hooks pre-edit gets WORSE after pretrain: correct extension routing at 0.5 becomes wrong routing at 0.9; confidence is a two-valued constant, relatedFiles always empty, --json unparseable

Summary

hooks pre-edit gets worse when the intelligence store is populated. With no store, it routes correctly by file type at confidence 0.5. After hooks pretrain — the documented way to "bootstrap intelligence from your codebase" — the same file routes to the wrong agent at confidence 0.9.

Alongside that: confidence is a two-valued constant rather than a measurement, relatedFiles is empty for every source file tried, and --json emits unparseable output.

Version: agentic-flow 2.1.2, Node v22.23.0, macOS 15.

1. Pretraining degrades routing and raises confidence

Same repository, same files, only difference is whether .agentic-flow/intelligence.json (produced by hooks pretrain) is present:

WITH the pretrained store
  scripts/erp_cli.py       documentation-specialist   0.9
  src/roster.cjs           frontend-developer         0.9

WITHOUT it (file moved aside, then restored)
  scripts/erp_cli.py       python-developer           0.5
  src/roster.cjs           coder                      0.5

scripts/erp_cli.py is a Python CLI; src/roster.cjs is a Node module. The un-trained answers are right. The trained answers are wrong, and carry higher confidence.

In an empty directory with no store at all, routing by extension is entirely sensible:

same.py     python-developer         0.5
same.cjs    coder                    0.5
same.ts     typescript-developer     0.5
same.rs     rust-developer           0.5
same.html   frontend-developer       0.5
same.json   config-specialist        0.5
same.md     documentation-specialist 0.5
same.go     go-developer             0.5

(Eight files, identical one-byte contents, differing only in extension.)

So the baseline behaviour is a clean extension map, and the "intelligence" layer overrides it with a worse answer. Whatever the store contributes, it is not improving the decision — it inverts a correct one while increasing the confidence attached to it.

2. confidence is a constant, not a measurement

It takes exactly two values across everything tested: 0.5 without a store, 0.9 with one.

src/roster.cjs             0.9      src/scoring.cjs            0.9
scripts/erp_cli.py         0.9      scripts/stock_alert.py     0.9
src/weekly-dashboard.html  0.9      package.json               0.9
CLAUDE.md                  0.9

Seven files, four languages, three agents suggested — one confidence value. The README example advertises 📊 Confidence: 94.2%, which suggests a computed score; no input produced anything other than 0.5 or 0.9.

3. relatedFiles is empty for source files

The documented output includes:

📁 Related Files:
   - src/api/validation.ts
   - src/types/user.ts

That section never printed in any run. The JSON confirms the key exists but is unpopulated:

src/roster.cjs              relatedFiles=0
src/scoring.cjs             relatedFiles=0
scripts/erp_cli.py          relatedFiles=0
scripts/stock_alert.py      relatedFiles=0
src/weekly-dashboard.html   relatedFiles=0
package.json                relatedFiles=0
CLAUDE.md                   relatedFiles=2

Only a Markdown file produced any. Meanwhile the memories array is populated — with chunks of README.md, Cargo.toml and tsconfig.json when editing src/roster.cjs, at scores 0.46–0.49. Those are not related to the file being edited, which is consistent with the hash-based embedding producing near-uniform similarity.

4. --json is not parseable

agentic-flow hooks pre-edit src/roster.cjs --json 2>/dev/null | jq .

fails. 118 bytes of banner are written to stdout before the JSON, with stderr already discarded:

[IntelligenceBridge] RuVector intelligence layer initialized
[IntelligenceBridge] Features: Micro-LoRA, MoE Attention
{
  "success": true,
  ...

For a flag whose only purpose is machine consumption, this makes it unusable without a sed '1,/^{/d'-style workaround. The banner should go to stderr, or be suppressed entirely under --json.

5. The known constructor failures fire on every invocation

[RuVectorIntelligence] Attention init failed, using fallback: TypeError: MoEAttention is not a constructor
[RuVectorIntelligence] Optimizer init failed: TypeError: AdamOptimizer is not a constructor
[IntelligenceBridge] RuVector intelligence layer initialized
[IntelligenceBridge] Features: Micro-LoRA, MoE Attention

Both classes exist on the package's default export and are being read off the namespace — see #186 and #191. Worth noting the banner claims "MoE Attention" as an active feature on the line directly after MoE attention failed to construct.

Suggested fix

1. Investigate why a populated store overrides a correct extension-based route with an incorrect one. If the retrieval scores are near-uniform (as the memories output suggests), the store's contribution is noise and should not outrank the deterministic mapping.
2. Report a real confidence, or stop reporting one. A field that only ever takes two values invites callers to threshold on it.
3. Populate relatedFiles, or drop it from the documented output.
4. Keep --json output pure on stdout.

A regression test worth having: assert that pre-edit on foo.py suggests a Python agent both with and without a pretrained store. That single assertion covers items 1 and 2.

Related

• #185 — hooks pretrain duplicates its memories on every run; this issue is about what that store then does to routing.
• #186 — hooks route returns a constant researcher @ 95% with NaN factors. Same "confident constant" shape at a different entry point.
• #191 — the namespace-vs-default import that causes the constructor failures above.

Start at the hooks pre-edit entry point and inspect .agentic-flow/intelligence.json handling, comparing routing with and without the pretrained store. Reproduce the foo.py case and run agentic-flow hooks pre-edit src/roster.cjs --json | jq .; done should preserve the correct route, provide meaningful output fields, and keep JSON stdout parseable.