Fuse
Project

Roadmap

The planned direction for Fuse, framed as intent rather than commitment, with shipped work recorded in the changelog.

The roadmap states where Fuse intends to go, ranked by impact on agent efficiency. Each planned item names its primary lever: accuracy, token cost, or round-trips. Nothing here is a commitment or a delivery date. Work that has shipped is recorded in the Changelog, not on the roadmap.

This page is for stakeholders gauging direction and engineers deciding where new work fits.

The roadmap states where Fuse intends to go, ranked by impact on agent efficiency. Each planned item names its primary lever: accuracy, token cost, or round-trips, plus rough effort and value so stakeholders can compare options. Nothing here is a commitment or a delivery date. Work that has shipped is recorded in the Changelog, not on the roadmap.

This page is for stakeholders gauging direction and engineers deciding where new work fits.

Effort and value (how to read)

Effort is a t-shirt size for one engineer who already knows Fuse, including tests and docs, not a calendar commitment.

SizeTypical spanMeaning
S2 to 5 daysLocalized change, low integration risk
M1 to 2 weeksOne subsystem, harness regression expected
L2 to 4 weeksCross-cutting or tuning-heavy
XL1 to 3+ monthsNew subsystem, ML, or agent eval loop

Value rates expected impact on agent efficiency for the pinned benchmark corpus and typical .NET agent workflows. It is not a measured forecast.

ValueMeaning
HighLikely large move on recall, precision, tokens, or round-trips for many tasks
MediumClear win in a subset of repos or modes; smaller headline benchmark shift
LowProduct polish, guardrails, or measurement without direct retrieval lift

Re-bundling a large ONNX model, full MSBuild compilation graphs, and end-to-end agent eval each add weeks beyond the table estimate.

Planned Themes

ItemEffortValuePrimary lever
Auto-scope tool (task + budget, picks skeleton / focus / search internally)L to XLHighRound-trips; accuracy when agents pick the wrong mode
Session-delta emission (only new material across turns)LHighToken cost on multi-turn sessions
Hybrid retrieval with learned embeddings (opt-in rerank on BM25 top-K; sideload or small bundle, not mandatory ~90 MB default)L opt-in; +M to L if bundled againHigh on query mode for large repos with weak lexical overlapQuery recall (illustrative: Newtonsoft.Json query recall ~32% today)
Eval harness (task-success: tokens-to-correct-answer with and without Fuse)XLHigh for prioritization; Low direct user impact until usedConfidence; steers all other items

Scoping Recall and Precision

BM25-only query scoping today reaches about 49% mean recall and 2% mean precision at a 50,000 token budget on the pinned PR corpus; change scoping with a git base reaches 87% recall at 50% precision. The tables below target that gap without returning to a full-repo dump.

Retrieval and seeds (recall on query mode)

ItemEffortValueExpected effect
Query expansion from the symbol index (stems, *Validator-style patterns, path tokens)MMedium to HighBest ROI without ML; helps NL queries that omit type names; modest mean recall lift, larger on FluentValidation-style repos
Dual recall (BM25 plus embeddings, sparse retrieval, or symbol-alias merge)L to XLHighBroad query recall lift when second signal is semantic; cost scales with chosen backend
Cross-encoder or lightweight rerank on top-KL to XLHigh on hard reposTargets AutoMapper and Newtonsoft.Json gaps where grep vocabulary beats BM25 today
Optional query rewrite (task description to keywords and type hints)MMediumHelps prose-heavy agent prompts; LLM-backed rewrite adds policy and cost

Primary lever: accuracy on query mode. Secondary: round-trips when one call replaces blind grep loops.

Dependency graph (recall without false expansion)

The syntax-level Roslyn graph is the ceiling for focus, changes, and query expansion. Large repos with conditional compilation and partial classes trail smaller ones in recall.

ItemEffortValueExpected effect
Tighter false-edge control (less noise from stripped comments and literals)MMediumFewer wrong expansions; helps Newtonsoft.Json-style graphs
Edge confidence and typing (strong vs weak edges, weighted expansion)M to LMedium to HighBetter precision at same budget; steadier recall on large corpora
Compiled reference resolution (MSBuild or Roslyn when a solution exists)XLHighStructural ceiling lift for cross-assembly and resolved types; maintenance heavy

Primary lever: accuracy on changes and focus, especially on large corpora.

Expansion and ranking (recall vs precision trade)

Expansion protects recall by pulling neighbors; neighbors outside the ground-truth changed set hurt benchmark precision even when they help an agent.

