hooks pretrain --depth is ignored: depths 1, 5, 200 and 1000 give identical results, and the no-flag default differs from any explicit depth; Languages line never printed

Summary

hooks pretrain --depth <n> ignores <n> entirely. Depths of 1, 3, 5, 200 and 1000 all produce byte-for-byte identical results. Only the presence of the flag changes anything — and it changes the result substantially versus omitting it, which suggests the default path and the explicit-depth path are different code, not the same code with a default of 50.

Separately, the documented 🌐 Languages: output line is never printed.

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

Reproduce

Each run against a freshly deleted store, so the memories duplication in #185 does not confound the numbers:

run() {
  rm -f .agentic-flow/intelligence.json
  npx agentic-flow hooks pretrain "$@" 2>&1 | grep -oE 'Co-edits found: [0-9]+'
  python3 -c "import json;print('sequences=',len(json.load(open('.agentic-flow/intelligence.json'))['sequences']))"
}
run                 # no flag
run --depth 1
run --depth 5
run --depth 200
run --depth 1000

(no flag)        Co-edits found: 2     sequences=2
--depth 1        Co-edits found: 100   sequences=15
--depth 5        Co-edits found: 100   sequences=15
--depth 200      Co-edits found: 100   sequences=15
--depth 1000     Co-edits found: 100   sequences=15
-d 3             Co-edits found: 100   sequences=15

A depth of 1 and a depth of 1000 cannot legitimately yield the same co-edit analysis of the same repository. 100 also looks like a fixed cap rather than a measurement.

patterns (45), memories (6) and dirPatterns (41) are identical across every run, as expected — only the git-derived fields move.

The default path differs from --depth 50

The documented default is 50. But omitting the flag yields 2 co-edits, while any explicit --depth yields 100. If the flag simply defaulted to 50, --depth 50 and no flag should agree. They do not, so the two paths appear to be independent implementations.

--skip-git and --skip-files do work

For contrast, the sibling flags behave correctly:

--skip-git      Co-edits found: 0     sequences=0    patterns=45  dirPatterns=41
--skip-files    Files analyzed: 0     patterns=0     sequences=2  dirPatterns=18

--skip-git zeroes exactly the git-derived fields; --skip-files zeroes the file analysis. (Minor oddity: --skip-files still reports dirPatterns=18 after analysing 0 files.)

So the option-parsing layer is fine — it is specifically the depth value that never reaches the git analysis.

Languages: is never emitted

The documented output includes:

🌐 Languages: TypeScript, JavaScript, Python

grep -c 'Languages' over the output of a plain pretrain returns 0, and the field is absent from --json too. Either the feature is unimplemented or the docs describe a different version.

For completeness, the same output still shows ⏱️ Duration: undefinedms, noted in #185.

Suggested fix

1. Thread the --depth value through to the git-history analysis, and make the no-flag path use the documented default of 50 so that --depth 50 and no flag agree.
2. If 100 is an intentional cap on co-edits, surface it — a user asking for depth 1000 and receiving a capped 100 has no way to tell that from a genuine measurement.
3. Emit Languages: or drop it from the docs.

A test asserting that two sufficiently different depths produce different co-edit counts would pin item 1 permanently.

Related

• #185 — pretrain duplicates memories on every run against an existing store. Every measurement above used a fresh store to avoid it.
• #196 — the store this command produces makes hooks pre-edit routing worse than no store at all.

Start with the fresh-store shell reproduction for hooks pretrain, comparing omitted, --depth 50, and widely different depth values. Trace the CLI entry point and git-history analysis to verify that the depth reaches the analysis, the documented default is consistent, and any cap is reported. Check normal and --json output for the documented Languages: line, and add the suggested regression test for differing depths.