UNDERSTORY LABS
L2 Assess · github trending · Jul 27, 2026

affaan-m/ECC

defermedium confidence
01 · Fit to Stack
partial fit

ECC targets Claude Code specifically, which field-notes already uses for CCR orchestration, so the surface fits. But 'fit' here means installing a third-party JavaScript package that auto-applies behavioral 'instincts' and persists cross-session 'memory' inside the exact harness that runs unattended pipeline stages - that is a materially different integration than a normal library dependency.

02 · Fit to Vision
weak fit

The operator values practical over experimental, self-hosted, budget-conscious tooling that runs mostly unattended. A six-month-old, single-source-verified package that silently alters agent behavior and carries state across sessions is the opposite of the boring reliability that unattended automation needs - an upstream bug or bad update would corrupt scoring calibration across every future weekly run with no local visibility into why.

Implementation Options
01Hand-authored calibration memory (no ECC dependency)
field-notes2-4 hourslow risk

Skip installing ECC. Extract the underlying idea - persistent scoring calibration notes so the enrichment agent does not re-derive intuition every run - and implement it directly: a short CLAUDE.md/skill file in field-notes listing calibration rules already learned (e.g. 'evergreen resource repos default to signal 2-3', 'be skeptical of repos with star counts implausible for their age'), updated manually after each run.

+Zero supply-chain exposure - no unaudited package touching the harness
+Immediately actionable, fully under operator control
+Solves the actual problem (calibration drift) without adopting the vehicle
-No automatic persistence - operator has to remember to update the file
-Loses ECC's 'instinct' auto-trigger mechanism if that pattern-matching turns out to be genuinely valuable
02Sandboxed trial of ECC in an isolated session
field-notes3-5 hoursmedium risk

Install ECC only in a throwaway scratch repo/session (never in field-notes' production CCR path), run it against a handful of historical enrichment items, and inspect exactly what its 'security layer' and 'instincts' do to file system/shell access before considering any closer integration.

+Gives real evidence about whether the architecture works as described
+Contained blast radius if the package misbehaves
-Time spent evaluating a package whose headline growth metric (233,888 stars in six months, created January 2026) is not independently verifiable and is implausible for a niche dev-tooling repo
-Still requires trusting the sandbox is actually isolated from credentials/tokens on the host
01 · Value

The real, narrow value here is not ECC itself but the problem it names: field-notes' weekly enrichment re-derives scoring intuition from the rubric every run instead of accumulating it. A hand-written calibration-notes file captures that value directly without a new dependency.

02 · Why It Matters

This item is a useful test case for how the pipeline should treat agent-tooling submissions generally: single-source GitHub self-description, a headline star count that does not hold up to scrutiny for the stated creation date, and a pitch that specifically asks to be given more automatic control over the same harness the enrichment pipeline runs on. That combination should raise scrutiny, not lower it, regardless of how well the pitch matches the stack.

03 · What Implementation Looks Like

Add a short 'Calibration notes' section to field-notes' enrichment prompt or a new .claude/skills/enrichment-calibration.md file listing 2-5 scoring heuristics already learned from prior weeks (e.g. treat implausible star-count-to-age ratios as a red flag, default evergreen list repos to low signal). Update it by hand after runs where a scoring miss is noticed. No new package, no change to the CCR trigger, no automatic behavior injection.