Local-first 🧠memory for Hermes Agent that lives in your Obsidian vault and remembers project context. Nightly 😴 dream passes turn repeat corrections into confirmed preferences with measurable confidence. Adapters ship for Claude Code, Codex, and OpenClaw, with an MCP server for anything else.
Source pipeline integrity and operator tooling - variant audit trail
Consultant: Claude Code (claude -p), prompt in cli-output/prompt.md, raw
output in cli-output/claude.md. Three variants were requested for the wave
architecture as a whole (where shared abstractions live), not for the worth
of individual units. The fallback consultant (Codex) was not invoked because
the primary run succeeded with three parseable variants.
Variant 1: Two shared kernels, cluster-local units
- Approach: extract exactly two new shared abstractions, each riding in
with the first unit that needs it - a gitignore-style path-scope engine
(
src/core/fs/ignore.ts) introduced by the hygiene-scan unit and consumed by ingest scoping, and a diagnostics-signal model (issue class + detector + optional fixer + next-command hint) introduced bydoctor --repairand consumed by the status snapshot. Everything else lands as local changes to its existing module, in four clusters with a short dependency spine. - Trade-offs:
- Pro: shared logic gets exactly one home without any infrastructure-only commit; each abstraction ships inside a feature commit with tests.
- Pro: maximal prefix-shippability; a stalled wave leaves
mainreleasable after any unit. - Pro: byte-identical opt-out is easy; nothing changes uninvoked.
- Con: the ingest path stays four separate touch points rather than one explicit pipeline; a future wave may want that consolidation.
- Con: the diagnostics model shape is set by the repair unit and may need a small extension when the snapshot unit arrives.
- Complexity: medium
- Risk: low
Variant 2: Source-pipeline spine
- Approach: treat the five intake units as stages of one explicit
discovery pipeline abstraction (
SourcePipelinecontext with typed hooks: scope, gate, pre-extract, dispatch, reconcile); operator tooling reads a shared health-signal registry; remaining units stay peripheral. - Trade-offs:
- Pro: strongest conceptual match to the wave theme; one choke point for what enters the vault.
- Pro: reconciliation and gating become trivially testable pipeline stages.
- Con: the pipeline scaffold is an infrastructure commit the repo's
conventions discourage, and it forces refactoring
ingest.ts/batch-plan.tsbefore any user-visible unit ships; a stalled wave leaves a half-migrated pipeline. - Con: highest risk to byte-identical opt-out, since existing ingest is rerouted through new machinery even when no new flag is set.
- Con: hygiene importing from
brain/ingestinverts layering; the ignore engine must live below both anyway.
- Complexity: large
- Risk: high
Variant 3: Eleven islands, convention-only sharing
- Approach: every unit lands independently in its own existing module
with no new shared homes; ignore parsing lives inside
hygiene/and ingest imports it across layers; the snapshot re-formats six verbs' outputs itself; fixers pair with doctor checks ad hoc. - Trade-offs:
- Pro: cheapest to execute, easiest to parallelize, every commit small and independently revertable.
- Pro: near-perfect prefix-shippability with zero scaffolding.
- Con: the ignore engine under
hygiene/withingest/importing it violates one-directional layering, or gets duplicated. - Con: the snapshot re-deriving signals from six verbs duplicates presentation logic and drifts; hints get hardcoded per call site.
- Con: defers exactly the consolidation debt the post-v1.30.1 refactor paid down.
- Complexity: small
- Risk: medium
Consultant recommendation
Variant 1. "Variant 1 delivers the two abstractions this wave genuinely reuses (ignore composition, diagnostics signals with hints) in their correct one-home locations, while avoiding Variant 2's infrastructure-first commit and its threat to the byte-identical opt-out and atomic-feature-commit conventions. Unlike Variant 3, it keeps layering one-directional - the ignore engine sits below both hygiene and ingest, and hints travel with issue definitions instead of being duplicated in the snapshot. Its short dependency spine (1->2, 9->10) preserves the constraint that a partially completed wave still ships as a coherent prefix of v1.34.0." (quoted verbatim from the consultant output, arrows ASCII-normalized)
Orchestrator decision
Variant 1 is adopted without override. It is the same anchor-owned shared
abstraction pattern that succeeded in the v1.32.0 and v1.33.0 waves, applied
to this wave's shape: here the two shared pieces (path-scope engine,
diagnostics-signal model) each ship inside the first consuming feature
commit, keeping every commit release-visible and the layering
one-directional. The accepted ordering constraints (P1 before P2, O2 before
O3) are recorded in plan.md.