SEA Milestones — when mindX flags an evolution moment

SEA (Strategic Evolution Agent — agents/learning/strategic_evolution_agent.py) runs improvement campaigns continuously. Most successful campaigns are routine: an audit finds nothing actionable, or a small fix lands without fanfare. A few are milestones: real evolution moments where the system shifts meaningfully.

Why the distinction

The autopublishing pipeline (docs/publications/README.md) routes routine SEA successes through the orchestrator's terse generic template ("What I learned: …"). Milestones get routed through AuthorAgent.compose_milestone_article — a richer, first-person, cypherpunk2048-voice article that pulls BDI plan id, validation pass/fail counts, audit findings addressed, and before/after metrics. AuthorAgent is the canonical writer (precedent at agents/learning/improvement_journal.py:76-87); the orchestrator owns the publish, ledger, and rate-limit.

How SEA decides

SEA._is_milestone(status, data) returns True iff all of these hold:

status == "SUCCESS" (FAILURE / NO_WORK never qualifies)
campaign_data.validation_results.passed >= 1
campaign_data.audit_results.findings_resolved >= 1
campaign_data.validation_results.resolution_rate >= THRESHOLD

Where THRESHOLD defaults to 0.8 (matching the EXCELLENT grade band at strategic_evolution_agent.py:1112). Operators can tune via env:

MINDX_SEA_MILESTONE_RESOLUTION_THRESHOLD=0.9   # stricter — fewer milestones
MINDX_SEA_MILESTONE_RESOLUTION_THRESHOLD=0.6   # looser — more milestones

The flag is set once in _conclude_campaign and travels on the existing sea.campaign.concluded coordinator event (is_milestone: true|false). No new event topic. Backwards-compatible with any subscriber that ignores the field.

How publishing routes

PublicationOrchestrator.on_sea_campaign_success checks the flag:

kind = "sea_milestone" if data.get("is_milestone") else "sea_campaign_success"

Routine sea_campaign_success → orchestrator's internal _compose_sea_article

template (300-500 words). Status defaults to publish. Rate-limited by MIN_GAP_S.

sea_milestone → AuthorAgent.compose_milestone_article(campaign_summary)

(600-1200 words). Status defaults to publish (env override: MINDX_PUBLICATION_MILESTONE_STATUS). EXEMPT from MIN_GAP_S — a milestone is rare by construction (~weeks apart) and must not be coalesced into a routine SUCCESS that happened to land in the same 6-hour window.

Trigger-id collision avoidance: routine SUCCESS uses <campaign_run_id>, milestone uses sea_milestone_<campaign_run_id>. Same underlying campaign can never end up published twice with the same id.

Conservatism

The heuristic is intentionally biased toward false negatives (missed milestones are silent; false positives become noisy public rage posts). If SEA fires zero milestones for a stretch, that is fine — the routine sea_campaign_success path still publishes the everyday improvements. The bar for "milestone" should reflect real evolution moments, not normal operation.

Tuning loop