feat(mcp): V10 MCP consolidation — single keeper, 21-tool surface, 2-command setup#381
Open
feat(mcp): V10 MCP consolidation — single keeper, 21-tool surface, 2-command setup#381
Conversation
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
Open
7 tasks
- Re-home + rename dev-coordination ontology TTL:
packages/mcp-server/schema/dev-paranet.ttl ->
packages/mcp-dkg/schema/dev-context-graph.ttl (paranet terminology dropped;
one stray paranetId? -> contextGraphId? in the signature-example comment).
- Port DKG_ADAPTERS plugin-loader shim from packages/mcp-server with the
cleaner signature (server, client, config); drop the DKG_SPARQL_ONLY env
gate (no backwards-compat shims for first dkg-v10 PR).
- Implement P0 parity tools toward OpenClaw adapter parity:
dkg_assertion_create / write / promote / discard / query
dkg_memory_search (re-implemented per parity-matrix v0.7 path B, ~280
LoC; trust weights WM=1.0 / SWM=1.15 / VM=1.3 cross-pointed to
packages/adapter-openclaw/src/DkgMemoryPlugin.ts:285-301 for drift
discipline; eventual dedup path is shared packages/dkg-memory).
- Add dkg mcp setup + dkg mcp serve subcommands to the umbrella CLI:
setup is idempotent and re-runnable (per-client state-aware prompts:
registered / stale / not-registered; --force / --print-only / --yes
flags; skips dkg init re-prompts if ~/.dkg/config.yaml already exists);
serve is the runtime entrypoint invoked by MCP-aware clients per the
written { "command": "dkg", "args": ["mcp", "serve"] } config block;
mcp-dkg main() now exported with argv parameter and self-exec guard so
the umbrella wrapper can import + dispatch cleanly while the standalone
dkg-mcp bin still works for clients that prefer the direct invocation.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Wave-2 prerequisite for the eventual deletion of `packages/mcp-server/`.
Updates the autoresearch adapter to consume mcp-dkg's `loadAdapters` shim
landed in wave-1 (`packages/mcp-dkg/src/adapters.ts`) instead of the
retired mcp-server's pluggable loader.
Architectural call (option Z of three considered): contained per-adapter
daemon-fetch shim. The legacy `DkgClientLike` contract had four methods
(`query`, `publish`, `createContextGraph`, `subscribe`) and mcp-dkg's
`DkgClient` only implements `query`. Bloating the canonical client with
the three missing helpers (option X) would bleed adapter-specific
endpoints into a surface every other tool consumer reads. Hoisting a
shared `DkgDaemonHelpers` module (option Y) is cleaner long-term but
out of scope for this PR. Option Z keeps mcp-dkg's `DkgClient` lean and
contains the legacy contract baggage in the one consumer that needs it.
Concretely:
* `src/types.ts` — `DkgClientLike` shrinks from four methods to one
(`query`, object-arg shape per mcp-dkg's `DkgClient.query`); a new
`DkgConfigLike` exposes the `api` + `token` fields the daemon-fetch
shim reads. Both interfaces are structural so the adapter never takes
a hard dep on `@origintrail-official/dkg-mcp`'s concrete classes.
* `src/tools.ts` — `registerTools(server, client, config)` replaces the
legacy `(server, getClient, contextGraphId?)`. Private `daemonPost` +
`daemonCreateContextGraph` + `daemonSubscribe` + `daemonPublish`
helpers POST to the same daemon mcp-dkg already talks to via
`config.api` + `config.token`. SPARQL calls now pass mcp-dkg's
object-arg `query({ sparql, contextGraphId })` shape.
* `test/tools.test.ts` — full harness rewrite. Replaces the legacy
`mock.publish.calls` / `mock.subscribe.calls` assertions with a
`globalThis.fetch` stub (`createDaemonFetchStub`) that captures URL +
method + parsed body per call. Tests now assert on the captured
daemon HTTP calls for the write path; the SPARQL read path still
asserts on the `client.query` mock with the new object-arg shape.
Scope reduction: the legacy 3rd `contextGraphId` override parameter
to `registerTools` is dropped per parity-matrix v0.5 §4.21. Each
adapter targets its own canonical CG. A documenting comment at the
bottom of the test file flags this is reversible as an `opts` bag if
a real consumer surfaces the need.
Test result: 37/37 passing (was 15/38 with 23 contract-mismatch
failures). End-to-end loader smoke verified via `DKG_ADAPTERS=autoresearch`
+ `loadAdapters()` against the built `dist/` — all six autoresearch
tools (`autoresearch_setup`, `_publish_experiment`, `_best_results`,
`_experiment_history`, `_insights`, `_query`) register cleanly through
mcp-dkg's adapter pipeline.
