Skip to content

sdlc-process 0.4.13 — observer reporting integrity

Choose a tag to compare

@cchatfield cchatfield released this 08 Oct 22:48
adf515d

The MAIN observer now explicitly checks substantive events against successful same-session SDLC widget updates and inspection before the next normal reply or dependent action. It compares success claims with visible results, retains unresolved findings, distinguishes apologies from corrections, and checks loaded-source freshness and concurrent writer conflicts. Unchanged turns, pending tools and honest capability gaps do not create heartbeat obligations.

Session-start/setup distinguish installed/discovered plugins from successful native loading of the actual source. The bundle pins Planning bfa23a05205f98a3bd9ba9d3d57f6bae89afdcd6 (planning#41), which preserves build authority while separating ticket publication and readiness and extends existing shared-worktree coordination. Version: sdlc-process 0.4.13. Core engineering and status runtime versions are unchanged.

Validation:

  • 88 Node tests, 9 Python tests and the 74-file generator drift check pass. Fifteen new replay fixtures cover failures and compliant/partial/unknown controls.
  • Haiku 5.5 replay uses exact saved instruction bytes. Following independent diagnosis of private-assessment trigger/recovery ambiguity, one pinned four-case trial (f15/f3/f4/f21) recovered native finding identities and emitted the expected private assessments without claiming correction; unchanged pending work stayed silent. Earlier trials include missed annotations and a false-pass finding with an incorrect reference. Those failures remain recorded as variance; this is not a universal reliability claim. Several baseline instructions also detected the test contradictions, so the release does not claim every positive case was previously missed.
  • A real Claude print MAIN session uses native Skill invocation and full source/reference reads against the generated candidate. This is loading evidence, not complete readiness, interactive first-turn coverage or live MAIN observer-delivery qualification. The runtime delivery mechanism is unchanged.

Native transcript and replay logs are disposable. Maintained cases and their acceptance criteria live in sdlc-process/tests/observer-behavioral-prompts.md. User authorized these focused corrections and retained PR/merge/release/install authority. No live session157b intervention or environment changes.

Commands (run against the pinned disposable candidate):

node --test tests/*.test.mjs sdlc-process/tests/*.test.mjs
python3 -m unittest discover -s tests -p 'process*.py' -v
python3 tools/build-process.py --planning /projects/sidkik/ep/planning-wt-authority-continuity --core /projects/sidkik/ep/core --check
claude --print --plugin-dir /tmp/reporting-integrity-evidence/merged-correction/subject/sdlc-process --plugin-dir /tmp/reporting-integrity-evidence/merged-correction/subject/sdlc-status --model claude-haiku-5-5 --tools Skill,Read --allowedTools Skill,Read --permission-mode dontAsk --setting-sources '' --strict-mcp-config --mcp-config '{"mcpServers":{}}' --output-format stream-json --verbose

The native command ran from /tmp/reporting-integrity-evidence/merged-correction/native/work with its disposable CLAUDE_CONFIG_DIR. Its stdin explicitly requested native loads/full references only and prohibited writes, setup repair, delegation and observer launch. It was a bounded loading probe, not a blind autonomy test.