This page traces the projectโ€™s decisions, incidents, and documentation migration.

๐Ÿ“š The two logs

LogWhat it holds
๐Ÿ›๏ธ Decision LogDated choices and the reasoning behind them, newest first
๐Ÿ”ฅ Incident LogReal failures, their root causes, and the safeguards they produced

They are worth reading together: most of the durable decisions in this project were forced by a specific outage, and the incident log names which one.

๐Ÿ•ฐ๏ธ Timeline at a glance

PeriodTheme
Early, undatedFoundations โ€” configuration unification, dependency injection, Pydantic v2, async CPU-only adapters
2025-07 to 2025-08Stabilization and the local-first turn; serialization correctness
2025-09Artifact structure, crew expansion, the zombie-process crisis and the liveness rules it produced
2025-10Observability built in response to an undiagnosable outage; total-failure incident from three missing imports
2025-11 to 2025-12Optimization and verification; the image-generation bottleneck identified and still open
2026-02SFX provider chain reworked; truth-first documentation policy adopted
2026-07Documentation migrated out of the repository into this wiki
2026-08This wiki rebuilt from source, and sixteen blockers verified against code โ€” four of them found by the rebuild itself

๐Ÿ“ฆ The pre-migration archive

On 2026-07-14 a bulk migration ingested 498 documents from the repository, emptied its docs/ and vault directories, and produced this wikiโ€™s first skeleton. Those originals still exist, unpublished:

LocationContents
wikis/stmx/source-archive/2026-07-14/305 archived originals, content-addressed by name and hash
wikis/stmx/protected-backend/32 protected sources plus their manifest
wikis/stmx/source-registry.jsonMaps all 498 inputs to their disposition and successor page
content/90-evidence/migration-manifest.json641 entries tracking archived, active and protected state

They are deliberately not published, for three reasons. They are stale โ€” several assert behaviour the code contradicts. They are heavily duplicated, with the same decision log appearing five times across parallel copies. And publishing roughly 340 additional pages would drown the explorer, the graph and search for no gain in truth.

The originals remain readable on disk for anyone auditing provenance, and ๐Ÿงพ Evidence & Provenance explains how to trace a given page back to its inputs. Everything of durable value in them has been distilled into the two logs above.

โš ๏ธ What the old documents get wrong

Worth knowing before reading any archived original, or the repositoryโ€™s remaining prose:

Claim in older documentsActual state
Output is written to a directory named for plural outputsThe configured default is out/ โ€” see ๐Ÿƒ Runbook
Various protocols and phases are โ€œ100% completeโ€Sixteen blockers are open and no pipeline run completes on this checkout โ€” see ๐Ÿ—บ๏ธ Roadmap
The suite sits at roughly 79.5% with two errorsTwelve errors, and a pass rate near 85% that varies by one flaky test โ€” see ๐Ÿงช Testing
Configuration precedence is a simple four-tier overrideThe mechanism differs from the outcome โ€” see โš™๏ธ Configuration

This is not a criticism of the authors; it is the ordinary result of documentation written per-session and never reconciled. It is also precisely why this wiki cites code paths for every claim.

2 items under this folder.