Reposition searches saved issues, pull requests, comments, and changed files, then brings cited evidence into your next review. Find related work, inspect references to prior fixes, and see what the cache does and does not cover.

This is an early local tool with a Python API and CLI. The default uses SQLite FTS5 and needs no model, server, credentials, or third-party Python package. Nothing is posted to GitHub, and nothing from a cache is executed.

Python 3.10+ with SQLite FTS5 support is required. From a checkout, one command runs a complete synthetic example:

./reposition demo

It indexes the included sample in memory, shows cited search results and follows a reference to a proposed fix. No installation, credentials or downloads are needed, and no database is saved. To try another query:

./reposition demo 'advisory metadata'

For the immutable triage-cache workflow, ./reposition demo --cache creates a temporary fixture, indexes it, searches it and retrieves a larger verified source fragment. It handles the snapshot, corpus and fragment IDs and removes its files when finished.

Starting without a checkout? Clone it, enter it and run the demo:

git clone https://github.com/Univeracity/reposition.git
cd reposition
./reposition demo

On Windows, use py reposition demo; elsewhere, python3 reposition demo also works. Installed packages provide the same demo as reposition demo.

Ready for your own data? See using Reposition for indexing, installation and filters, or the immutable-cache workflow for triage-o-mator snapshots and corpora.

• Title-weighted FTS5 search over source text, with component filters and one highest-ranked component per distinct item.
• Stable snapshot digests derived from actual records and coverage metadata.
• Source URI, revision, exact excerpt offsets, and SHA-256 hashes for each citation.
• Visible candidate caps, omitted hits, excerpt truncation, and unknown coverage.
• Incoming and outgoing literal references, including missing or external targets. A phrase such as “fixes #40” is retained as a source claim with its surrounding wording; negated or uncertain wording remains a generic reference.

Leading #40 anchors select that item's summary before lexical results. Quoted phrases are preserved; ordinary terms use any-term matching. Results are ranked evidence, not automated duplicate decisions or calibrated confidence scores.

The cache-index, cache-query and cache-retrieve commands work directly with triage-o-mator's immutable evidence cache. They provide component-aware ranked fragments, exact snapshot/corpus selection, source verification and progressive reads across all nine components. The complete JSON response has an enforceable UTF-8 byte budget. Derived indexes live beside the acquisition cache and have their own storage ceiling.

See the workflow and contracts, including resource limits and an optional bin/cache integration patch. This API keeps acquisition and approvals with the caller; real-cache and independent review evaluation remain the next qualification steps.

python -m pip install -e '.[tokens,tfidf]'
reposition search 'archive integrity' --tokens 1024 --encoding o200k_base
reposition search 'archive integrity' --method tfidf --json

The budget measures the complete formatted evidence text, including citations, headers, coverage, and footer. With --json, the structured JSON wrapper and full ranked records are additional output; pass evidence.text to a bounded consumer. Choose the actual consuming tokenizer encoding. Character budgets never claim to be token budgets.

TF-IDF is an experimental comparison arm, fitted once per loaded snapshot using the same SQLite tokenizer and title/text fields. It does not replace FTS5. The initial experiments found mixed tradeoffs; experiment notes explain the evidence and its limits. Semantic search and Vyral integration are possible extensions, not current runtime requirements.

from reposition import Index, load_snapshot, render

snapshot = load_snapshot(
    "examples/github-cache.json", format="github", repository="example/packages"
)
with Index("repo.sqlite") as index:
    index.import_snapshot(snapshot, replace=True)
    result = index.search("same filename older bytes", component="summary")
    evidence = render(result, budget=8000)
    print(evidence.text)

A TUI or agent can use SearchResult and Evidence without parsing CLI prose. Reposition is currently offline: acquisition, incremental updates, and review actions belong to the calling workflow.

python -m pip install -e '.[dev,tokens,tfidf]'
python -m unittest discover -s tests -v
ruff check .
python -m build
python scripts/check_dist.py
python scripts/smoke_wheel.py dist/*.whl
python benchmarks/evaluate.py examples/github-cache.json examples/cases.json \
  --format github --repo example/packages --methods fts5 tfidf --tokens 1024

The benchmark reports retrieval recall and citation retention separately at the chosen complete evidence budget (--chars or --tokens). Each case includes the rendered text, source-bound excerpts, omitted hits, and rendering time. Citation retention does not establish that an excerpt supports a correct review decision; independent review is still needed. The report's JSON wrapper is outside that evidence budget.

See design and next steps. Source code is MIT licensed, with the adapted Vyral query policy under Apache-2.0 and the adapted triage-o-mator evidence validator under MIT as described in NOTICE. This project is separate from the GitHub-to-Notion CLI also called Reposition.