Refs: parity-matrix v0.5 §4.21, roadmap §3 step 3.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Per user direction (2026-04-30): agents should see provenance per
memory hit. Two new first-class fields plus an existing one promoted
to explicit emission, while the SKILL.md combined-string `layer`
contract stays untouched.
Preserves SKILL.md §6.3 combined-string layer contract:
`agent-context-wm | agent-context-swm | agent-context-vm | project-wm
| project-swm | project-vm`. The same SKILL-canonical principle that
drives the audit's `dkg_sparql` → `dkg_query` rename — SKILL.md is the
agent-facing contract, mcp-dkg follows it. parity-analyst caught the
divergence on the prior attempt (Option A) which split the combined
string into separate `contextGraphId` + 3-element-`layer` axes; Option
B keeps the combined string and ADDS the split fields alongside.
Per-hit shape:
- `contextGraphId: string` — first-class, no change from prior
(already on `Hit`)
- `layer: string` — combined SKILL form, no change from prior
(already on `Hit`)
- `trustWeight: number` — NEW: numeric weight applied to the score
(WM=1.0, SWM=1.15, VM=1.3); explicit so
consumers don't re-derive from the layer
string
- `path: string` — NEW: synthetic `dkg://{cg}/{layer}/{hash}`
mirroring the adapter shape
(`packages/adapter-openclaw/src/DkgMemoryPlugin.ts:410`)
so cross-surface tooling sees identical
identifiers
- existing `score`, `entityUri`, `snippet` unchanged
Text-mode rendering surfaces all three provenance pieces up front,
deriving the human-readable CG and tier from the combined `Hit.layer`
at render time (`tierFromCombinedLayer` helper):
### 1. [agent-context · VM · weight=1.30 · score=0.87]
`urn:dkg:memory:abc123`
<snippet>
Layer rendered uppercase via the helper (no schema change). CG
rendered as-is. Numeric weight + score show two decimal places.
The header's `searchedLayers` and per-layer `breakdown` continue
to emit the SKILL combined strings (`agent-context-wm` etc.) so
consumers parsing the rendered text see the contract verbatim.
Trust weighting and trust-tier dedup unchanged: tier determined by
the view, not the CG, so an agent-context VM ties with a project
VM. Source-of-truth pointer to
`packages/adapter-openclaw/src/DkgMemoryPlugin.ts:285-301,391-398`
preserved in the file header alongside qa-engineer's Case 3 drift-
detector note.
Test result: 101 mcp-dkg tests passing (same pre-existing
`capture-hook.test.ts` syntax-error failure, unchanged from prior
commit). Smoke-tested the rendered output end-to-end against a stub
`DkgClient` returning seeded bindings across all six layers — VM >
SWM > WM trust ranking holds, header lines emit SKILL combined
strings, per-hit lines emit the human-readable provenance triple.
Refs: roadmap §3 wave-2; SKILL.md §6.3; matches dispatch-brief text
spec verbatim.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Wave-2 rename pass per parity-analyst's tool-inventory-audit §7
items 5, 10, 11. SKILL.md (`packages/cli/skills/dkg-node/SKILL.md`)
is the source-of-truth contract for the V10 agent-facing tool surface;
mcp-dkg was the divergent surface for these three.
* `dkg_sparql` → `dkg_query` (audit §7 item 5 + SKILL.md line 175,
`POST /api/query`). Includes a TWO-AXIS SCHEMA MIGRATION beyond
the rename: the legacy `layer: 'wm' | 'swm' | 'union' | 'vm'` enum
conflated two orthogonal axes. The new shape exposes them
separately:
view: 'working-memory' | 'shared-working-memory' | 'verified-memory'
includeSharedMemory?: boolean
The `'union'` mode (`view: 'working-memory' ∪ SWM`) now requires
explicit `view: 'working-memory' + includeSharedMemory: true`. The
daemon-side wire shape already accepts both as separate fields per
`client.ts:133-183`; this is a public-tool-surface alignment only,
no daemon change needed. Surviving call sites of the legacy
`includeSharedMemory: true` path (in `dkg_get_entity` and
`dkg_get_agent`) continue to work — those tools call
`client.query` directly with the same two-axis shape.
* `dkg_list_projects` → `dkg_list_context_graphs` (audit §7 item 10
+ SKILL.md line 174, `GET /api/context-graph/list`). Description
string now opens with "List every context graph this DKG node
knows about (also called 'projects' in the DKG node UI)" so MCP
consumers see the UI/canonical equivalence on the tool surface
itself — they don't need to read SKILL.md to learn that "project"
in the UI maps to "context graph" on the wire.
* `dkg_list_subgraphs` → `dkg_sub_graph_list` (audit §7 item 11 +
SKILL.md line 175 + adapter `DkgNodePlugin.ts:2373` — both
agree on the canonical verb-noun shape).
