Skip to content

RELEASE_NOTES.md

Latest commit

 

History

History
62 lines (50 loc) · 3.33 KB

RELEASE_NOTES.md

File metadata and controls

62 lines (50 loc) · 3.33 KB
Raw

Imprint — Release Notes

v2.0.1 — CI/security patch

  • Bump CI/release Go toolchain 1.25.10 → 1.25.11, clearing the stdlib vulnerabilities GO-2026-5039 (net/textproto) and GO-2026-5037 (crypto/x509). No functional changes; rebuilds the v2.0.0 binaries on the patched toolchain.

v2.0.0 — Token economy, three-tier memory & Codex support

The biggest release since the initial plugin: Imprint becomes a token-economy + high-value-memory plugin, and gains first-class Codex support alongside Claude Code.

Token economy meter (Phase 1)

  • Append-only token_ledger + injection_log (migration 006), idempotent saving credit.
  • Every background LLM spend point instrumented (compress, consolidate, graph, reflect, summarize, root) → ledger + budget.
  • Budget ceiling (per session/day) that pauses background work and trims injection before overspending, without breaking the main path.
  • GET /imprint/economy (plan-aware: currency for API, "fôlego" for subscription) + Economy UI tab. "Memory used" via cheap file/concept co-occurrence.

Three memory layers + intuition (Phase 2)

  • Explicit layers: base (compressed observations) → refined (memories) → rooted (intuitions) (migration 007).
  • Intuitions are born only by convergence of many insights, injected resident at max priority, auto-weaken on contradiction, with a hard residency cap.
  • Inspection screen (always visible, manually demotable/deletable) — invariant 11.
  • Lazy on-demand injection (/imprint/inject/lazy) and precedence handling (specific refined overrides generic intuition).
  • Memory governance (A5): export, purge-by-repo, full reset — drops docs from search indexes too.

Spend Haiku only on what becomes insight (Phase 3)

  • Pre-compression importance gate: trivial observations are captured deterministically (regex pre-pass, no LLM) into the base layer.

Code graph as a relevance signal (Phase 4)

  • BlastRadius over the existing graph boosts lazy injection — editing a file surfaces memories about structurally related files. Pure-Go (no tree-sitter / CGO).

Codex support

  • Codex plugin (.codex-plugin/plugin.json, hooks, marketplace, codex-hook / codex-watch binaries) with the official Codex hook format (command + commandWindows) and MCP config (mcp_servers).
  • Full feature parity with Claude Code — the only difference is the background model.

Cheap GPT-5 background (Codex's "Haiku")

  • New OpenAI provider tuned for GPT-5 reasoning models (max_completion_tokens, no temperature, reasoning_effort). Default gpt-5-mini; gpt-5-nano for the cheapest tier.
  • Codex ChatGPT-OAuth: reuses an existing codex login (no API key) — reads ~/.codex/auth.json, refreshes via auth.openai.com, calls the ChatGPT backend Responses API. Or use OPENAI_API_KEY. Auto-detected; zero-config.
  • Provider chain: anthropic → codex-oauth → openai → openrouter → llamacpp.

Compatibility

  • Claude Code / Anthropic flow is unchanged and takes priority when an Anthropic key is present; the new providers only activate when their credentials exist. Both work in the same install. All migrations idempotent; per-repo isolation preserved.

Note: the Codex ChatGPT-OAuth path is validated against mocks (SSE/headers/refresh); confirm against a live codex login on first use.