coilyco-flight-deck/voice-flow-learning-loop
Archived
1
0
Fork 0
Code Issues 11 Pull requests Projects Releases Packages Wiki Activity Actions

Loop 2 layer 2.5: move snippet-frequency mining into repo-recall + Luca #1

New issue
Open
opened 2026-05-23 20:55:36 +00:00 by coilysiren · 0 comments
coilysiren commented 2026-05-23 20:55:36 +00:00
Owner
Copy link

Originally filed by @coilysiren on 2026-05-20T12:54:38Z - https://github.com/coilysiren/voice-flow-learning-loop/issues/17

🤖 Filed by Claude Code on Kai's behalf.

Problem

#16 landed a standalone Python script (scripts/snippet-frequency.py) that walks ~/.claude/projects/**/*.jsonl directly, applies filters, and counts word-n-grams. The script works, but it duplicates capability that already exists in Kai's stack:

  • repo-recall already indexes session JSONL across every project. The script re-walks raw files instead of querying the index.
  • Luca is the natural-language consumer layer over repo-recall data. The frequency-mining behavior should be a Luca tool / dispatch route, not a per-repo script.

Building it standalone fragments the substrate: every consumer that wants snippet candidates has to re-implement the JSONL walk, the harness-wrapper stripping, the subagent-prompt heuristics. Putting it in repo-recall (for the indexing primitive) + Luca (for the natural-language surface) reuses Kai's existing infrastructure and earns its place in the Luca-asker fleet.

This effectively collapses layers 2 and 3 of the original README plan: layer 2 was the standalone script, layer 3 was the Luca dispatch route wrapping that script. The right design is layer 2 = the repo-recall primitive + the Luca tool.

Scope

  1. In repo-recall: add (or surface) a primitive for "give me user-role plain-string content across all sessions, with optional repo / cwd / time filters." This is the indexer's job. The harness-wrapper stripping and subagent-prompt filtering belong here so every downstream asker gets clean signal.
  2. In Luca: add a snippet_candidates tool (or close-named) that consumes the repo-recall primitive, counts n-grams above a length floor, collapses substring overlaps, and returns ranked candidates. The natural-language surface is whatever Luca already does (dispatch route, MCP tool, ask phrasing).
  3. Migrate this repo's runs: the next mining run uses Luca, not the standalone script. The script becomes either a reference implementation kept for archive, or is deleted once the Luca path is in place.

Done when

  • repo-recall exposes the indexing primitive (whatever shape - MCP tool, JSON query, dispatch route).
  • Luca exposes the snippet-candidate tool.
  • A 2026-05-20+1 mining run lands in docs/snippets-mining-runs/ produced via Luca, with shape comparable to the layer-2 run already committed.
  • scripts/snippet-frequency.py is either deleted or marked as a reference implementation in its own header.

Why now, before layer 3

The remaining layers (4-8) are all natural-language-shaped and operate on candidate lists - length-weighted ranking, session-aware weighting, phrase-shape extractors, trigger-name proposals via LLM, conflict dedup. Every one of those is a Luca tool that takes "the candidates" as input. If the candidate-generation step lives in a per-repo Python script, every later layer has to either reimplement it or shell out. Moving it into Luca up front means layers 4-8 are pure Luca compositions, not script wrappers.

Depends on: #16 (delivered the reference implementation that defines what the Luca tool needs to produce). Unblocks: layers 4-8 (renumbered as Luca-composition layers in a followup).

_Originally filed by @coilysiren on 2026-05-20T12:54:38Z - [https://github.com/coilysiren/voice-flow-learning-loop/issues/17](https://github.com/coilysiren/voice-flow-learning-loop/issues/17)_ > 🤖 Filed by Claude Code on Kai's behalf. **Problem** #16 landed a standalone Python script (`scripts/snippet-frequency.py`) that walks `~/.claude/projects/**/*.jsonl` directly, applies filters, and counts word-n-grams. The script works, but it duplicates capability that already exists in Kai's stack: - **repo-recall** already indexes session JSONL across every project. The script re-walks raw files instead of querying the index. - **Luca** is the natural-language consumer layer over repo-recall data. The frequency-mining behavior should be a Luca tool / dispatch route, not a per-repo script. Building it standalone fragments the substrate: every consumer that wants snippet candidates has to re-implement the JSONL walk, the harness-wrapper stripping, the subagent-prompt heuristics. Putting it in repo-recall (for the indexing primitive) + Luca (for the natural-language surface) reuses Kai's existing infrastructure and earns its place in the Luca-asker fleet. This effectively collapses layers 2 and 3 of the original README plan: layer 2 was the standalone script, layer 3 was the Luca dispatch route wrapping that script. The right design is layer 2 = the repo-recall primitive + the Luca tool. **Scope** 1. **In repo-recall:** add (or surface) a primitive for "give me user-role plain-string content across all sessions, with optional repo / cwd / time filters." This is the indexer's job. The harness-wrapper stripping and subagent-prompt filtering belong here so every downstream asker gets clean signal. 2. **In Luca:** add a `snippet_candidates` tool (or close-named) that consumes the repo-recall primitive, counts n-grams above a length floor, collapses substring overlaps, and returns ranked candidates. The natural-language surface is whatever Luca already does (dispatch route, MCP tool, ask phrasing). 3. **Migrate this repo's runs:** the next mining run uses Luca, not the standalone script. The script becomes either a reference implementation kept for archive, or is deleted once the Luca path is in place. **Done when** - repo-recall exposes the indexing primitive (whatever shape - MCP tool, JSON query, dispatch route). - Luca exposes the snippet-candidate tool. - A 2026-05-20+1 mining run lands in `docs/snippets-mining-runs/` produced via Luca, with shape comparable to the layer-2 run already committed. - `scripts/snippet-frequency.py` is either deleted or marked as a reference implementation in its own header. **Why now, before layer 3** The remaining layers (4-8) are all natural-language-shaped and operate on candidate lists - length-weighted ranking, session-aware weighting, phrase-shape extractors, trigger-name proposals via LLM, conflict dedup. Every one of those is a Luca tool that takes "the candidates" as input. If the candidate-generation step lives in a per-repo Python script, every later layer has to either reimplement it or shell out. Moving it into Luca up front means layers 4-8 are pure Luca compositions, not script wrappers. **Depends on:** #16 (delivered the reference implementation that defines what the Luca tool needs to produce). **Unblocks:** layers 4-8 (renumbered as Luca-composition layers in a followup).
coilysiren added
P4
 and removed
P3
 labels 2026-05-31 07:01:25 +00:00
Commenting is not possible because the repository is archived.
No Branch/Tag specified
Branches Tags
main
No results found.
Labels
Clear labels
  
P0
now / blocking / urgent

  
P1
next / important

  
P2
backlog / will do

  
P3
someday / low

  
P4
icebox / frozen / reopen if it becomes real

No labels
P0
P1
P2
P3
P4
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference
coilyco-flight-deck/voice-flow-learning-loop#1
No description provided.