. COMMITTED rarely emerges from implicit satisfaction.
-
length correlates with commits: short threads rarely commit (2.7%). the 31-60 turn range has highest rate (9.9%).
-
steering doesn’t prevent commits: steered threads that commit show high investment (91-200 avg turns). steering signals persistence, not abandonment.
-
power user effect: 2 users account for 72% of commits. commit patterns may reflect individual workflow habits more than universal signals.
-
spawn/task threads commit differently: structured migration tasks (with explicit instructions) often reach commit, suggesting task formulation matters.