Per the no-backwards-compat-shims rule for the first dkg-v10 PR
(memory: `feedback_no_backwards_compat`): no aliasing, no
re-export under the old names. The legacy registrations are
removed cleanly.
Two cross-reference doc-string updates: `dkg_assertion_query` and
`dkg_memory_search` had description text mentioning `dkg_sparql`
as the precision SPARQL alternative — both now say `dkg_query`.
No tests pinned the old names (verified by `grep` across
`packages/mcp-dkg/test/` + `packages/mcp-dkg/scripts/` — no
matches), so no test-side updates were needed. 101 mcp-dkg tests
still pass post-rename. Smoke-tested `dkg_query` end-to-end:
all three argument patterns (working-memory + includeSharedMemory,
verified-memory only, defaults) route through to
`client.query` correctly.
Refs: parity-analyst audit §7 items 5/10/11, audit §5 naming
table, SKILL.md §4a tool table, no-backwards-compat-shims rule.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Wave-2 drop pass per parity-analyst's tool-inventory-audit §7
items 13-19 + the user's 2026-05-04 sign-off on the sugared
coding-project tools.
Recoverable via `git show pre-v10-tool-drop:<file>`. Sugared
dev-coordination tools (decision/task/comment) may be re-introduced
after root CLAUDE.md cleanup if needed.
Tools dropped (10 total):
V9-era / no SKILL.md analog (7):
1. dkg_review_manifest — manifest-pipeline preview tool. Tool
wrapper at `tools/review.ts` deleted;
`packages/mcp-dkg/src/manifest/` library
code KEPT (consumed daemon-side by
`routes/context-graph.ts:976-1075` +
`daemon/manifest.ts:140`). Option (a)
cheap path per user 2026-05-04.
2. dkg_annotate_turn — capture-hook chat-turn annotation
3. dkg_get_ontology — V9 starter-ontology lookup; `import-
ontology.mjs` workflow not in active use
per user 2026-05-04
4. dkg_get_chat — V9 capture-hook chat-turn read
5. dkg_set_session_privacy — V9 chat-session privacy toggle
6. dkg_request_vm_publish — V9 human-gates-VM marker (superseded by
the ungated `dkg_shared_memory_publish`
landing in #16)
7. dkg_search — V9-coded predicate set; superseded by
`dkg_query` (#17 rename) and
`dkg_memory_search` (wave-1 P0)
Coding-project sugar dropped per user 2026-05-04 (root CLAUDE.md is
a coding experiment being removed in a separate task; these tied
mcp-dkg to dev-coordination terminology that doesn't earn its keep
on a generic MCP):
8. dkg_propose_decision
9. dkg_add_task
10. dkg_comment
Source-file deletions:
- packages/mcp-dkg/src/tools/writes.ts (5 tools, all dropped)
- packages/mcp-dkg/src/tools/annotations.ts (2 tools, both dropped)
- packages/mcp-dkg/src/tools/review.ts (1 tool, dropped)
- packages/mcp-dkg/docs/RECONCILIATION.md (V9 look-before-mint
protocol coupled to dropped sugared writes)
In-place edits:
- tools.ts: drops `dkg_search` + `dkg_get_chat` + the
`expandPrefixed` helper used only by `dkg_search`
- index.ts: removes `registerWriteTools` /
`registerAnnotationTools` / `registerReviewTools` imports and
invocation calls
Test-suite deletions (all four files pinned dropped surfaces):
- test/normalise-slug.test.ts — imported `normaliseSlug` from
the deleted `tools/annotations.ts`
- test/uri-helpers.test.ts — imported `toUri` from same
- test/capture-hook.test.ts — asserted on `dkg_annotate_turn`
string inside V9 ontology agent-guides
- test/starter-ontologies.test.ts — asserted on `dkg_search` +
`dkg_annotate_turn` strings inside V9 ontology agent-guides
`vitest.config.ts` adds `passWithNoTests: true` so the empty test
suite exits clean. qa-engineer's verification-plan v6.x (Task #19)
is the follow-up that wires fresh fixtures for the post-drop
surface (assertion CRUD + memory-search trust ranking + dkg_query
two-axis schema).
V9 infrastructure intentionally kept (out of wave-2 scope per
team-lead brief): `hooks/capture-chat.mjs`,
`scripts/smoke-{annotate,writes}.mjs`, `templates/ontologies/`,
`packages/mcp-dkg/src/manifest/{install,publish,fetch,schema,
templates}.ts`, the `dkg-mcp join <invite>` CLI subcommand. These
have daemon-side consumers and/or are slated for separate cleanup
PRs.
