Codify Nemo's strategies as skills with Strategy Registry seeding

[seamus-brady]

Summary

Final PR of the sub-agent-resilience plan. Implements Fix 5: codifies the four orchestration strategies Nemo registered locally as skill content + Strategy Registry seeding, so every fresh instance has the floor patterns at boot rather than re-discovering them through CBR.

Together with PR #166 (truncation guard + auto-save) and PR #167 (referenced_artifacts + checkpoint), this closes the loop on the 2026-04-26 incident.

What changed

Two new skills

• .springdrift_example/skills/orchestration-large-inputs/SKILL.md — covers reconnaissance-first, search-then-read, and parallel-after-reconnaissance. Targets the cog loop (agents: cognitive) for orchestration decisions on large inputs.
• .springdrift_example/skills/when-to-use-writer/SKILL.md — covers synthesise-in-root judgment. Explicit "use writer when" / "synthesise directly when" criteria with the 2026-04-26 lesson written in.

SkillSeeded strategy source variant

• New variant on StrategySource distinguishes skill-derived floor strategies from operator-curated (OperatorDefined), CBR-mined (Proposed), and agent-observed (Observed) ones.
• Encode/decode pair in strategy/log.gleam updated.
• The 4-variant tag is meaningful for telemetry — operators can see at a glance what came from where.

Boot-time seeding

• New src/strategy/seed.gleam module:
  • floor_strategies() — pure function returning the four (name, description, tags) triples
  • build_seed_events(now) — turns the triples into StrategyCreated events
  • seed_if_empty(dir) — writes the events to disk only when the registry has zero events; idempotent across boots
• Wired into springdrift.gleam boot path next to librarian.init_captures.
• Operator-curated strategies and CBR-mined Proposed strategies are never overwritten — the seed only fires on a genuinely empty registry.

Plan moved to implemented

docs/roadmap/planned/sub-agent-resilience.md → docs/roadmap/implemented/sub-agent-resilience.md with a "What Shipped" preamble that documents all five fixes across the three PRs.

Test plan

• gleam build clean, no warnings
• gleam test — 2083 passed (9 new), no failures
• floor_strategies includes all four expected strategies
• All four have non-empty descriptions and the orchestration tag
• seed_if_empty writes 4 strategies to an empty registry; the resolved registry contains them with source: SkillSeeded
• seed_if_empty on a populated registry returns 0 and does NOT overwrite existing entries (regression test seeds an OperatorDefined entry first, then asserts it survives)
• Repeated seed_if_empty calls are idempotent — second call returns 0, total entries stays at 4
• Both skill files exist on disk and are readable
• Operator: delete .springdrift/memory/strategies/ on a fresh deployment, restart, verify the four floor strategies show up in <strategies> block of the sensorium on the first cycle

Notes for review

• The seeded strategies reference patterns documented in the new skill files. Both must ship together — if a future change deletes a skill file but leaves the seeding, agents see strategy names pointing at nothing. The skill-files-exist tests guard the pairing.
• SkillSeeded source vs. OperatorDefined is a deliberate distinction. An operator who reads .springdrift_example/skills/ and copies the patterns into their own config still creates OperatorDefined entries (different intent, different telemetry). The codebase-shipped floor stays as SkillSeeded so the provenance is clear.
• seed_if_empty runs at boot regardless of whether the strategy registry is configured-enabled — it's idempotent and cheap, and registries that remain disabled won't surface the strategies in the sensorium anyway. Keeps the boot path simple.

🤖 Generated with Claude Code