Collapse Phases 4 (summary commit) + 5 (brain note) into a single
capture-driven Phase 4. The skill now assembles one brain:capture payload
(insights + tickets + summary + context) and dry-runs-then-executes;
capture owns the brain/gitea/ai-sessions writes, the I1 gate, the I5
audit record, and the supersession/read-after-write discipline server-side.
Adds the classification-gate lesson learned dogfooding capture this
session: effective classification = strictest across ALL targets
(wings, ticket repos, repos_touched); untagged repos (brain, ai-sessions)
fail-safe to confidential and get refused via claude.ai, so repos_touched
must stay to internal-default repos. Legacy inline path kept only as the
capture-unreachable fallback. Staleness prose removed (server owns it now).
Partial of #62; harness-token + harvest-adapter + fidelity-supersession
work remain.
Reconsidered Q4: instead of one global degrade-and-warn, the posture now
inherits from effective classification (Q1):
- confidential + audit-sink-down -> hard-refuse (no buffer; removes the
buffer-integrity question for confidential data)
- internal/public + audit-sink-down -> degrade-and-warn + durable local
buffer + ntfy + reconcile-on-recovery
- floor (all tiers): refuse if nothing can record the audit
Updated the I5 Gherkin scenarios + obligations row to match. Couples Q4
to the Q1 classification spine -> one coherent sensitivity model.
Q1 classification trust: model (C) — caller declares, server cross-checks
target tag, stricter wins, mismatch logged. Needs a classification
taxonomy + per-wing/repo tags (prerequisite, sub-task of #49).
Q2 harness origin: server-derived from authenticated principal;
context.harness is descriptive-only, never a gate input.
Q3 central relay: ships in v1 (needed for claude.ai/Crush/Pi/LLM Council)
+ I2 security-baseline ledger entry is v1 work.
Q4 audit-sink-down: degrade-and-warn + durable local buffer + reconcile
on recovery; refuse only if NOTHING can record the audit.
Updated the I1 sovereignty + I5 auditability Gherkin scenarios to match;
added classification-mismatch and origin-spoofing scenarios.
Pre-build spec for the uniform cross-harness capture path. Grounds the
design in the homelab invariants (I1-I5, now canonical on infra main):
- I1 sovereignty gate: refuse confidential capture via us-nexus harness
- I2: distributed-library form (no high-degree observer node) is
admissible; central relay needs a security-baseline ledger entry
- I5: capture is a privileged write path, must emit audit records
Gherkin scenarios cover the happy path, the sovereignty refusal,
supersession + staleness discipline (#45/#47), fail-closed validation,
best-effort partial-failure receipts, dry-run, and fidelity supersession.
Four open questions flagged for review (classification trust is the
highest-risk one).
Disciplined end-of-session closeout for Claude.ai chats: harvest →
ground-truth gitea → confirm issue actions → commit session summary →
brain orientation note → safe-to-archive verdict.
Finalized against current infra (git.d-ma.be, koala:30401 LiteLLM,
Authentik) and the brain_update/brain_get verbs. Phase 5 uses
supersede-by-slug + brain_get read-after-write with the batch
no-semantic-query-after-supersede discipline. session_log/brain_tunnel
inlined (now callable from claude.ai). Summary frontmatter is the
reduced live-capture schema with fidelity:live-capture to distinguish
from batch-export summaries.
Drop the context:sync:cursor task and remove .cursorrules +
.aider.conventions.md from the drift-guard file list in `check`, now
that the generator no longer emits them.
Cursor and Aider are not supported harnesses. Remove generate_cursor
and generate_aider, their no-arg calls, and their case arms. The active
adapters are claude (CLAUDE.md), agents (AGENTS.md), and system-prompt
(.context/system-prompt.txt).
mathias