README updated:
- "agent can do" examples reflect post-PR surface (12 tools)
- "Tools at a glance" table replaced (12 rows, dropped tools gone)
- "Layer semantics" → "View semantics" (matches the
#17 dkg_query schema migration: separate `view` enum +
`includeSharedMemory: boolean` axis)
Final tool surface: 22 (pre-PR) − 10 (this drop) + 5 (wave-1
assertion quintet) + 1 (wave-1 memory_search, already counted in
the 22) = 12 tools registered. Wave-2 add commit (#16) will land
9 more, bringing the post-PR surface to 21.
Build green. Tests exit clean (no failures; empty test suite
explicitly tolerated by `passWithNoTests: true`).
Refs: parity-analyst tool-inventory-audit §7 items 13-19, user
sign-off 2026-05-04, no-backwards-compat-shims rule, recovery via
`pre-v10-tool-drop` annotated tag.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Wave-2 add pass per parity-analyst's tool-inventory-audit §7. Brings
the consolidated mcp-dkg surface to its target shape: 21 tools,
matching the audit's "Net surface after applying recommendations"
(22 pre-PR − 10 dropped + 9 added; renames are net-zero).
P0 — SKILL.md Quickstart parity (1):
- dkg_context_graph_create — wraps `POST /api/context-graph/create`.
Description includes the "(also called 'projects' in the DKG node
UI)" UX note per the team-lead direction (mirrors the same lock on
`dkg_list_context_graphs` from the rename pass). Slugifies `name`
when `id` is omitted; rejects invalid slugs at the tool boundary.
P1 — high-value canonical write surface (3):
- dkg_shared_memory_publish — wraps `POST /api/shared-memory/publish`.
Final canonical-flow step (publish SWM → VM, clear SWM). UNGATED
per matrix v0.6 / user lock 2026-04-30 — no `agent.canPublishToVm`
flag; matches the OpenClaw adapter shape exactly. Optional
`registerIfNeeded` upgrades a local-only CG to on-chain
registration before publish; defaults `clearAfter` to false on
subset publishes so unpublished roots aren't dropped from SWM.
- dkg_publish — wraps the two-call helper
`/shared-memory/{write,publish}`. "I have fresh quads, publish
them now" one-shot per SKILL.md §4a. URI vs literal auto-typing
on object terms mirrors the adapter's `handlePublish` byte-for-byte.
- dkg_assertion_import_file — wraps
`POST /api/assertion/{name}/import-file` (multipart/form-data).
Daemon's extraction pipeline turns markdown / PDF / DOCX / etc.
into RDF triples. Extension-based content-type inference covers
.md/.markdown/.pdf/.docx/.html/.htm/.txt/.csv; unmatched falls
through to application/octet-stream.
P2 — curation / health (4):
- dkg_status — `GET /api/status`. Node diagnostic.
- dkg_wallet_balances — `GET /api/wallets/balances`. Pre-publish
"do I have funds" check; per-wallet balances + chain context.
- dkg_subscribe — `POST /api/subscribe`. Required for joining
peer-shared CGs; optional `includeSharedMemory: false` to skip
SWM sync.
- dkg_sub_graph_create — `POST /api/sub-graph/create`. Strict
create (mirrors adapter); for idempotent semantics during
sugared writes, `client.ensureSubGraph` is the internal path.
P3 — debug/audit (1):
- dkg_assertion_history — `GET /api/assertion/{name}/history`.
Lifecycle descriptor (author, extraction status, promotion
state). 404 surfaces as a tool error.
DkgClient extensions (`client.ts`, +220 lines):
The 9 tools needed methods that mcp-dkg's `DkgClient` didn't have
before. Added: `getStatus`, `getWalletBalances`, `subscribe`,
`createContextGraph`, `createSubGraph`, `publishSharedMemory`,
`publishQuads`, `importAssertionFile`, `getAssertionHistory`,
`registerContextGraph`. All thin wrappers over the same daemon HTTP
routes the OpenClaw adapter's `DkgDaemonClient` calls — wire shapes
mirror byte-for-byte so cross-surface tooling stays consistent.
`importAssertionFile` is the only one that bypasses `request()` —
multipart/form-data needs `FormData` rather than `JSON.stringify`,
so it builds the form locally and shares the auth-header /
base-URL plumbing with `request()` for behavioural parity.
File layout:
- tools/setup.ts (new) — context-graph + sub-graph + subscribe
- tools/health.ts (new) — status + wallet balances
- tools/publish.ts (new) — publish + shared_memory_publish
- tools/assertions.ts (extended) — import_file + history added
alongside the wave-1 P0 quintet
- index.ts — three new register* invocations
Build green. Smoke-tested `main()` exports cleanly (arity 0). Final
tool registration count by file:
tools.ts: 6
tools/assertions.ts: 7
tools/setup.ts: 3
tools/health.ts: 2
tools/publish.ts: 2
tools/memory-search.ts: 1
Total: 21 — matches audit §7 "Net surface" target.
