Architecture Decision Records

This log captures the why behind copilot-experiments' design. Each ADR is a short, immutable record of one decision: its context, the decision itself, and the consequences. We follow the lightweight format popularized by Michael Nygard.

Conventions

• One file per decision, numbered sequentially: NNNN-short-title.md.
• Status: Proposed → Accepted → (later) Superseded by ADR-XXXX / Deprecated.
• ADRs are append-only. To change a decision, add a new ADR that supersedes the old one (and update the old one's status); don't rewrite history.
• Use 0000-template.md as the starting point.

Index

ADR	Title	Status
0001	Record architecture decisions	Accepted
0002	The filesystem is the source of truth	Accepted; amended by ADR-0015
0003	A derived SQLite index for cross-run queries	Superseded by ADR-0020
0004	The Copilot session log is the primary data source	Accepted
0005	A MockInvoker keeps the harness offline-testable	Superseded by ADR-0015 for Pier runs
0006	Separate analysis data from its rendering	Accepted
0007	Ship CLI (Rich) analysis first; defer the web explorer	Accepted
0008	--dry-run is an ephemeral, validating plumbing check	Superseded by ADR-0015 for Pier runs
0009	Copilot is always invoked with an absolute workspace path	Superseded by ADR-0015 for Pier runs
0010	Keep secrets and bulky debug logs out of stored results	Accepted
0011	Token economics from session.shutdown, costed in AIU	Accepted
0012	A task suite is an axis of an experiment	Superseded by ADR-0015
0013	Auth preflight, harness-failure status, and richer trial diagnostics	Accepted
0014	SWE-bench as a task source with decoupled Docker grading	Superseded by ADR-0016
0015	Adopt Pier for sandboxed agent evaluations	Accepted
0016	Use DeepSWE for large benchmark protocols	Accepted
0017	Import DeepSWE as a Pier dataset config	Accepted
0018	Adopt pytest-cov for local coverage analysis	Accepted
0019	Use nested Pier run directories	Accepted
0020	Remove the legacy native harness	Accepted