Skip to content

v3.8.0 — kj start + inbound secret redaction

Choose a tag to compare

View all tags
@manufosela manufosela released this 11 Jul 15:28
v3.8.0
61cbd15
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

[3.8.0] - 2026-07-07

Minor. kj start — one entry point to the autonomous squad: the user states an intent (and at most the project's maturity) and the Brain does the rest, read-only, before proposing anything.

Added

  • kj start [intent] — the single door to the Brain (epic KJC-PCS-0061). The user never learns or invokes doctor/check/harden/onboard/rag/qmd; the Brain orchestrates them. A cheap AI layer decides the WHAT (a structured intent from a closed set) and KJ maps it to an existing saved command with confirmation for anything that writes — the model never runs commands on its own.
    • Maturity classifier (src/start/maturity.js, KJC-TSK-0568): infers new | existing | legacy from deterministic signals (scaffolding only, real code + tests + CI, or neglected health) to shape what gets proposed; the guess is confirmed with the user, never asked cold.
    • Read-only sweep (src/start/sweep.js, KJC-TSK-0568): runs the assessment commands without touching the user's code (init only ever writes .karajan/).
    • Deterministic synthesis + decider role (src/start/assessment.js, src/start/start-decider-role.js, KJC-TSK-0569): condenses the sweep into a stable report and asks the decider for the next intent.
    • Orchestration loop + dispatch (src/commands/start.js, KJC-TSK-0570): each writing intent maps to a saved command + args; degrades gracefully so kj start never crashes.
  • kj advanced command namespace (KJC-TSK-0582): kj --help is trimmed to the core commands, with the full surface indexed under kj advanced and guarded by a help-parity gate.

Security

  • Inbound secret redaction at the review boundary (KJC-TSK-0583): a hardcoded credential in the working tree used to reach the (possibly cloud) reviewer model verbatim inside the diff. src/guards/secret-redactor.js now masks every secret before the model sees it, reusing the existing CREDENTIAL_PATTERNS catalog (deterministic, no AI cost, no-op on clean text). It preserves an identifiable prefix (AKIA***REDACTED***, sk-ant-***REDACTED***) so the reviewer can still flag "move this to .env" without seeing the value. Closes the inbound counterpart to the existing outbound (commit) secret guard.
Assets 8
Loading