Tests: still 0 in mcp-dkg's package (the four tests pre-existing
before #18 all pinned dropped surfaces). qa-engineer's
verification-plan v6.x is the follow-up that wires fresh fixtures
for the post-drop + post-add surface.
Refs: parity-analyst tool-inventory-audit §7 items 1, 2, 3, 4, 6,
7, 8, 9, 12; SKILL.md §3 Step 1, §4a, §7; matrix v0.6 ungated
publish lock; user sign-off 2026-05-04.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…ocks Follow-up to #16 (e71afa0). parity-analyst's audit v1.1 specified verbatim-quote requirements on four tool description strings that my ship-from-audit-§7 commit met semantically but not literally. Fixing the divergence so an agent reading SKILL.md and the tool description sees identical reconciliation language. `dkg_context_graph_create` (tools/setup.ts) — opens with the v1.1 locked sentence "Create a context graph (called 'projects' in the DKG node UI). Returns the new CG's id, name, and on-chain registration status." Earlier shipped wording ("Create a new context graph on the DKG node (also called a 'project'...)") said the same thing in different words; the verbatim quote is the agreed lock. `dkg_list_context_graphs` (tools.ts) — opens with the v1.1 locked sentence "List all context graphs the node knows about (called 'projects' in the DKG node UI)." Earlier wording was equivalent but not byte-aligned with audit §7 item 10's quote. `dkg_publish` and `dkg_shared_memory_publish` (tools/publish.ts) — audit v1.1 required both descriptions to quote SKILL.md §4a line 182's disambiguation verbatim so agents pick the right tool without re-reading SKILL.md. Both now lead with the canonical phrasing: `dkg_publish` = "I have fresh quads, write+publish now" (two HTTP calls); `dkg_shared_memory_publish` = "canonical step-4 finalizer for 'publish existing SWM' (one HTTP call)". Per-tool details follow. Tool surface and behaviour unchanged — semantic content of every description is the same as e71afa0. This commit is purely description-string alignment to lock cross-doc consistency between the MCP tool surface and SKILL.md. Build green. Tool count and registrations unchanged (still 21). Refs: tool-inventory-audit v1.1 description-string locks, SKILL.md §4a line 182, §6 line 297. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…t v1.1 locks Follow-up to e71afa0 + 2d69cef. Diffed every shipped tool shape against parity-analyst's audit v1.1 schema-source-of-truth message ("Q1+Q2 confirms + all schema sources locked"). Two divergences vs the locks: `dkg_sub_graph_create` — shipped strict (via `client.createSubGraph` which propagated the daemon's "already exists" 409). v1.1 locked idempotent semantics: route through `client.ensureSubGraph` which swallows the duplicate-name error so a re-run is a no-op. Matches the wave-1 `createAssertion` `alreadyExists: true` pattern. Tool description updated; per-call success message ("Sub-graph 'X' ready in 'Y'.") no longer claims a fresh create. `dkg_subscribe` — shipped with `includeSharedMemory` left undefined when the caller omits, relying on the daemon's `true` default. v1.1 spec'd "Pass `{contextGraphId, includeSharedMemory: true}` by default per adapter conventions" — same effective behaviour but the tool contract should make the default explicit. Switched the zod schema to `.default(true)` so the agent sees the default in the tool description without having to read the daemon code. Side-effect cleanup: `client.createSubGraph` (the strict variant added in e71afa0's `DkgClient` extension to back the now-removed strict tool) is unused after this rewire. Per the project no-features-beyond-task rule, dropped from the client. Re-add is one diff hunk if a future strict-path tool needs it. Build green. Tool surface and behaviour-on-duplicate-name now match audit v1.1 locks. Tool count unchanged at 21. Refs: tool-inventory-audit v1.1 schema-source-of-truth Q&A on sub_graph_create idempotency + subscribe defaults. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…nalyst self-correction parity-analyst's audit v1.1 schema-source-of-truth message had recommended idempotent semantics for `dkg_sub_graph_create` via `client.ensureSubGraph`, which I aligned to in 8185998. parity-analyst's follow-up review of the post-#16 commit chain self-corrected: strict is the right call. Three independent signals all point that way: 1. SKILL.md is silent on idempotency for this tool. 2. The daemon route 409s on duplicate sub-graph name. 3. The OpenClaw adapter exposes the strict path. The earlier "idempotency parity with `dkg_assertion_create`" framing was wrong: assertion-create's `alreadyExists: true` shape is a daemon-side affordance (the route returns the existing assertion's URI, not a 409). sub-graph