Summary / The 27 components
03 Repository reconnaissance
Understanding a repository's layout, build system, entry points, configs and tests before touching anything.
Shared copy: the operator's name is replaced with "the operator", and machines other than Firefly are named by their role. Nothing else is changed.
What is built now
Nothing in the harness. The code graph (code_nodes in her memory database) holds about 2,300 nodes for three of her repositories; her main runtime is not indexed.
What is missing
The manifest parser (pyproject, setup.cfg, tox, pytest.ini, CI files, every claim with a file and line), the per-workspace index refresh, and the manifest header injected into REPAIR (arm h3).
Reviewer Claude, 2026-10-04; close-out verdicts 2026-10-23
Reviewer's read: Closed 10-22: built as the hx localize stage; no detectable gain
The spec's six-signal escalation gate and ego-graph injection are Round 2. The Round 1 version is small, a 500-token header. There is a reason to doubt the gain: a 09-18 review of her worker sessions found that 530 of 559 already explore the repository with grep and find before editing. Reconnaissance may already be happening inside the session, and the real gap was cold-start workers with no memory of the codebase.
Close-out, 10-22 (pre-registered, boxed, Claude): built for real as stage S1 (localize) of the close-out bundle arm hx (A12): h0b + localize pre-pass + repro-first with failure digest + repro-triggered reflect/repair (2 rounds) + critic with one revision. hx against h0b, verified-hard 38 + pro-hard 20, 3 seeds: 103 against 94 cells (verified 65/114 against 55, pro 38/60 against 39). Gain +3.0 tasks of 58, paired bootstrap CI [−1.0, +7.0]; 4 regressions on tasks h0b always solved (limit 2); 2.9× h0b cost per cell. No detectable gain at this model. The pre-registered rule closes every bundled component together, with no halving round. Terminal: not re-run.
How it is measured no detectable gain (hx, 10-22)
Arm h3 (manifest header) against h0-v41, with file and entity recall@5 against the gold patch as a diagnostic.
Adoption rules shared by every arm: How the tests work.
Research & attempts
Score history
No scores yet. The first entry is the baseline for suite v1.
Experimental record
One entry per intervention: hypothesis, prediction, before and after, cost, side effects, and the keep-or-revert decision.
No interventions recorded yet.