Skip to content

Latest commit

 

History

History
177 lines (129 loc) · 6.31 KB

File metadata and controls

177 lines (129 loc) · 6.31 KB

Artifacts and cache

Needle reuses typed, validated repository evidence. It does not cache arbitrary model responses.

Identity boundaries

Semantic identity

Semantic identity describes what evidence means:

  • exact repository-scoped subject;
  • predicate and facets;
  • semantic world;
  • canonical payload;
  • contract and validator definitions;
  • input artifact identities when derived.

Model, reasoning, prompt, worker, transport, pricing, usage, and wall time do not normally belong to semantic identity.

Execution provenance

An execution attempt records how evidence was produced: worker profile, prompt profile, App Server version, usage, cost, commands, timing, repair, and fallback. Different attempts may produce the same semantic artifact.

Operational cache state

Resolution state records whether evidence was selected, stale, rejected, contradicted, bypassed, or recomputed. It is not part of artifact identity.

Contracts

Current semantic outputs are:

  • ImplementationLocation: primary and supporting exact code locations;
  • RuntimeFlow: ordered producer, carrier, transformation, precedence, and consumer steps;
  • FocusedTests: representative runner, argv, cwd, identifier, selection, and structural evidence.

Workers return typed proposals. Validators derive authoritative coverage from payload structure and observed evidence. A worker cannot declare that it satisfied an obligation.

Validation certificate

An artifact validation certificate binds:

  • contract and validator definition;
  • canonical subject and semantic world;
  • payload and artifact identity;
  • input artifacts;
  • exact dependency closure;
  • validator-derived coverage;
  • evidence and freshness requirements.

Changing the validator, contract, dependencies, or world invalidates the certificate for authoritative reuse.

Sufficiency certificate

A reuse sufficiency certificate proves that a selected plan satisfies the current need. It contains required obligations, selected artifacts and claims, satisfaction steps, freshness, world compatibility, contradiction state, residual state, and proof-engine definition.

Proof replay invokes no model. Unknown is unsatisfied.

Artifact-level reuse

Resolver order:

exact request
  -> artifact coverage
  -> claim coverage
  -> mixed artifact/claim plan
  -> typed partial plan
  -> fresh worker
  -> native fallback or bypass

Primary outcomes:

Resolution Meaning
ExactHit Exact artifact request identity reused
CoverageHit One artifact satisfies an equivalent semantic need
CompositeHit Several artifacts jointly satisfy the need
PartialHit Valid evidence covers part of the need; only missing obligations run
Miss No valid reusable plan exists
Stale Candidate dependencies are no longer fresh
Ambiguous Exact subject authority cannot be established
Contradicted Active evidence conflicts
Rejected Candidate failed validation or authority rules
Bypass The proof resolver cannot safely decide or policy forbids compute

Full authoritative hits require positive net reuse value. A valid but economically unsupported plan remains advisory.

Claim-level reuse

Validators can extract smaller trusted claims from an artifact:

  • one location claim per exact location;
  • one runtime-flow claim per typed step plus certified order relations;
  • one focused-test claim per representative test identifier and command.

Claim identity is content-addressed from claim kind, contract definition, and canonical payload. Origin artifact, route, worker, model, prompt, and pricing remain provenance.

Claim certificates retain subject, world, freshness, origin, dependency closure, and contradiction membership. A stale origin artifact does not automatically make an independent claim fresh: only the claim's selected dependency closure can do that.

Current authority is deliberately narrow: exact canonical positive primary ImplementationLocation, contract-complete RuntimeFlow, and representative open-world FocusedTests claims can produce ClaimHit, contribute to ClaimCompositeHit, or seed a claim-aware PartialHit after explicit bounded capability promotion. Built-in capability modes remain Shadow by default. Runtime-flow and focused-test claim authority is covered by deterministic offline validation only; no provider-backed claim-authority observation exists.

Cache scopes

SnapshotExact is the safe fallback. It requires the exact captured source snapshot.

WorktreeSemantic requires a validator-proven claim-to-dependency closure. An unrepresentable search negative, evidence gap, ambiguity, or unknown dependency downgrades scope or bypasses reuse.

Partial recomputation

A partial hit may run only operators already declared by the route plan for missing typed obligations. The worker receives:

  • missing obligations;
  • covered artifacts and claims as bounded input;
  • a restricted output schema;
  • an instruction not to repeat certified discovery.

Relevant mutation invalidates affected evidence and proofs. Irrelevant mutation preserves independent evidence. The resolver never fills an unstructured mandatory residual with guessed equivalence.

Focused tests

Identifying a representative test and executing it are separate facts. FocusedTests can be certified as located from trusted static evidence. Execution is optional unless the contract explicitly requires it. If executed, the exact approved command produces separate command evidence and may upgrade the certificate to executed.

Concurrency and failures

  • Single-flight coalesces the same artifact request.
  • Negative attempts are keyed narrowly to the exact attempt identity.
  • Plans use bounded exact bitmask selection and deterministic tie-breaking.
  • Hot immutable caches are content-addressed; SQLite remains authoritative.
  • Memory and cold paths must make the same validity decision.

Main projection

The main receives a bounded FrontierView containing selected evidence, obligations satisfied and missing, resolution, proof identity, pending needs, and a prohibition on repeating covered discovery. Worker transcripts, raw logs, and full test output never enter the main context.

Evidence and status

See Project status for current authority and validation, and Benchmarking for evidence interpretation.