creation has no such daemon-side affordance — `ensureSubGraph` is a CLIENT-side wrapper that catches the 409, originally added for the wave-1 sugared writes (now dropped). It should not have leaked onto the public tool surface. Reverting: - `client.createSubGraph` re-added (was dropped in 8185998 once the tool stopped consuming it). Strict semantics; daemon's 409 propagates as `DkgHttpError`. - `dkg_sub_graph_create` tool now calls `createSubGraph` again. Comment block + tool description rewritten to reflect strict-is-honest. Tool description: "Strict create — the daemon returns an error if a sub-graph with this name already exists." - `client.ensureSubGraph` retained on the client (no current consumer after #18, but parity-analyst's review explicitly preserves it for "internal use only"; the sugared writes that used it may be re- introduced after root CLAUDE.md cleanup, per Task #18 commit body). Build green. Tool surface and 21-tool count unchanged. Behaviour divergence on duplicate-name only: the tool now propagates the daemon's 409 as a tool error rather than swallowing it. Net commit chain on `chore/v10-mcp-consolidation`: - 8185998 had `dkg_sub_graph_create` idempotent (this commit reverts that aspect) - 8185998's `dkg_subscribe` explicit-true-default change STAYS (parity-analyst approved that one) Refs: parity-analyst e71afa0-review schema-shape self-correction on 2026-05-04. Audit v1.1 sub-graph-create disposition will reflect the self-correction in matrix v0.8 (Task #20 Part 2): "strict per adapter and daemon, idempotency lives in `client.ensureSubGraph` for internal use only." Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… lock) Final lock per parity-analyst matrix v0.8 §4.18 (2026-05-04). Reverts 2a80394 which itself reverted 8185998 — three flips in this thread, matrix v0.8 §4.18 is the locked disposition that closes it. Final answer: idempotent via `client.ensureSubGraph`. The create-family-wide view is what makes it the right call: - dkg_assertion_create → daemon-idempotent (`alreadyExists: true`) - dkg_context_graph_create → daemon-idempotent (returns existing CG) - dkg_sub_graph_create → daemon-strict (409 on duplicate) parity-analyst's source-of-truth verification against `packages/cli/src/daemon/routes/{context-graph,assertion}.ts` confirms two of the three create-family tools are daemon-side idempotent. The strict odd-one-out is `dkg_sub_graph_create`. Wrapping it at the client level via `ensureSubGraph` gives agents uniform "all *_create tools are safe to retry" semantics — better UX than asymmetric behaviour. Adapter parity loses to UX consistency on this one; matrix v0.8 §4.18 documents the deliberate divergence. This commit: - rewires `dkg_sub_graph_create` tool to call `client.ensureSubGraph` - drops `client.createSubGraph` (unused after the rewire; same drop pattern as 8185998) - rewrites the in-tool comment block to record the create-family rationale + matrix v0.8 §4.18 lock for future readers - tool description: "Idempotent — a pre-existing sub-graph with the same name is silently reused, no error." History note: leaving the four flip commits (e71afa0 strict ship → 8185998 idempotent → 2a80394 strict revert → THIS idempotent re-revert) intact rather than rebasing. parity-analyst preferred the honest history, and a reviewer reading the chain forward sees the real lock → align → self-correct → re-align discovery cycle. Each commit body cross-references the prior step so the chain is self-documenting. Build green. `client.ensureSubGraph` retained on the client (unused internally after #18, but is now the public path's backing — sole consumer is the `dkg_sub_graph_create` tool). Refs: parity-analyst matrix v0.8 §4.18 lock, daemon route source verification at routes/context-graph.ts:466,598 + routes/assertion.ts:1273-1277 (two of three create-family tools daemon-idempotent), audit v1.1 + the create-family-wide rationale that supersedes the post-e71afa0a adapter-only-source-of-truth flip. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
