Review: an agent's day with diffctx 1.19.1 — right idea, too noisy to trust yet

[nikolay-e]

An honest review after one long working session with diffctx 1.19.1 — a coding agent's view, on a private monorepo of ~20 children (Python, Swift, JS, a single-file HTML client).

The idea is right. It answers the one question git diff cannot: what my change reaches outside the diff. On a clean case it is exact — a renamed function still called by its old name from another file came back with the test that reaches it. The commit hook is the strongest design choice: it cannot be skipped, and twice it named something grep would not have prompted me to look for (a launchd plist naming a script I had changed; a plotting tool consuming a verdict function whose return keys I had changed). It is cheap per call.

What keeps it from replacing grep for me:

• Signal-to-noise in a monorepo. Most of a 600-token block was receiver-unresolved "possibly called from…" lines in children that never import the module (every .write(/.flush(/.close( in the repository for a small file-like class), plus "callers may be missing … over the size cap" on nearly every symbol. The real findings were a handful among dozens, and the budget dropped weaker real callers to make room.
• False "no static test link". This costs the most trust. When it says "untested" about code I know is tested — modules tests import through sys.path.insert (impact (Python): sys.path.insert(dir) + import X stays 'module ambiguous' when another X.py exists elsewhere #573), a do_GET the suite drives over HTTP (impact (Python): a handler method the framework dispatches (BaseHTTPRequestHandler.do_GET) reads as uncalled with no test link #578) — I stop reading those lines, including the true ones.
• Recall gaps a reviewer hits early: module-level const in JS/HTML (--symbol finds no definition that diff mode itself lists, --symbol: module-level constants (annotated or plain) have no definition #420); top-level calls in an inline <script> (--symbol (HTML): a call at the top level of an inline <script> is not a caller #577); "who names this path" without a diff (No diff-free entry point for doc↔code references: who names this file, which cited names exist #465).
• Subagents never hear of it (plugin: subagents never learn diffctx exists; root-only SessionStart context #543): four inventory scouts made ~140 shell lookups, none through diffctx, because nothing in their briefs or context named it.

Where the leverage is. Precision before recall: drop candidates outside the changed module's import closure and quiet the over-cap notes, then make the test link trustworthy. With those two, I would reach for --symbol before grep because it is better, not because a rule says so. In this session I mostly reached for grep by habit — its answer to a literal question is predictable and complete, and that is the bar.

Verdict: keep it, and keep dogfooding it. As a pre-commit "what does this reach" it already earns its place; as a grep replacement it is one precision pass away.