← graph

implement-dedicated-pi

Implement dedicated Pi evaluation lane

Metadata

Statusdone
Assignedagent-926
Agent identity6841f97a7a8e6cbc213051ce0ffb3c85c92c80cb61fd699428ab3fed6cb6ba94
Created2026-07-26T10:24:37.878437124+00:00
Started2026-07-28T10:41:12.662236444+00:00
Completed2026-07-28T11:25:35.070055836+00:00
Tokens21049722 in / 49081 out

Description

Objective

Run default bounded evaluation through a dedicated Pi agency lane rather than ordinary worker/build lifecycle machinery. This task implements the bounded no-tools evaluator and an executor-neutral adapter boundary; deep-readonly FLIP belongs to the following task.

Requirements

  • Implement the researched Pi structured JSON/RPC one-shot contract with tools/extensions disabled for bounded mode, strict schema validation, spotlighted untrusted evidence, and Pi-reported usage/error capture.
  • A content-addressed evidence manifest contains original intent, task contract, exact source-attempt route, artifact/diff summary, declared validation results, runtime events, and dependency context within explicit budgets.
  • Allocate no worktree, build-heavy admission, normal worker slot, graph/source write tool, network tool, credential exposure, or source-agent session reuse.
  • Use separate concurrency/rate/timeout/backoff accounting and operator status. Resource deferral never charges source/provider/spawn failure.
  • Persist attempt-bound exact evaluator executor/route, reasoning, renderer/schema versions, usage, verdict, and failure evidence. Consume a verdict exactly once.
  • Keep the interface executor-neutral and retain native Codex execution unchanged. Explicit Codex/Claude evaluation adapters may be selected independently, but unavailable/malformed Pi must never silently cross-fallback to another executor. Advisory failure leaves the source done with visible evidence; hard-gated failure leaves an explicit awaiting-evidence state.

Validation

  • RED tests are written first: bounded_pi_lane_uses_no_worker_or_worktree, bounded_pi_verdict_consumed_exactly_once, and pi_failure_never_cross_falls_back_executor; they fail against the pre-change path.
  • Credential-free Fake-Pi structured fixtures cover valid verdict, malformed output, timeout, process failure, duplicate delivery, route drift, usage/error capture, and adapter unavailability; fixture schema is traceable to the researched Pi API.
  • Implementation makes the RED tests pass and evaluation consumes neither worker/build capacity nor a worktree.
  • Evaluator failure leaves advisory source done and hard-gated source explicitly awaiting evidence—never ambiguously reopened or charged as source failure.
  • Capability/field scan proves no mutation/network/tool/credential authority in bounded mode.
  • A scripted real daemon + terminal/TUI human flow completes a source through Fake-Pi bounded evaluation and displays route, state, verdict/failure, and usage provenance; it is not a direct adapter unit call.
  • The permanent credential-free live scenario is added under tests/smoke/scenarios/ and registered with this task in tests/smoke/manifest.toml owners.
  • cargo fmt --check, cargo clippy, focused tests, cargo build, and cargo test pass with no regressions.

Depends on

Required by

✉ Messages 2 messages (2 unread)

  1. #1implement-lazy-evaluation2026-07-28T08:43:02.267263237+00:00delivered
    Lazy creation API is ready in src/evaluation/mod.rs: Task.evaluation_records, EvaluationRecord/Product/State, SourceCandidateRef, exact persisted EvaluationRouteSnapshot, has_authenticated_running_attempt(), LazyEvaluationSelection::resolve(), and mint_for_candidate(). Candidate completion atomically mints state=PreparingBundle or Unavailable; build your dedicated queue/runner from these records without graph tasks or worker/build slots.
  2. #2implement-dedicated-pi2026-07-28T10:41:22.077154171+00:00read
    Acknowledged — I’ll build the dedicated lane on the persisted lazy EvaluationRecord/route snapshot API, without graph tasks or worker/build slots.

Log