#21 R2 — cheap-modified per team-lead's unified ruling. The five starter ontologies under `packages/mcp-dkg/templates/ontologies/` have node-ui as a live consumer (`packages/node-ui/src/ui/lib/ ontologyInstall.ts` Vite-glob-imports both `ontology.ttl` and `agent-guide.md` from each starter dir, and `CreateProjectModal` exposes them in a UI dropdown). So the directory cannot be fully deleted — node-ui glob imports would break. Two-part sweep: (1) `ontology.ttl` — KEEP as V10-relevant RDF schemas. Surgically sweep the 6 dropped-tool refs in two files: book-research/ontology.ttl line 102 — `skos:definition` on slug-normalisation: "Apply BEFORE comparing dkg_search results" → "Apply BEFORE comparing free-text recall results (`dkg_memory_search`)". coding-project/ontology.ttl: - line 50 (file-level `#` comment): "Used by `dkg_annotate_turn` to express what each chat turn was ABOUT" → rewritten to describe the predicates as written by the capture-chat hook and read by `dkg_query` / `dkg_memory_search`. - line 63 (`chat:mentions` rdfs:comment): "query dkg_search first" → "use `dkg_memory_search` (free-text recall across WM/SWM/VM)". - line 78 (`chat:proposes` rdfs:comment): "freshly-minted ... in the same dkg_annotate_turn call" → "freshly-minted ... in the same write batch". - line 302 (`UriPattern` rdfs:comment): "Read by dkg_annotate_turn at runtime" → just "Documents the URI shape the agent should produce". - line 336 (`slug-normalisation` skos:definition body): "BEFORE calling dkg_search" → "BEFORE comparing free-text recall results (`dkg_memory_search`)". (2) `agent-guide.md` — REPLACE all 5 with a 1-paragraph stub each. The agent-guides were V9 capture-hook + sugared-write walkthroughs (7-15 dropped-tool refs each in the V9-flagship files; the 3 "clean" files were still V9-protocol-shaped, just incidentally without dropped-tool name strings). Per parity-analyst's audit, applying the stub to all 5 keeps the user-facing surface (modal dropdown) consistent — agents selecting different starters at project-creation time get one consistent experience instead of "this starter has detailed V9 walkthrough; that one has a stub." Stub template (per-domain one-liner pulled from each starter's existing opening paragraph for accuracy): # <Domain> agent guide This starter ships an RDF ontology for <domain> projects — <one-line domain description>. The formal schema lives in `ontology.ttl` alongside this guide. For V10 MCP tool usage, see [packages/cli/skills/dkg-node/SKILL.md](...). The tool surface to use against this ontology: - `dkg_assertion_create` + `dkg_assertion_write` — populate (WM) - `dkg_assertion_promote` — share with peers (SWM) - `dkg_shared_memory_publish` — finalize on-chain (VM) - `dkg_query` — SPARQL read; `dkg_memory_search` — free-text recall The longer per-domain agent-guide walkthrough format will return when the V10 ontology endpoint and per-project annotation workflow stabilise. Five starters: book-research, coding-project, narrative-writing, pkm, scientific-research. All five down to ~19-20 lines each. Out of scope for this commit (handled in R3 and R4 follow-ups): - `hooks/capture-chat.mjs` (R3) - root `AGENTS.md` + `.cursor/rules/dkg-annotate.mdc` (R4) - final cross-package `git grep` verification (lands at end of R4) Verification: `grep -E '<13 stale names>' packages/mcp-dkg/templates/` returns zero matches post-sweep. node-ui's `ontologyInstall.ts:178` existence check still passes (ttl + guide both present in all 5 starter dirs). node-ui display strings (`ontologyInstall.ts:77-86` hard-coded) are untouched — the modal still shows "Coding Project ontology", "Book Research ontology", etc., per parity-analyst's R2 scope concur. Refs: parity-analyst tool-inventory-audit on starter ontologies, parity-analyst R2 scope concur, team-lead unified R2/R3/R4 ruling 2026-05-04. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…apture only #21 R3 — cheap-tighter per team-lead's unified ruling. The V9 capture- chat sub-system grew to 1201 lines because it bundled three concerns: (a) chat-turn HTTP capture — write {chat:Session, chat:Turn} triples into the project's chat sub-graph. V10-relevant; keep. (b) V9 prompt injection — sessionStart `additionalContext` summarising the annotation protocol, per-turn reminders telling agents to call `dkg_annotate_turn` / `dkg_search` / `dkg_get_ontology`. Retired with the dropped sugared writes; rip. (c) V9 sugared-write companion machinery — pending-annotation queue, mention-regex backstop, agent self-register, V9 ontology fetch. Retired with the dropped tools; rip. Per team-lead's brief, this commit keeps (a) and rips (b) + (c). Net file size: 1201 → 642 lines (-559 lines, -47%). Concretely dropped (~600 lines across helpers + handlers + dispatch): - `selfRegisterAgent` — V9 onboarding helper that wrote an `agent-self-register-<slug>` assertion on first sessionStart. - `applyPendingAnnotations` — V9 race-fix that applied annotations queued by `dkg_annotate_turn(forSession=...)` on the next turn write. The tool that queued them is gone. - `buildPerTurnReminder` — V9 prompt-string builder that injected "after your reply, call `dkg_annotate_turn` with `forSession: ...