docs(maintainer): blueprint completeness and strategy review - #1766
Conversation
A deliberately critical pass over what the product actually is today, from four angles the user asked about: someone who does not know AI, someone using an agent for the first time, multi-agent/collaboration, and commercialisation. Every claim carries a file path or a number that can be re-derived; anything unverifiable is marked as such. Three findings worth the read on their own: - The trust layer is declarative, not verifiable. 440 lessons: 398 carry `provenance`, but 96% of those values are uncheckable labels (`community` / `external`); only 22% carry `evidence_level` and 0.7% carry `evidence_refs`. Meanwhile CI is green on a PR whose `source:` points at a repository that does not exist (#1713) — structure, DCO and injection scanning cannot check truth. Today the trust bottleneck is maintainer attention, not machinery, and that does not scale to the corpus size the flywheel needs. - "Multi-agent collaboration" is not what exists yet. What exists is one knowledge base read by several agents: no task layer, no handoff, no conflict detection, and `agent_type`/`client_id` are self-declared (AGENTS.md §3.3 says so). The missing piece is a lease/receipt layer, which could sit on the existing events table. - Both candidate labels score 0.5/1 today. "Improves efficiency" has cases but no measurement; "useful to large organisations" has potential but no deliverable (RFC-280 Vision 4 mentions enterprise deployment, and there is no PRD behind it). Also records the two cheap moves that would move the labels to 1: quantify hit-vs-missed tasks from data already in `me_events`, and write two pages of self-hosting notes (data boundary / audit / quota) to test whether B-class demand exists. Signed-off-by: Ikalus1988 <136884451+Ikalus1988@users.noreply.github.com>
PR Reviewer Guide 🔍Here are some key observations to aid the review process:
|
PR Code Suggestions ✨No code suggestions found for the PR. |
🧾 Audit Report — PR #1766 (d5622fe)📊 Quality Score🔏 DCO Audit✅ All commits signed-off. 📏 PR Size
🔐 Secret Scan✅ No hardcoded secrets detected. 📦 Dependency Audit⏭️ Skipped; no Python/JS dependency files changed. 🧪 Test Suite✅ PASS — 53% coverage 📋 Lesson Schema✅ All lessons valid. ⚖️ Verdict✅ All gates passed. Ready for merge. Scope: |
🎉 Merged — Thank you!Your contribution has been merged into main. PR: #1766 — docs(maintainer): blueprint completeness and strategy review What's next:
Welcome to the MisakaNet contributor community! 🧠 |
|
✅ Merged! Thanks again, @Ikalus1988. docs(maintainer): blueprint completeness and strategy review (+184 lines, 1 files) Quick question — did any MisakaNet lesson help you this time? No need to reply if nothing comes to mind. ⚡ |
User description
A critical pass over what the product is today, from the four angles asked for: someone who does not know AI, a first-time agent user, multi-agent/collaboration, and commercialisation. Every claim carries a path or a number; unverifiable items are marked.
The three findings that matter
1. The trust layer is declarative, not verifiable. Audited all 440 lessons:
provenanceevidence_levelevidence_refsprovenance.sourcevaluescommunity289 /external80 /internal14 / … → 96% uncheckable labelsAnd the gates cannot see the difference: #1713 was 24/24 green while its
source:pointed atmodelcontextprotocol/mcp-memory-service, a repository that returns 404. CI checks structure, DCO and injection — not truth. So the bottleneck for "verifiable failure memory" is currently maintainer attention, which does not scale to the corpus size the flywheel needs. The cheapest structural reinforcement is a resolvability check onsource:URLs (~10 lines + one CI step); it will not catch semantic falsehood, but it removes the cheapest and most numerous class of pollution.2. "Multi-agent collaboration" is not what exists yet. What exists is one knowledge base read by several agents. No task layer, no handoff, no conflict detection, and identity is self-declared (
AGENTS.md§3.3 saysagent_type/client_idare not attribution evidence). The missing piece is a lease/receipt layer, and it could sit on the events table that already exists.3. Both candidate labels score 0.5/1. "Improves efficiency" has cases but no measurement (no A/B, no hit-vs-missed comparison); "useful to large organisations" has potential but no deliverable —
docs/rfc-280-90-day-roadmap.mdVision 4 mentions enterprise/federated deployment with no PRD behind it. The two cheap moves that would take both to 1: quantify hit-vs-missed tasks from datame_eventsalready records, and two pages of self-hosting notes (data boundary / audit / quota) to test whether that demand is real.The document also carries the 5-layer puzzle inventory, per-segment breakpoints for the non-technical user (the first-question example is English error text while the same corpus answers natural-language Chinese questions with zero hits), the flywheel diagram with its two leaks, and ten weakest assumptions each with a falsification test.
Docs only — no code. Says plainly which numbers are unverified (
cloud.google.comlink unreachable from this machine).PR Type
Documentation
Description
Adds critical product blueprint review across four user angles
Audits trust layer: 440 lessons, 96% uncheckable sources
Identifies three hard gaps in current product state
Recommends six next steps ordered by leverage ÷ cost
Diagram Walkthrough
File Walkthrough
blueprint-and-strategy-review-2026-09-16.md
Blueprint completeness and strategy reviewdocs/maintainer/blueprint-and-strategy-review-2026-09-16.md
first-time agent, multi-agent, enterprise)
evidence_level,0.7% carry
evidence_refs, 96% ofsource:values are uncheckable labelssucceed and where they fail
Chinese/natural-language search, first-call telemetry, self-hosting
notes, minimal task lease layer, GUI last)