ItemEffortValueExpected effect
Tighter admission thresholdsS to MMediumHigher benchmark precision; risk of recall drop if over-tightened
Adaptive depth (deepen only high-scoring branches)MMediumBetter recall per token than fixed depth everywhere
Selective reverse hops for query (callers/tests when seed is strong, capped)MMediumRecall lift for tests and callers without focus-wide broadening
Two-phase expansion (must-have deps first, optional neighbors if budget remains)M to LMedium to HighPrecision at fixed budget with less recall sacrifice than blunt depth cuts
Learned or feature-based reranking after expansionL to XLMedium to HighFeature plumbing M to L; learned rank XL; rescores tail before budget cut

Primary lever: precision at a fixed budget without sacrificing headline recall on small repos.

Emission tiers (precision inside selected files)

File selection can stay recall-oriented while emission trims noise inside each file.

ItemEffortValueExpected effect
Stronger member-level packing for queryS to MMediumSame file recall, fewer tokens and less sibling noise in matched files
Path and test policy (tests only when query, seed, or change signals them)S to MMediumPrecision and token savings; may drop useful test context if heuristics are crude
Score-tiered emission (full / outline / skeleton by rank within the included set)M to LHighLarge token reduction at same file set; improves effective precision per token

Primary lever: token cost and effective precision (less noise per token) at unchanged file recall.

Workflow and product surfaces (agent round-trips)

ItemEffortValueExpected effect
Task-shaped defaults (PR and branch work steers to change scoping when a git base exists)SHighNo algorithm change; moves many PR tasks from ~49% query recall to ~87% changes recall on the harness
Guided multi-call flows (TOC or skeleton, then focus or changes on large repos)S to MMedium to HighExtra round-trip but better accuracy when single query at 50k is insufficient

Primary lever: round-trips and accuracy in live agent sessions without requiring 100% single-call recall.

Measurement

ItemEffortValueExpected effect
Per-repo recall dashboards in harness outputSLow direct; High for devSurfaces Newtonsoft.Json and AutoMapper regressions hidden by the mean
Layer 2B regression guard when retrieval shipsS; M if wired to CI compare gateLow direct; High for devCatches lexical localization regressions (grep baseline is 58% on 12 questions today)
End-to-end task success (see Planned Themes eval harness)XLHigh for prioritizationValidates that harness recall moves translate to agent task success

Primary lever: confidence that accuracy work improves real tasks, not only harness aggregates.

Phased bundles (effort vs value)

Illustrative sequencing for one engineer. Calendar time overlaps if items run in parallel.

PhaseScopeEffort (cumulative)ValueNotes
A Quick winsQuery expansion, admission tuning, tiered emission v1, task-shaped MCP defaults, per-repo harness reports~3 to 5 weeksHigh value per weekNo ML; best first release
B Expansion and graph hygieneAdaptive depth, two-phase expand, edge weights, false-edge fixes~4 to 8 weeks after AMedium to HighLarge-repo focus and changes/focus more than raw query
C Semantic retrieval v2Opt-in hybrid rerank on BM25 top-K (no mandatory bundle)~6 to 10 weeksHigh on query on hard reposReuses prior hybrid design lessons; packaging optional
D Compilation graphMSBuild or Roslyn references where available~3 to 6 monthsHigh structural ceilingIndependent track; edge cases and maintenance
E Learned ranking and task evalLTR or cross-encoder plus agent eval harness~3 to 6+ monthsHigh when eval proves liftDepends on eval design

Fastest benchmark moves: task-shaped defaults (S effort, High value, already shipped capability); query expansion plus top-K rerank (M plus L, High value on query); tiered emission (M to L, High token value). Slowest but deepest: compiled graph (XL) and full task eval (XL).

Sequencing intent (not dates)

Near-term emphasis: phase A items. Medium-term: phase B and opt-in hybrid rerank (phase C). Long-term: compilation graph (phase D), dual recall, learned ranking, and task eval (phase E).

Recently Shipped (no longer on the roadmap)

These items moved to the Changelog when they shipped:

  • Roslyn structural analysis for C# (skeleton, dependency graph, type location, outlines, route maps, semantic markers) is the default path; regex remains for content reduction, the project graph, and pattern detectors.
  • Symbol-level scoping (Type.Member focus seeds and member-granularity query retrieval).
  • A persistent SQLite store at .fuse/fuse.db for reduction output, analysis, and relevance tokens.

How This Relates to Shipped Work

The roadmap describes intent. Once an item ships, it is removed from the roadmap and documented in the Changelog with the version that introduced it. To learn what Fuse does today rather than what is planned, read the changelog and the reference pages.

What This Does Not Cover

This page does not specify designs or release dates beyond the phased bundles under scoping recall and precision. Effort sizes are illustrative, not estimates tied to staffing or calendar. Shipped behavior is documented in the Changelog.

Next

Read the Changelog for the history of shipped work, or Contributing to propose or build a planned item.

On this page