29 lines
2.3 KiB
Markdown
29 lines
2.3 KiB
Markdown
# Goal: Lean Athanor optimization — Spire 0.2.1
|
|
|
|
**Set**: 2026-10-05, Igor session [2026-10-05-21-athanor](../journals/2026-10-05-21-athanor.md)
|
|
**Builds on**: [lean-athanor](./lean-athanor.md) (shipped as Spire 0.2.0). Feedback from dirigent and spacecraft implementation passes on 0.2.0 — and Codex's report on how it performs — feeds into this.
|
|
|
|
## Success criteria
|
|
|
|
### Learn from the first 0.2.0 runs
|
|
- [ ] Run logs, journals and reviews from the first dirigent/spacecraft (and Codex) implementation passes on 0.2.0 are read, and what they show is reflected — fixes where the skill misled, nothing encoded from single incidents.
|
|
|
|
### Trunk and forks is tested, not assumed
|
|
- [ ] Preflight on a throwaway branch: is `fork` available; does a foreground fork block and return its result; what context and model does it start with; can it fork again; where does a background fork's completion go. Results land in `implementation/references/spawning-claude-code.md` with dates.
|
|
- [ ] A trial on a real multi-phase plan (venue: dirigent or spacecraft, chosen with the user) compares trunk notes vs. forks for complex blocks. The skill keeps forks as an option or promotes them based on the result.
|
|
|
|
### Project settings have a home
|
|
- [ ] The `spire:project` template has a place for build settings the implementation skill reads: gate and test commands, worktrees on/off, explicit-path staging, owner-model defaults per block kind, standing brief constraints, whether the user sees the build plan before the run.
|
|
|
|
### Models and roles
|
|
- [ ] Decided whether pinned 5.5 planner/reviewer role agents are worth adding, or whether the stamped-id check in Igor is enough.
|
|
- [ ] Pinky (brainstorm partner) exists as an agent pinned to Opus 4.6 — the "understanding the human" interpreter — or the decision not to build it is recorded. All other Opus-pinned Spire agents stay on 5.5.
|
|
|
|
### Small fixes
|
|
- [ ] The "Planned implementation" delegation row (Igor skill and `model-tiers.md`) says the plan must have been attacked first.
|
|
|
|
### Later (not 0.2.1)
|
|
- Mining journals for deeper insights.
|
|
- Orchestrator introspection of subagent token use / signals into running work (dirigent is exploring this). Until then, context sizes stay guesstimates.
|
|
- Swappable plan-writing approaches (grill-me, our own playbook).
|