This release adds 2 notable features for engineering teams evaluating rollout.
Published 1mo
AI Agents & Assistants
✓ No known CVEs patched
✓ No known CVEs patched in this version
Topics
ai-agents
aider
anthropic
autonomous
ci-cd
claude
+13 more
cline
code-review
devops
gemini
github-action
github-issues
loki-mode
multi-agent
openai-codex
openapi
pull-request-review
sdlc
spec-driven-development
Summary
AI summaryOptional caveman output-token compressor integration added for Claude Code skill.
Full changelog
Added
- Optional output-token compressor integration (caveman). caveman
(https://github.com/JuliusBrussee/caveman, MIT, vendor-less pin) is a Claude
Code skill + SessionStart hook that instructs the model to compress its OUTPUT
tokens only (prose style), keeping all technical substance. Loki ACTIVATES it
on free-form generation (the main RARV development loop and parallel dev
streams) to cut output-token cost, and HARD-SUPPRESSES it on every
parsed-output trust-gate subcall so determinism is never affected: the
completion-council votes (VOTE:), the code-review verdict (^VERDICT:), the
adversarial probe, the merge-conflict resolver, and the USAGE.md regen all run
with compression OFF (CAVEMAN_DEFAULT_MODE=off). The carve-out is DEFAULT-SUPPRESS
by construction: a singleexport CAVEMAN_DEFAULT_MODE=offin
autonomy/lib/claude-flags.sh(the one module every Loki process tree sources)
makes EVERY claude subcall inherit suppression, so any parsed trust-gate call --
existing or future, in any tree (run.sh, completion-council, council-v2,
voter-agents, grill) -- is suppressed unless a site explicitly opts into
compression. Activation (the compression level) is then set per-invocation ONLY
at the known free-form generation sites (main RARV loop, parallel dev stream).
This fails safe: a missed activation site only loses token savings, never
corrupts a verdict. A tree-wide audit test asserts no parsed claude subcall is
left unsuppressed.- Claude-provider-only. On Codex / Cline / Aider, caveman is unavailable and
the run is byte-identical to before (clean degrade, honest stderr). - Default-on, opt out with
LOKI_CAVEMAN=0. Level viaLOKI_CAVEMAN_LEVEL
(defaultfull; alsolite,ultra,wenyan*). Version pin via
LOKI_CAVEMAN_VERSION(default1.9.0); vendor-less -- Loki ships no copy of
caveman and bootstraps the pinned version on demand (idempotent, cached
marker under.loki/, opt out withLOKI_CAVEMAN_AUTO_BOOTSTRAP=0). - NEVER applied to trust gates. When the legacy completion-prose match is in
use (LOKI_LEGACY_COMPLETION_MATCH=true), main-loop activation is disabled so
compression can never mangle the prose completion-promise grep. The default
completion path (theloki_complete_taskMCP tool / completion signal file)
is immune to compression and keeps caveman on. - Savings are real but bounded (output tokens only, not input/thinking). There
is no price API, so Loki discloses the savings CLASS, never a dollar figure. - Parity-locked across both routes (autonomy/lib/claude-flags.sh predicates and
loki-ts/src/providers/claude_flags.ts mirror), with a determinism / moat
carve-out proof in tests/test-caveman-flags.sh and
loki-ts/tests/providers/caveman_flags.test.ts.
- Claude-provider-only. On Codex / Cline / Aider, caveman is unavailable and
Weekly OSS security release digest.
The CVE patches and breaking changes that affected production tools this week. One email, every Sunday.
No spam, unsubscribe anytime.
Share this release
About Loki Mode
Multi-agent autonomous SDLC framework. Spec to deployed app. PRD, GitHub issue, OpenAPI/JSON/YAML, or one-line brief. 5 AI providers, 11 quality gates.
Related context
Related tools
Beta — feedback welcome: [email protected]