Skip to content

v0.7.2

Choose a tag to compare

View all tags
@mrilikecoding mrilikecoding released this 11 Apr 07:11
· 72 commits to main since this release
v0.7.2
57b1a31

Cycle 10 WP-F verification-surfaced remediation. Nine latent defects in the v0.7.0/v0.7.1 hook infrastructure caught by the methodology running against its own corpus in enforcement mode. All fixes land in this release. Invariant 8's compound defense is now operational end-to-end on any corpus that runs the plugin.

Hook infrastructure fixes

  • All RDD hook scripts now read input from stdin, not $1. Since v0.6.0, every input-dependent hook (invariant-reminder.sh, orientation-trigger.sh, sizing-check.sh, epistemic-gate.sh, tier1-verify-dispatch.sh) has been silently non-functional in Claude Code's actual runtime because hooks.json passed input via $INPUT (expected as a shell variable) while Claude Code delivers the JSON payload via stdin. Synthetic tests masked the divergence by invoking scripts directly with $1. Only tier1-phase-manifest-check.sh worked by accident — it reads cycle-status.md from disk and falls through on empty input — and that accidental functionality was precisely what caught the gap during WP-F. hooks.json drops the "$INPUT" argument from every command; scripts read stdin when $1 is empty and stdin is not a terminal (backward compatible with direct invocation for test harnesses)
  • PostToolUse Tier 1 hook strips plugin namespace prefix before matching against TIER1_MECHANISMS. Claude Code dispatches plugin-provided subagents with the <plugin>:<name> form (e.g., rdd:susceptibility-snapshot-evaluator), but the mechanism list used bare names. Every prefixed dispatch silently skipped logging, breaking the compound check's cross-reference entirely. ${SUBAGENT_TYPE#*:} normalizes the name; the original prefixed form is preserved in the dispatch log's subagent_type field for forensic record
  • Stop hook schema conformance — replaced {"decision":"allow",...} with plain-text stdout (advisory mode, matching the convention of the project's other stdout-based hooks) and schema-valid {"decision":"block",...} (enforcement mode). Claude Code's Stop hook schema only accepts "approve" | "block" for the decision field; emitting "allow" failed JSON validation every invocation and the advisory notice never reached the agent
  • Stop hook E1 dispatch detection is cycle-sensitive. Matches on mechanism name AND substituted expected_path, mirroring the enforcement-mode compound check. Without this, prior-cycle dispatch log residue produced misleading "dispatched but no artifact" messages at the current cycle's phase boundary
  • PostToolUse hook emits JSON null (not string "null") for missing Output path lines. jq --argjson replaces jq --arg when the path is absent, so downstream queries like jq '.expected_path == null' work correctly
  • Distinguish user-tooling mechanisms from subagent dispatches. The manifest supports a mechanism_type: user-tooling field for aid-cycle-gate-reflection entries. In enforcement mode, the compound check's dispatch-log cross-reference no longer applies to user-tooling mechanisms — gate reflection notes are produced by the orchestrator in-context, not by an isolated subagent dispatch, so they have no dispatch log entry by design. The structural floor (E1/S1/S2/S3) still applies; the Susceptibility Snapshot at the same phase boundary is the complementary content-level defense

Migrate subcommand fixes

  • /rdd-conform migrate Step 6 includes docs/system-design.md in its mechanical path substitution file list. The system design document contains multiple pre-migration path references (Appendix A per-phase susceptibility snapshot briefs, Test Architecture table, Integration Contracts notes) that were previously not updated by the migration operation, leaving the file in an inconsistent mix post-migration
  • /rdd-conform migrate Step 5 rewrites internal relative links in the moved cycle-status.md. When the file moves from docs/ to docs/housekeeping/, its own ./essays/..., ./product-discovery.md, ./domain-model.md, ./system-design.md, and ./roadmap.md references need to be rewritten for the new directory depth. Order matters: ./essays/audits/ → ./audits/ must run before ./essays/ → ../essays/

Verification evidence

  • First live operational Tier 1 dispatch in project history. The build-phase susceptibility snapshot fired via the ADR-065 Skill-Structure anchor without ceremonial attention, produced a substantive artifact, was logged by the PostToolUse hook, cross-referenced by the Stop hook compound check, and cleared the build phase boundary block. ARCHITECT phase open question 1 (does the per-phase dispatch actually fire under task load?) is answered empirically: yes, and the harness layer captures it
  • First gate reflection note. The User-Tooling Layer graduated-artifact mechanism (ADR-066) produced its first real artifact at docs/housekeeping/gates/014-build-gate.md with all five required headers and four required fields, successfully satisfying the manifest check
  • Three preserved historical snapshots at docs/housekeeping/audits/ — .pre-prefix-fix.md (first live dispatch that surfaced Finding #7), .post-prefix-pre-stdin.md (second dispatch that surfaced Finding #9), and the canonical build snapshot (third, through the fully-fixed chain). The preservation pattern is itself a methodology commitment: Option B (preserve-and-re-dispatch) is the right remediation for artifact/log provenance gaps discovered at enforcement boundaries

Cycle 10 full closure

  • 57 scenarios verified across two releases (44 v0.7.0 + 13 v0.7.1). All PASS
  • Nine remediation commits during WP-F verification, each numbered in docs/housekeeping/cycle-status.md
  • Cycle 11 candidate research question surfaced: what categories of behavior are structurally unreachable by specification-conformant synthetic tests, and which warrant runtime-level testing infrastructure? A runtime smoke test that fires an actual subagent dispatch and verifies the dispatch log is populated would have caught Findings #7 and #9 before WP-F
Assets 2
Loading