Solve Context v0

MCP v0.3.0 is published on npm and GitHub; main protection is enforced. See the current completion checklist for fresh evidence and remaining live acceptance gates. Older checkpoints below are historical.

Status: implemented repository product foundation with real-source regression coverage; distributed in the tagged MCP v0.3.0 package; not yet proven by complete real-agent/provider measurements

Tracking epic: #898

Repository status reconciled through #920 on 2026-09-15.

Solve Context is a separate product layer for Claude Code and Codex. It is not part of SolveLang language semantics and it is not Solve Runners. SolveLang can dogfood it, but the context engine is designed for arbitrary repositories.

The product thesis is: prevent irrelevant context from being read first; compact only when necessary and only when correctness can be preserved.

Current repository source includes:

1. solvelang_context_plan — bounded deterministic task-aware planning;
2. solvelang_context_pack — content-addressed exact excerpts with source provenance and byte budgets;
3. solvelang_context_retrieve — stale-source-safe exact retrieval;
4. solvelang_context_handoff and validation — portable Claude ↔ Codex task state without automatic instruction-file mutation;
5. correctness-first structured compaction/expansion for JSON/log/diff payloads;
6. changed-path priority and bounded direct relationships from an integrity-validated supplied Solve Graph;
7. deterministic selection reasons, path/source ranges, source/excerpt hashes, omission/truncation truth and stable ordering;
8. pinned first-party real-source regressions from SolveLang;
9. independently pinned Chalk and node-fetch source subsets with equal-budget lexical/changed-path/graph-assisted comparisons;
10. an offline agent-run record/report contract for future approved real Claude/Codex measurements;
11. #920 pair-integrity rules requiring baseline and Solve Context arms to use the same fixture/provider/model/agent/record class, outcome-evaluation basis and required-evidence denominator.

The repository does not currently establish:

• verified managed-workspace installation or public Plugin Directory listing for the published MCP v0.3.0 package;
• real Claude/Codex provider-token savings;
• real-agent task-success improvement or quality equivalence across the full benchmark suite;
• provider cache-reuse improvement;
• end-to-end agent latency improvement;
• superiority over Headroom or another context product;
• a live provider proxy;
• a provenance/memory learner that is part of the released product.

Synthetic/pinned-source excerpt-byte reduction is not provider-token savings.

A context pack contains:

• a schema identifier;
• a content-addressed pack ID;
• a hash of task text rather than storing the task text as pack identity;
• a hard UTF-8 byte budget;
• measured selected bytes and explicit omission/truncation truth;
• normalized workspace-relative source paths;
• exact line ranges;
• SHA-256 identities for complete source and selected excerpt;
• stable retrieval handles;
• deterministic relevance reasons/score;
• exact selected source text.

The canonical schema is schemas/solve-context-pack.v0.schema.json.

Selection is local and deterministic. It combines lexical evidence with optional explicit changed paths and an optional supplied Solve Graph. Current behavior includes:

• local task-token/path/source matching;
• bounded lexical windows;
• deterministic whole-line budget fragments when a matching window is larger than remaining budget;
• changed-path priority before the discovery read cap;
• one-hop graph dependency/dependent/test evidence;
• bounded exported-declaration text windows for graph-selected neighbors where necessary to preserve relevant implementation evidence;
• global ranking of fragments using actual scores with stable tie handling;
• exact provenance for every included excerpt;
• explicit omissions rather than hidden summarization.

The graph input is local and integrity-validated but is not an authenticated statement of workspace freshness. The tool reports that distinction rather than pretending a checksum proves current workspace truth.

The evaluation stack now has three layers:

MCP @solvelang/mcp-server@0.3.0 is published on npm and GitHub, verified on 2026-09-20. It distributes the tagged release source; later main changes require separate distribution. Managed-workspace marketplace installation and public Plugin Directory listing remain unverified. v0.2.0 is historical.

External distribution verification remains a distinct milestone from repository source/tag state.

These are priorities, not promised dates:

1. add broader independent/blinded repository fixtures without weakening budgets/evidence;
2. fix only demonstrated selection defects exposed by those fixtures;
3. run separately authorized real Claude/Codex baseline-vs-context tasks covering all acceptance categories and both handoff directions;
4. collect truthful provider-token/latency/cache/selection/quality evidence;
5. verify managed-workspace installation and Plugin Directory listing for the published release;
6. consider optional provider-proxy or provenance/memory work only as separately reviewed expansion after core evidence is strong.

A provider proxy is not part of current v0 activation. If later approved, it must be fixture-locked for raw-byte/streaming fidelity: untouched material byte-identical, cache-hot prefixes preserved, Anthropic cache-control boundaries respected, OpenAI Responses fields such as phase preserved, numeric/tool/content ordering stable, UTF-8 handled at byte boundaries and streaming errors/termination truth retained. If safe transformation cannot be proven, optimization must fail open to the original content.

Solve Context succeeds when Claude and Codex complete coding tasks with less irrelevant context without measurable quality regression, while exact source evidence remains recoverable and provider/cache behavior stays truthful. Compression percentage alone is not the product goal.