docs(D-SAG-4): graph-backed Claude Code context and ONA feedback after #1509

[AdaWorldAPI]

D-SAG-2: addressable context + introspective execution evidence

Research-only draft successor to merged #1509. Starting from the real SoftwareGraph (one-hop, in-memory Quack-backed projection), specify how its compiler-grounded entities might become a local, addressable context provider for Claude Code, and how ONA's event→anticipation→operation→feedback→revision ideas might help rank and test legal optimization proposals.

No new query engine, ISA change, method registry, autonomous edits, or claim of token-free Claude. Local graph queries need no LLM inference, but model-visible results consume tokens.

Added

• D-SAG-2 architecture + falsifiable staged plan: qualified source-symbol vs semantic-method vs execution-evidence graphs, revision/worktree overlays, bounded context/impact/propose queries, source identity, privacy/prompt-injection limits, current one-hop/multi-hop blocker, low-overhead receipts, entropy/MDL boundaries, token/accuracy and optimization baselines.
• Ready-to-use Claude Code implementation prompt: exact source inspection order, minimal one-method executable vertical slice, typed refusals, independent oracle, disable-red negatives, measurement and stop/go criteria.
• .arxiv ONA + NARS primary source excerpts: pinned ONA commit dc4efd0; NAR, Cycle, Memory, Decision, Inference implementation functions; separated 2018 self-theory and 2020 ONA implementation; anticipation/negative confirmation and evidence revision caveats.
• .arxiv repository-code-graph excerpts: RepoGraph 2410.14684, RepoHyper 2403.06095, RANGER 2509.25257, KG code generation 2505.14394, and LSP as competing source index.

Reality check at #1509

• Quack/mask-risc answers actual one-hop questions over methods projected from MethodTable.
• No verified whole-repository symbol index, persistent graph, general two-hop or runtime receipts.
• H1's scalar/reference timing does not prove fused SIMD optimization.
• NAL-1..9 × 34 recipes audit: MCP receipt probe, revision laws against OpenNARS #1495's MCP receipt-reader gap is a promising bridge; Round 6: one SPOFC → do → learn cycle compiles onto existing folds and revision #1359's observe/revise/learn witness is narrower than a general self-optimizing policy.
• The audited seven-language Bible graph was planned, not implemented in the codebase; reuse semantic identity vs surface spelling only.
• SoC here means separation of concerns.

Required falsifiers in a future implementation

• Same-label/different-ClassId; stale source/body/seal; unresolved macro/dynamic dependency; unsupported multi-hop; forged Verified from Markdown; source prompt-injection kept inert.
• Receipt replay/overlap, false negative anticipation without a closed observation window, recursion from instrumenting the observer, cross-tier benchmark mismatch, semantic violation from a fast proposal.
• Compare current CLAUDE.md+grep, conventional source index, Quack graph slice, and hybrid with independently validated dependency recall, patch correctness, token count, cold index cost and latency.
• Compare ONA-inspired method-selection against simple static and historical empirical cost models; no performance claim without actual experiment.

Expanded, deliberately deferred research contracts (follow-up commits)

The initial four-file draft left several mechanisms at sketch level. The same PR now adds:

• Deferred technical contracts: exact D-SAG-1: software-awareness graph — lance-graph's own methods, programs and evidence as Quack-queried rows #1509 SoftwareGraph boundaries; source-symbol/semantic-method/execution identity; commit+dirty-worktree invalidation; snapshot-bounded resolve/context/impact/propose/expand/explain; token-budgeted foveated evidence selection; explicit uncertainty, omission and refusal; source-prompt-injection boundary; deterministic proposer truth grades; receipts, speculation vs measurements, recursive-instrumentation suppression; 15 non-vacuous negative fixtures; local retrieval/CPU/learning cost accounting and stop/go gates.
• Additional .arxiv attention and active-selection excerpts: Liu et al. 2307.03172 (context position), RepoHyper/RepoGraph retrieval, Song & Xu 1701.06725 (costly observation), Shi et al. 2606.09577 and Rakasi et al. 2608.14659 (uncertainty versus executable verification).
• Parent plan and Claude prompt now link to those additional contracts with an explicit instruction to implement W1 first, not build the entire speculative architecture.

Unshipped on purpose: no repository-wide index, source LSP adapter, context/MCP/Claude Code integration, multi-hop engine, optimized SIMD kernel, receipt pipeline, ONA runtime or self-modifying agent was implemented here. Literature excerpts are paraphrases and carefully bounded; no new result is measured by this PR.

Verification of this PR

Docs only: six new files plus links in the original plan and prompt; no Rust/runtime modifications or benchmarks claimed. Source links and branch comparison inspected. Execution gates live in the prompt and remain unrun.

Reading: Hammer & Lofthouse 2020, ONA architecture, pinned ONA, Wang, Li & Hammer 2018. Related: #1509, #1508, #1504, #1495, #1359, #1311.

Decision gate: no wider integration unless context retrieval preserves source-backed correctness and reduces model/context costs; no learned optimization machinery unless it outperforms simpler selection baselines.

[coderabbitai]

Warning

Review limit reached

Your organization has reached its usage spending cap. Adjust your spending cap in the billing tab.

Next included review available in 31 minutes.

Check out review usage here.

View limit details

Limit details: You’ve used the included review currently available. Your 62 included PR review attempts over the past 7 days set your current allowance at 1 review per hour.

Learn how review limits work.

Review configuration:

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: CHILL
• Plan: Essentials
• Run ID: 7bdc60c9-eb5f-4ec0-b7cb-143d390df06f

---

📥 Commits

Reviewing files that changed from the base of the PR and between 42fa2a3 and d1babe7.

---

📒 Files selected for processing (6)

• .arxiv/2026-10_sag_agent_attention_active_selection_excerpts.md
• .arxiv/2026-10_sag_ona_internal_experience_source_excerpts.md
• .arxiv/2026-10_sag_repo_graph_context_source_excerpts.md
• .claude/plans/software-awareness-agent-context-deferred-contracts-v1.md
• .claude/plans/software-awareness-agent-context-ona-v1.md
• .claude/prompts/d-sag-2-ona-agent-context-implementation.md

---

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Warning

Billing warning: we have not been able to collect payment for this subscription for more than 72 hours. Please update the payment method or pay any pending invoices in Billing to avoid service interruption.

---

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-12T19:27:42.467295Z	0da02da	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[AdaWorldAPI]

Superseded by #1523. Every commit from this branch is merged there unchanged except for the alignment edits.

Changes for this PR:

• propose() now returns ProposedMethod[], the canonical type wrapping main's awareness::Candidate.
• impact() now takes a ProposedMethod.
• The prompt file is renamed to d-sag-4-ona-agent-context-implementation.md to match its D-id.

Closing in favour of the capstone.

---

Generated by Claude Code