flow metrics · Monte Carlo forecast · kanban one-pager
Mined from git log on the working tree. Each feat: commit treated as one delivered card. Cycle Time = minutes between consecutive feat commits within an active session. Throughput = feats per active day.
State sourced from memory/v0.10.0-punch-list.md. 9 of 17 must-ship items merged. 8 open. No items currently in flight (clean WIP=0 — typical between bursts).
27 feat-to-feat intervals across 3 active days. Heavy right tail dominated by inter-burst gaps (~2d, ~4d). When in a burst, median is under 4 min — Gabriel ships in rapid-fire mode, then stops cold.
10 000 simulated futures. Each calendar day rolls active vs idle with P(active) = 3/8 ≈ 0.375 (observed base rate). Active days draw feat count from empirical sample [3, 5, 19]. Stops when cumulative feats ≥ 8 remaining must-ship items.
Smallest-to-biggest, value-first. Group A unlocks disky web, the agent-visible win. Group B is MCP polish. Group C is infrastructure (tests, docs).
| # | Item | Group | Est feats | Est days* |
|---|---|---|---|---|
| 1 | disky install-mcp subcommand (idempotent JSON patch) | infra | 1 | 0.2 |
| 2 | disky web starts FastHTML on :9000 | web | 1 | 0.3 |
| 3 | disky web renders dogfood layout + action buttons | web | 1 | 0.4 |
| 4 | Button → POST → cleanup --apply --reversible → SSE | web | 1 | 0.5 |
| 5 | MCP resources/list + resources/read | mcp | 1 | 0.4 |
| 6 | MCP prompts/list + 3 canned flows | mcp | 1 | 0.3 |
| 7 | MCP progress notifications (scan + cleanup) | mcp | 1 | 0.4 |
| 8 | tests/mcp_protocol.rs + AGENTS/CHANGELOG + smoke tests | infra | 1 | 0.5 |
* one active day per unit. Calendar days multiply by ~2.7 given current cadence.
Recovered from previous research sessions. Hard commitment for v0.10.0 stays the punch list — these are next-cycle (v0.11.0+).
--summarize scalar — shippeddisky old --older-than — shippeddisky filter --where (partial saved views) — shipped.diskyignore (already slipped to nice-to-have)df-stylekind classification beyond extensionCARGO_HOME global + sccache — kills the 1-2 GB target/ per projectlibc — 1.5-2× traversal (jwalk readdir → bulk syscall)