v1.44.0

Highlights

This release ships three coordinated features that cut the per-session prompt-injection footprint by ~1,700-2,600 tokens per Claude Code session while making agent reviews converge instead of churning.

Review ledger contract — reviews that converge (#1029)

The 4R review lenses and judgment-day used to churn: every pass re-sampled the diff with fresh context, so each round "found" issues that existed all along. Now every review agent follows a persisted-ledger contract, replicated across all 13 adapter orchestrator variants and both execution modes:

• Exhaustive first pass — each lens sweeps until N consecutive dry passes (default 2, ceiling 4).
• Persisted findings ledger ({LENS}-{NNN}, severity, status) in the session's artifact store.
• Scoped re-review — verify ledger findings and fix-touched lines only; new findings on untouched lines log as quality signals and never reopen the loop.

Battle-tested during its own development: five judgment-day cycles ran on this contract across two repos before release, converging in 2-4 rounds every time.

Engram protocol dedup — one canonical source, slim where safe (#1030)

The Engram memory protocol reached Claude Code sessions three times over (system-prompt section, SessionStart hook, MCP server instructions). Now:

• One canonical marker-sectioned asset replaces three drifting copies; Codex's three surfaces render from the same source.
• Claude Code's system-prompt section slims to 11 lines (~1,045 tokens/session saved), gated on the installed engram binary being >= v1.4.0 — the first release whose MCP server serves the instructions channel. Older, pre-release, or unparseable versions keep the full text.
• engram setup gains a side-effect-free --protocol capability probe (5s deadline, stdin detached). Paired with engram v1.19.0's new --protocol=slim|full flag, the SessionStart hook can drop another ~900 tokens of duplicated prose.

Persona canonical channel — tone lives in the output style (#1032)

Claude Code and Kimi received the persona twice (system-prompt section + output style), and the two copies had drifted into different wordings. Now:

• The output style is the single canonical tone channel; the system-prompt persona section slims to an action/tooling residual (~727/~515 net tokens saved per Claude session, ~487/~330 for Kimi).
• The drifted wordings reconciled as a union — nothing lost, enforced by a migration test with fixtures frozen verbatim from the pre-change assets.
• The 14 other adapters and the shared generic neutral asset are untouched.

Also included

• fix(codex): generate compatible permissions profile
• fix(ci): align Claude Context7 E2E expectations
• docs(windows): recommend PowerShell installer

Install / upgrade

brew upgrade gentle-ai

All three features carry their full SDD artifact sets and adversarial-review ledgers under openspec/changes/archive/2026-07-08-*/ — 54 review findings were found, fixed, and verified across the three changes before merge.