Skip to content

v0.16.0 — Memory-driven reasoning effort, code smells, agent-doc discovery, Codex plugin

Choose a tag to compare

View all tags
@mliotta mliotta released this 02 May 22:48
· 120 commits to main since this release
v0.16.0
3d1bb22
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Highlights

This release shifts neo from "reasoning helper that retrieves memory" to "reasoning helper that monetizes memory." Familiar queries get cheaper; hard novel queries get more thinking. The architectural feedback loop closes natively. And the prompt now sees what your AI tools have already documented about your codebase.

Added

  • gpt-5.5 default, replacing gpt-5.3-codex. Routes through the existing /v1/responses path; reasoning-plus-message output shape handled unchanged.

  • Memory-driven reasoning effort for OpenAI gpt-5* models. Each query's reasoning.effort is sized from the strength of the memory hit:

    Memory + difficulty Effort
    ≥3 patterns, avg confidence ≥ 0.8 low
    Some patterns, avg confidence 0.5–0.8 medium
    No relevant patterns OR avg confidence < 0.5 high
    No patterns AND difficulty == "hard" xhigh

    Cap with NEO_REASONING_EFFORT={none,low,medium,high,xhigh} for cost control. End-to-end measurement on a familiar query: Reasoning effort: low (patterns=5, avg_conf=0.91) — neo's learning monetizes directly into inference cost.

  • Code-smell detection during context assembly. Surfaces TODO/FIXME/HACK/XXX markers, Python stubs (pass / ... / raise NotImplementedError), bare except:, swallowed exceptions, and hardcoded credentials (OpenAI / AWS / GitHub / Slack token shapes) in the prompt under "KNOWN ISSUES IN NEARBY CODE." Per-file cap of 8 + global cap of 20 keeps growth bounded.

  • Project-local agent instruction discovery — automatic, no config. Reads from every major AI tool ecosystem:

    • Claude / Claude Code: CLAUDE.md, .claude/CLAUDE.md, .claude/agents/*.md, .claude/commands/*.md
    • Codex / AGENTS.md spec: AGENTS.md, .github/AGENTS.md, .codex/**/*.md
    • Cursor: .cursorrules, .cursor/rules/**/*.md and .mdc
    • GitHub Copilot: .github/copilot-instructions.md
    • Windsurf, Continue, Augment, Spec Kit, Aider, Codeium

    Surfaces under "PROJECT-LOCAL AGENT CONTEXT" — guaranteed inclusion regardless of relevance ranking.

  • Architectural quality feedback loop. At session save, neo snapshots three structural metrics — import-graph cycles (Tarjan SCC), god files (LOC + function-count thresholds), max nesting depth — and at outcome detection it diffs against current state. A regression weakens the accept/boost or strengthens the modify/penalty by 0.1; an improvement does the reverse. Confidence becomes a signal of "helped the codebase," not just "got accepted."

  • Codex plugin at plugins/neo/ with the same six skills as the Claude Code plugin ($neo, $neo-review, $neo-optimize, $neo-architect, $neo-debug, $neo-pattern). Repo marketplace at .agents/plugins/marketplace.json — install via codex plugin marketplace add Parslee-ai/neo. Persistent semantic memory in ~/.neo/ is shared across both plugins.

Test surface

71 net new unit and integration tests across the new modules: reasoning_effort, code_smells, agent_context, architecture_metrics, plus arch-delta modulation in fact-store. 562 tests total green on Python 3.11+ at release time.

Tag

v0.16.0 from commit 3d1bb22 on main.

Assets 4
Loading