Headline numbers
| Metric | Value |
|---|---|
| Sessions | 193 — 106 root + 87 subagent child sessions |
| Active days | 18 — every one a weekday, zero weekend sessions |
| User prompts | 374 |
| Assistant messages | 7,444 |
| Tool calls | 9,507 (9,412 completed · 91 errors · 4 still running at snapshot) |
| Skill invocations | 96 across 49 sessions (71 explicit, 25 implicit), 21 distinct skills |
| Tokens | 572.4M — of which 502.7M (88%) are prompt-cache reads |
| Cost | $8.18 (local models free; paid spend on OpenRouter models) |
By day (created)
| Day | Sessions | Tool calls | Prompts |
|---|---|---|---|
| Mon Sep 07 | 16 | 674 | 33 |
| Tue Sep 08 | 6 | 277 | 16 |
| Wed Sep 09 | 28 | 501 | 46 |
| Thu Sep 10 | 14 | 607 | 30 |
| Fri Sep 11 | 17 | 612 | 27 |
| Mon Sep 14 | 8 | 243 | 18 |
| Tue Sep 15 | 8 | 515 | 22 |
| Wed Sep 16 | 11 | 507 | 15 |
| Thu Sep 17 | 3 | 501 | 16 |
| Fri Sep 18 | 5 | 368 | 13 |
| Mon Sep 21 | 10 | 307 | 17 |
| Tue Sep 22 | 5 | 322 | 14 |
| Wed Sep 23 | 2 | 108 | 9 |
| Thu Sep 24 | 9 | 575 | 13 |
| Fri Sep 25 | 32 | 2,230 | 50 |
| Mon Sep 28 | 4 | 426 | 7 |
| Tue Sep 29 | 2 | 110 | 7 |
| Wed Sep 30 | 13 | 624 | 21 |
| Total | 193 | 9,507 | 374 |
By project (working directory)
| Sessions | Directory |
|---|---|
| 107 | ~/projects/panseg-fork |
| 40 | ~/projects/panseg-timedimension |
| 20 | ~ (home) |
| 13 | opencode worktrees (io-zarr-tiff-coverage, elf-multicut-compat, issue-146-patch-size-assert) + other panseg branches |
| 7 | ~/projects/colony_size |
| 5 | ~/projects/panseg |
| 1 | ~/test/zmesh_installation |
By agent and model
- Agents: build 91 · general 54 · explore 31 · plan 15 · (unattributed 2)
- Models (session-level): Qwen/Qwen3.8-27B local vLLM 130 ·
~z-ai/glm-flash-latest(OpenRouter) 39 · qwen3.8-27b@q3_k_xl (LM Studio) 19 · z-ai/glm-5.3 2 · mimo-v2.6-flash-free 1 · unattributed 2
Subagents
87 child sessions were created (54 general, 31 explore, 2 build), launched via 79 task + 9 subagent tool calls. The dominant pattern: code-review always fanning out into two parallel @general subagents (“Standards review” + “Spec review”), plus @explore research sessions.
Tool calls (9,507)
| Tool | Calls | Note |
|---|---|---|
| bash / shell | 5,267 | 4,667 bash (v1 sessions) + 600 shell (v2 sessions) — same tool, renamed between versions |
| read | 1,983 | |
| edit | 1,114 | |
| grep | 487 | |
| write | 139 | |
| glob | 123 | |
| todowrite | 90 | |
| webfetch | 89 | |
| task | 88 | 79 subagent launches (v1) + 9 (v2) |
| skill | 71 | explicit skill invocations |
| question | 38 | agent asked you a question |
| websearch | 9 | |
| worktree_create | 6 | +1 worktree_delete |
| execute | 2 |
Error rate: 91/9,507 = 0.96%. Errors by tool: read 22 · webfetch 20 · bash 13 · edit 12 · task 7 · websearch 5 · question 3 · glob 3 · grep 3 · skill 1 · write 1 · shell 1.
Top sessions by tool calls: viewer display input tab fork #1 (324) · micro_sam pipeline (231) · viewer display input tab (230) · issue 17 preprocessing tabs (213) · proofreading single timepoint (211).
Skills (96 invocations, 21 distinct)
Explicit = skill tool call (the agent loaded the skill itself).
Implicit = skill body injected via a command (/skill-name or skill-invoking command) — no tool call is recorded, detected by matching the SKILL.md opening against your user messages.
| Skill | Total | Explicit | Implicit |
|---|---|---|---|
| implement | 17 | 9 | 8 |
| code-review | 15 | 15 | 0 |
| wayfinder | 11 | 0 | 11 |
| grilling | 9 | 9 | 0 |
| domain-modeling | 9 | 9 | 0 |
| tdd | 6 | 6 | 0 |
| to-tickets | 4 | 2 | 2 |
| diagnosing-bugs | 4 | 4 | 0 |
| unslop | 4 | 4 | 0 |
| handoff | 3 | 2 | 1 |
| setup-matt-pocock-skills | 2 | 1 | 1 |
| customize-opencode | 2 | 2 | 0 |
| writing-for-agents | 2 | 2 | 0 |
| ask-matt | 1 | 0 | 1 |
| grill-me | 1 | 0 | 1 |
| resolving-merge-conflicts | 1 | 1 | 0 |
| research | 1 | 1 | 0 |
| skill-usage-report | 1 | 1 | 0 |
| prototype | 1 | 1 | 0 |
| opencode | 1 | 1 | 0 |
| (unknown — failed call) | 1 | 1 | 0 |
Workflow patterns visible in the data:
- wayfinder is 100% implicit — you typed it 11 times as a command, and it in turn pulls in
grilling+domain-modeling(9+9, nearly all explicit tool calls) as its standard trio. - The delivery loop:
to-tickets→implement(±tdd) →code-review(always, 15×) →unslop/handofffor polish. - 49 of 193 sessions (25%) used at least one skill; skill use concentrates in the panseg projects.
- The 1 failed skill call (Sep 25, name not recorded) is the only error among the 71 explicit invocations.
- Note: the
skill-usage-reportskill as of Sep 30 only read the v1 message tables and thus reported 95 invocations (it missed an explicitopencode-skill call in a v2 session on Sep 30). It has since been rewritten (Oct 2) to use the v2session_v2/session_messagestorage and now reports 96 for September, matching this table.
Infrastructure they provide
Glossary.md
Shared language between you and agent
Architecture Design Records (adr)
Fix choices in a findable way for agents
Ticket system
- Define where are the tickets
- Define labels
- Triage state machine
Workflows
User invoked
to-spec Plan to spec
to-tickets Spec / Conversation split into tickets
implement Take ticket and implement
wayfinder Create a roadmap to a big feature. Creates research, prototyping, and grilling tickets. Ends with map, from which a spec can be created.
Automatic
code-review Complience with the standards and spec, in parallel subagents.
grilling Plan a feature, ask the user until everything is defined.