Skip to content

ContextForge v0.6.0b1

Latest
Latest

Choose a tag to compare

View all tags
@github-actions github-actions released this 13 Oct 07:19
v0.6.0b1
e27e2f7
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

What's Changed

Added

  • Added Index v3 structural-first generations containing deterministic
    CodeMaps, a provenance-bearing relationship graph, PageRank/connectivity
    metrics, structural retrieval postings, and a complete orientation map.
  • Added sparse grounded Semantic Cards with code/documentation/config/test
    profiles, item-level validation, one bounded repair, deterministic fallback,
    priority scheduling, and a path-neutral content-addressed cache.
  • Added persisted exact/BM25/graph retrieval with CandidateCards, bounded
    agentic search/symbol/graph/map evidence planning, Working Set and diff
    boosts, and deterministic fallback on every provider or plan failure.
  • Added TokenEstimator, ContextBudget, MAP/SUMMARY/SLICE/FULL representations,
    Context Capsule v2, and stable <contextforge schema_version="2"> prompts.
  • Added public Python map/retrieval/compiler APIs, contextforge map, Capsule v2
    CLI options, read-only MCP and HTTP map/search/symbol/compile operations, and
    Bridge 2.1 with advertised artifact schemas.
  • Added benchmark required ranges, optional files, useful-token accounting,
    file precision/recall, precision@5, range/token precision, model/provider
    calls, lifecycle latency, and ungrounded-claim rate.
  • Added typed evidence/provenance projections to repository maps, normative
    OrientationMap/RepositoryMap schemas, CLI map --kind, and the additive
    benchmark index_v3_capsule pipeline with grounded/dropped-claim metrics.
  • Added off|auto|required model-assisted Evidence Plans, Bridge 2.2 planning
    negotiation, and paired same-model answer regressions with an ordinary
    full-file baseline, range-validated citations, three-vote blinded
    groundedness judging, and phase-separated token accounting.
  • Added closed task-evidence roles and ID-only CoverageLedger diagnostics across
    retrieval, planning, and materialization. Bridge 2.2, MCP, HTTP, and CLI
    compilation results expose additive evidence-coverage diagnostics without
    changing legacy Bridge or artifact contracts.
  • Added bounded multilingual planner query expansion from supplied repository
    vocabulary, including expression/result-ID diagnostics, without domain
    dictionaries or model-created source facts.
  • Added semantic-retrieval comparison reporting and per-file deterministic
    scheduler/value diagnostics, separating grounded retrieval value from model
    request cost.
  • Added source-bound function descriptions and deterministic function-level
    discovery over immutable CodeMaps, plus compact numbered source prompt export
    through the Python API and context create --compact-prompt.

Changed

  • Index build/update now publish a usable structural generation before semantic
    enrichment. Enrichment failure, timeout, or cancellation leaves that
    structural generation active; old immutable generations remain until explicit
    index clean.
  • Index/manifest/record schemas are version 4. Earlier schemas are status-readable as
    rebuild_required, excluded from retrieval, rejected by index update, and
    replaced only by a full build without semantic migration.
  • Task-based context suggest and context create default to Index v3 retrieval
    and Capsule v2 when available. --legacy-discovery and --legacy-handoff
    retain the previous flow; manual ContextPackage v1 creation is unchanged.
  • Repository orientation, architecture, conventions, and feature maps are now
    deterministic aggregations of graph, CodeMaps, and grounded cards, without
    separate repository-wide model calls.
  • Semantic enrichment now uses analyzer 7 / semantic-card-v3.4, declaration-
    aware one-to-four chunk planning, full-request/schema token accounting, one
    scheduler-owned repair authority, and lexical anchors for ranking prose.
  • Enriched retrieval generations now store a digest-bound semantic overlay over
    structural postings, preserving the public RetrievalIndex API while avoiding
    a second near-complete postings corpus.

Fixed

  • Kept descriptive verbs after "which function" as topical search terms while
    retaining explicit named identifiers and directed caller/callee requests.

  • Retained verified planner discovery after later provider failures and honored
    explicitly disabled structured-response repairs without a hidden local retry.

  • Reduced planner menus and previews while retaining available required evidence,
    resolved source anchors and complete local diagnostics.

  • Materialized mandatory source above the automatic soft target when it fits
    inside the unchanged hard context budget.

  • Stabilized config-key digests, parse-error reuse, and no-op updates; added
    Kotlin structural extraction and compact bounded occurrence/card evidence.

  • Enforced source-spanning planner previews, session-wide round/action/token
    ceilings, strict whole-plan fallback, model-order materialization, and safe
    FULL → SLICE → MAP downgrade without weakening freshness or hard budgets.

