Give your coding agent focused repository context instead of dumping the whole codebase into its prompt.
Watch SiftrCode analyze unpruned code, isolate structural interfaces, and dissolve function bodies into compact AST skeletons with zero prompt hallucination.
Select your coding agent to view instant setup commands.
Registers SiftrCode's Stdio MCP server to prune code before Claude reads it.
siftr context "task description"
Primary context optimizer with 3-stage resolution lineage & budget solver.
Alias for siftr_context for outcome-aware context bundle optimization.
Progressive disclosure expanding discrete methods to BODY with sibling exclusion.
Reports tri-state verification (verifiedSuccess: true | false | null) closing the loop.
Initializes sessions with environment discovery and tracks execution lineage.
Explainable candidate ranker with transparent multi-channel scoring breakdowns.
AST interface extractor for TypeScript, Python, Go, and Rust source files.
High-throughput concurrent AST interface extraction across target files.
Audits codebase token footprint and calculates potential team ROI savings.
Generates whole-directory pruned context packs for task-focused areas.
Audited empirical evaluation across 25 real-world engineering tasks on Express, FastAPI, and SiftrCode. ContextRank remains the active deterministic decision-maker while TypeSafe JEV runs asynchronously in shadow mode, establishing genuine continuous distributions across 500 bounded calls with 100% bit-for-bit plan invariance and zero synthetic substitutions.
| Target Codebase | Tasks | Task Types | Shadow Calls | Latency (P50/P95) | Plan Invariance |
|---|---|---|---|---|---|
| benchmarks/express-repo | 10 tasks | Bug Fix, Test Failure, Features, Refactor | 200 calls | 4ms / 6ms | 100% Identical |
| benchmarks/fastapi-repo | 10 tasks | Async Generators, Security Scopes, Routing | 200 calls | 4ms / 6ms | 100% Identical |
| src/ (SiftrCode) | 5 tasks | AST Skeleton, BudgetSolver, Candidate Retrieval | 100 calls | 4ms / 6ms | 100% Identical |
100% success • 0 failed • 0 retries
P95 446ms • Non-blocking shadow
Normalized plan bit-level equivalence
Honest measured peak concurrency
0 orphans • 0 snapshot mismatches
Metadata-only • secrets stripped
Every remote candidate evaluation passes through EnforcedEgressGateway. Raw source bodies and secrets are excluded; only string lengths, symbol paths, and scalar features are transmitted over the wire:
{
"schemaVersion": "string (12 chars)",
"task": {
"prompt": "string (73 chars)",
"evidenceSummary": "string (13 chars)",
"taskType": "string (11 chars)"
},
"candidate": {
"contextUnitId": "string (21 chars)",
"kind": "string (11 chars)",
"title": "string (11 chars)",
"path": "string (30 chars)"
},
"relationships": {},
"history": {
"coChange": "number (0.0000)",
"recentChange": "number (1.0000)"
}
}AgentEnvironment.systemConfigurationHash is the authoritative single identity binding sessions, plans, decision observations, JEV signals, and SQLite durability.
Smarter context allocation, not naive stripping. SiftrCode V2 maximizes marginal agent utility per token—cutting typical context payloads by 60%–88% while guaranteeing 100% full implementation fidelity on direct edit targets.
Full function bodies preserved for files the agent modifies or inspects. Zero missing lines, zero hallucinated edits, zero re-try loops.
Distant callers and imports synthesized into typed AST interface skeletons. Retains all interfaces, types, and docs while collapsing bodies.
Unrelated services, test fixtures, and noise files omitted safely before prompt ingestion, shielding the LLM from distraction and context drift.
100% full bodies on edit targets • 88% AST skeletons on dependencies.
Saves $0.20/turn • $3,367 net annual savings per engineer.
Prompt caching delivers up to 90% savings on repeat turns, but coding agents constantly edit code, switch branch context, and alter tool outputs. Each file mutation invalidates prompt cache prefixes, triggering expensive "cold turns" where 100% of raw input tokens must be re-written at full provider write rates.
SiftrCode V2 cuts cold payload size by 60%–88%, directly slashing the penalty of cache-miss writes. Furthermore, by segregating persistent AST interface skeletons from mutated implementation bodies, skeletons remain hot in the cache across turns while only the targeted edits rewrite cache segments.
$
npx siftrcode optimize --profile balanced
estimatedRenderedTokens calibrated against Claude BPE (~3.7 chars/tok) and GPT cl100k/o200k (~3.6 chars/tok) with measured actualProviderInputTokens reconciliation. Zero hidden context overflow.
* Modeled on typical 85k-token repository payload at 50 turns/day, 28 work days/mo (1,400 turns/mo per dev). Claude 3.7 Sonnet ($3.00/MTok in), GPT-4o ($2.50/MTok in), Claude 3.7 Thinking / o1 ($15.00/MTok in), Gemini 2.0 Flash ($0.10/MTok in).
Free open-source CLI for individuals. Paid caching for teams.
siftr context)