kineti

Kineti 0.6.1 native gate candidate

Record checks, detect changed inputs, and block participating CI commands when evidence or reported spending fails policy.

0.6.1 is an unreleased repair candidate targeting Linux and macOS, x64 and arm64. It needs fresh acceptance of its exact source, binaries and installed packages, followed by separate artifact and accounting approval for publication. kineti-cli changes from its older assistant lineage to this gate. Older published releases remain available. Harness execution, model routing, delegation, swarm, dashboard, memory, evaluator services and protected writes are removed from the active product.

Install and initialize

After approved 0.6.1 publication, choose one installation route:

Cargo (requires Rust for installation):

cargo install kineti-cli --version '=0.6.1' --locked

npm (requires Node and enabled installation scripts for installation):

npm install -g [email protected]

The already published 0.6.0 packages remain available: Cargo 0.6.0 and npm 0.6.0. See the GitHub release and release closeout for dated 0.6.0 publication and acceptance evidence. That approval and those artifacts do not establish 0.6.1 acceptance.

From your project directory, preview initialization with your explicit goal:

kineti init --goal "Your explicit project goal"

Review the preview, then apply initialization:

kineti init --goal "Your explicit project goal" --apply

The installed executable needs neither Node nor Bun. The npm package contains four checksummed binaries; installation verifies/selects one and removes unused payloads. Disabled scripts leave an explicit installation error.

From source, use cargo run --manifest-path core-native/Cargo.toml -p kineti-cli --bin kineti --.

init previews changes, destinations and backups. Review before --apply. New projects require an explicit goal; existing locked goals and historical records stay intact. Missing fields may be initialized; damaged records require separate reviewed recovery. When Codex is detected, init proposes native MCP registration and one instruction skill. --codex-home <directory> chooses an explicit destination. Repetition preserves unrelated settings; edited owned settings are refused. init --uninstall --apply removes owned integration content and preserves project records. Restart existing Codex sessions after changes. Cursor and Antigravity configuration is preserved, but those hosts are unverified for this candidate. There is no automatic native-tool interception. Dated 0.6.0 Codex integration checks cover only 0.163.0-alpha.11.2 and stable 0.164.0; no intervening version is implied supported. Fresh 0.6.1 host checks must bind the new executable and exact Codex version.

Record and check

kineti test run --label tests --required-test tests -- cargo test
kineti test check --label tests --expect-argv '["cargo","test"]'
kineti ci --require-evidence tests
kineti spend log --stage build --model reported --tokens-in 0 --tokens-out 0 --usd 0.25
kineti spend check
kineti undo push --label cleanup -- rm generated-file
kineti undo rollback
kineti undo rollback --yes
kineti self-test

Configure required labels in .kineti/gate.json or pass them explicitly. Empty required checks cannot pass. --evidence-commands <json-file> binds each label to an exact argv array. --env-name NAME declares relevant environment inputs; PATH is bound by default. Receipts bind project inputs, dependency files, runtime identity, exact arguments, declared environment, session reference and observed results. Commands provide structured JSON; explicit --json is also accepted. evidence and saga remain aliases. Removed commands fail explicitly.

verify runs only its declared command after command-hash trust. Use verify --status and interactive verify --trust. It does not intercept other commands. CI checks evidence, project gates and reported spending policy.

Trust and recovery

A valid signature does not prove honest tests. A command can fabricate a passing summary; self-test demonstrates this limitation. Local keys do not protect against a same-account adversary. Native receipts use kineti-canonical-v2; legacy signatures use the original JavaScript serialization. Unmarked historical v3 receipts also retain that original signing format. Legacy evidence is readable and signature-checkable; unsupported legacy runtime/parser freshness cannot authorize the native gate. Run fresh native checks after migration.

Spending is reported accounting, not provider-bill verification. Integer millionths of USD are used, with a 95% total/stage stop. A lower configured safety factor can stop spending earlier; configuration cannot raise the 95% stop. Corruption, unknown costs and unresolved reservations fail closed. Reset requires audited --i-am-human confirmation; the flag does not prove a human acted and cannot create headroom.

spend recover previews recovery. Select a preserved snapshot inside the project with --source and --evidence. Apply requires the exact reviewed --proposal-sha256, --apply and --i-am-human. Originals remain unchanged and are backed up; a separate recovery view and resolution trail are recorded. Unknown records are never cleared. Native chain damage requires inspection. Incomplete recovery, altered recovery history, and changes to a reviewed recovery view block accounting until inspected.

Undo records intent/completion, runs newest first and continues after definite failures. Uncertain interruption blocks replay until inspected and resolved with undo resolve --action-id ID --outcome applied|not-applied --yes. If the history guard was interrupted, undo inspect previews its exact bytes; undo inspect --expected-sha256 HASH --yes records reviewed inspection and keeps a byte backup. This does not resolve uncertain action outcomes. There is no atomic authorization of later writes or protected-write guarantee.

Acceptance

The frozen 0.6.0 contract, behavior map and exit codes, and dated evidence matrix preserve the behavior and accounting baseline. The owner separately selected 0.6.1 preparation; its exact artifacts need fresh acceptance and separate publication approval. Final installed files for each route must be below 10,000,000 bytes per target, including initial integration assets. Download and unpacked-package sizes are separate. No unmeasured platform-size claim is made.

Exit codes: 0 success; 1 failed command/check or incomplete undo; 2 invalid input; 3 spending/integrity/recovery stop; evidence 4 stale, 5 missing, 6 unsupported legacy, 7 invalid/untrusted; verify 9 missing/untrusted command.

Dated older results, including the historical 631-test result, are separate from the new suite inventory. Historical documents and Git history remain available; they are not current product guarantees.

Operational evidence check results do not authorize test gates. Test checks and CI require test receipts. Declared label commands override default test commands; operational checks never inherit the default test command. Value options accept --name=value; arguments after -- remain exact, including empty arguments.