  • Counted response schemas embedded by plain-JSON providers in preflight and
    bounded every agentic round against the provider's complete context window,
    shrinking candidates, module inventory, and previews before dispatch.

  • Corrected paired efficiency measurement to compare Capsule input with the
    ordinary full required/working-file payload rather than the hand-selected
    oracle, while preserving the oracle and blinded judge as quality gates.

  • Added verified non-call reference edges, module-level environment-key facts,
    scoped config-consumer projections, and centrality metrics that exclude both
    inferred and synthetic source-test relationships.

  • Grounded model-inferred relationships in supplied content-addressed targets
    and source evidence, with item-level validation and source/target rename
    rebinding. Updated semantic priority scheduling and behavioral barrel routing.

  • Prevented centrality-only task material, added bounded planner representation
    influence, and enforced a 30% automatic Capsule soft ceiling without weakening
    explicit-material or hard-budget rules.

  • Restricted retrieval graph scoring to verified/best-effort structural edges,
    while retaining inferred neighbors as non-authoritative CandidateCard context.

  • Allocated automatic capsule sections inside the post-envelope soft payload,
    seeded complementary MAPs before upgrades, and penalized duplicate concepts,
    ranges, and graph neighbors during marginal-utility selection.

  • Reported all current v3 repository maps after builds and added a digest-bound
    generated-artifact registry so unchanged package/Capsule/prompt outputs do
    not enter later index updates.

  • Serialized generated-artifact registry updates with a bounded, ownership-
    checked stale-recoverable lock so concurrent writers do not lose entries.

  • Completed twelve-language Python/Tree-sitter import/call/reference extraction,
    ambiguity-aware
    provenance, and key-digest-scoped config-consumer relationships without
    storing configuration values.

  • Sharded graph and retrieval records into digest-bound 4 MiB artifacts,
    removed duplicated generation facts, and made warm retrieval load compact
    file projections plus embedded grounded evidence instead of every CodeMap and
    Semantic Card.

  • Replaced fill-to-30% automatic compilation with minimum-sufficient Evidence
    Plan coverage; the 30% allocation is now a soft ceiling while explicit
    material remains governed by the hard budget.

  • Reused digest-matched published Semantic Cards across unrelated structural
    updates, avoiding retries of unchanged fallback/partial files while still
    rebinding or dropping inferred targets against the current snapshot.

  • Distinguished the closed Semantic Card evidence-ID set from untrusted source
    container IDs, preventing compatible providers from returning transport IDs
    that cannot ground ranking claims.

  • Decoupled Bridge progress delivery from index cancellation: queue overflow
    coalesces intermediate events, counts dropped/coalesced updates, and always
    prioritizes terminal state. A slow progress consumer cannot cancel a job.

  • Separated Bridge client wait timeout, per-provider-attempt timeout, and whole
    operation timeout. Client timeout leaves the job tracked by operation_id
    with its writer lock held through completion/cleanup; explicit cancellation
    and shutdown still cancel it.

  • Resolved relative TypeScript emitted import suffixes only through the declared
    .js/.jsx/.mjs/.cjs to source-suffix table. Exact files still win,
    package imports remain external, and ambiguity remains unresolved with
    best-effort structural provenance.

  • Replaced the isolated Python test-path policy with a declarative shared file
    policy. Source-test edges now prefer resolved test imports, calls, and
    references, carry provenance, and remain excluded from centrality.

  • Constrained entrypoint-handler edges to verified callback or direct bootstrap
    shapes instead of promoting every entrypoint import/call.

  • Made every planner action demonstrate a measurable evidence-coverage gain;
    no-gain actions cannot extend a bounded session.

  • Validated plan sufficiency after actual capsule materialization. Missing
    planned items, evidence/ranges, mandatory role coverage, or task material now
    downgrade effective sufficiency with bounded reason codes.

  • Made automatic capsule selection coverage-first and prevented upgrades from
    displacing the final evidence for a required role. Added an automatic compact
    profile for one short exact file only when it is cheaper than the ordinary
    verified capsule.

  • Bound --force-reanalyze to semantic cache/card reuse. Structural reanalysis
    has an explicit internal control, and a semantic-force no-op does not publish
    an unchanged enriched generation.

  • Strengthened paired answer evaluation: assertion evidence support and
    lexical/identifier support are distinct from citation containment, and the
    blinded judge sees only answer assertions and cited material ranges.

Maintenance

Full Changelog: v0.5.1...v0.6.0b1

Contributors

waterflane
Assets 6
Loading