This release adds 3 notable features for engineering teams evaluating rollout.
Published 27d
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 summaryFix prevents rebuilding finished work when reusing an already‑completed PRD.
Full changelog
Reuse done-recognition: a no-PRD reuse run no longer rebuilds finished work
- The bug: a
loki startwith NO PRD over a project Loki already built and
completed reverse-engineered (reused) the prior generated PRD and then rebuilt
a full task queue and re-ran the RARV loop, re-doing finished work. The reuse
path never asked "is this reused PRD already satisfied by the current code?"
Observed live: a completed project (completion signal present, all tests green,
advisor said stop) still got an 11-taskprd-*queue and fresh iterations. - The fix: a single localized gate (
reuse_done_recognition_gate, in the new
autonomy/lib/done-recognition.sh) runs inrun_autonomous()between
load_stateand the queue/loop, only on a no-PRD reuse of an existing
generated PRD. It model-verifies whether the codebase already satisfies the
spec and routes to one of three outcomes:- done: re-runs the tests now, re-checks each requirement, refreshes the
verified-completion record, finishes through the normal completion path, and
tells the user clearly ("This project already satisfies its spec ... To
rebuild from scratch runloki start --fresh-prd; to extend it, edit the
spec or pass a new/changed PRD."). No wasted iterations, no stray delegate
branch (the gate runs before the start-sha/delegate block). - incomplete: writes a
prd_sha-guarded satisfied-requirements manifest so
populate_prd_queuebuilds ONLY the unsatisfied requirements (no full
replan). - inconclusive: falls through to the normal full build (safe default).
- done: re-runs the tests now, re-checks each requirement, refreshes the
- Trust moat intact (no fake-green): the positive
doneverdict is always
the model's, grounded in re-run reality, never asserted from a stale artifact.
A modeldoneis DOWNGRADED to build if the freshtest-results.jsonis red or
any requirement is unmet/uncertain. Theupdateaction (PRD stale by
definition) may never fast-stop as done; it only feeds the incremental queue.
The only deterministic shortcut is NEGATIVE (cheap signals that route to
BUILD, e.g. no completion footprint at all, or a provider that cannot verify). - Rollout: DEFAULT-ON.
LOKI_DONE_RECOGNITION=0disables the gate (legacy
reuse-then-build behavior). Trust-safe because any uncertainty falls through to
build, so the worst case is today's behavior. - Tests:
tests/test-reuse-done-recognition.sh(registered in
tests/run-all-tests.shandscripts/local-ci.sh) exercises the gate with a
stubbed model seam across done+green, done+red-downgrade, inconclusive,
incremental-only-unsatisfied, opt-out, update-never-fast-stops, the negative
fast-paths, and a static no-emoji/no-em-dash scan.
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]