`" on every beforeSubmitPrompt event. Retired in spirit and substance. - `extractMentionedUris` + the `URN_DKG_RE` regex — defensive backstop that auto-emitted `chat:mentions` for any urn:dkg:* URI in turn text "even if the agent forgets to call dkg_annotate_turn". The thing it backstopped is gone, so the backstop is too. - `buildSessionStartContext` — built the multi-paragraph V9 additionalContext that taught agents the annotation protocol on session start. Retired wholesale. - `handleSessionStart` — only did V9 self-register + additionalContext injection. Retired (the chat:Session triple write that USED to live here moves to the lazy bootstrap path inside handleAfterAgentResponse, which already did first-turn bootstrap for resumed-thread cases). - `handleSessionEnd` — only wrote a `dcterms:modified` timestamp on session close, which is best-effort and doesn't justify a hook event. Retired. What stays (the V10-relevant chat-turn capture path): - Constants + namespaces + RDF helpers (NS / T / P / LIT / URI). - Config loader (.dkg/config.yaml walk + env overrides + tokenFile expansion). - Session state cache (`~/.cache/dkg-mcp/sessions/*.json`). - Stdin / payload parsing (`pick`, `extractText`, `extractSessionKey`, `sanitiseSlug`, `extractMeta`). - Daemon HTTP path (`postJson`, `writeTriples`, `promoteEntities`, `perTurnAssertionName`, `shouldPromote`, `ensureSubGraph`). - `handleBeforeSubmitPrompt` — stash the pending user prompt for the next turn. Unchanged. - `handleAfterAgentResponse` — write the `chat:Turn` triples, bootstrap the `chat:Session` on first turn, promote to SWM if the session isn't private. Reduced (no pending-annotation apply, no mention-regex backstop, no `additionalContext` return). Cross-reference: predicates + sub-graph names stay byte-aligned with `packages/adapter-openclaw/src/ChatTurnWriter.ts` so a turn captured via either ingestion path lands in the same RDF shape (per team-lead's "same logic" direction in the brief). `.cursor/hooks.json`: dropped `sessionStart` + `sessionEnd` hook entries since the script no longer handles them. Kept `beforeSubmitPrompt` + `afterAgentResponse`. Comment updated to record the V9 retirement so a future reader knows the asymmetry is deliberate. Verification: zero dropped-tool refs in either file (verified via grep of all 13 stale tool names: `dkg_review_manifest`, `dkg_annotate_turn`, `dkg_get_ontology`, `dkg_get_chat`, `dkg_set_session_privacy`, `dkg_request_vm_publish`, `dkg_search`, `dkg_propose_decision`, `dkg_add_task`, `dkg_comment`, `dkg_sparql`, `dkg_list_projects`, `dkg_list_subgraphs`). `node -c capture-chat.mjs` parses clean. Module import smoke-test loads the 5 retained exports (`parseDotDkgConfig`, `pick`, `extractText`, `extractSessionKey`, `sanitiseSlug`). Out of scope for this commit (handled in R4 follow-up): - root `AGENTS.md` (R4) - `.cursor/rules/dkg-annotate.mdc` (R4) - README sections describing the capture hook's now-reduced surface (R4 — re-survey after this lands) - final cross-package `git grep` verification (lands at end of R4) Refs: parity-analyst tool-inventory-audit on V9 sub-system retirement, team-lead unified R2/R3/R4 ruling 2026-05-04 (cheap-tighter on R3 with ChatTurnWriter cross-reference for converged logic). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Deletes downstream V9 artefacts that referenced the 10 dropped MCP tools (#18), the 3 renamed tools (#17), or otherwise documented defunct V9 ecosystem flows. After this commit the cross-repo grep for the 13 stale tool names produces zero hits outside three intentional migration-context comments (tools.ts, assertions.ts, vitest.config.ts skip-list) — those stay because they document the V9→V10 migration path for future readers. Deletes: - AGENTS.md (107 lines, 9/13 dropped-tool refs in instructions) - .cursor/rules/dkg-annotate.mdc (sister V9 rule file) - packages/mcp-dkg/scripts/smoke-writes.mjs (exercises 5 dropped tools end-to-end via JSON-RPC; would fail at first call) - packages/mcp-dkg/scripts/smoke-annotate.mjs (exercises dkg_get_ontology + dkg_annotate_turn; same) - scripts/import-integration-tasks.mjs (V9 phase-task seed, one-shot historical data) - scripts/send-test-chat-turn.mjs (V9 phase-4 two-machines smoke) - docs/TWO-LAPTOP-DEMO.md (V9 demo doc on feat/cursor-dkg-integration branch; defunct flow) Edits: - packages/mcp-dkg/README.md — capture-hook section pruned to the 2-event surface (sessionStart/sessionEnd dropped from both Cursor and Claude Code config blocks, mirroring .cursor/hooks.json) - packages/mcp-dkg/src/manifest/install.ts — 2 stale dkg_review_manifest mentions in JSDoc rewritten to refer to the CLI / daemon-route preview surface; buildReviewMarkdown stays live - packages/mcp-dkg/src/tools/memory-search.ts — user-facing tool description was recommending dropped dkg_search; trimmed - packages/node-ui CreateProjectModal.tsx — comment updated to reference canonical V10 write flow (assertion_create + write + promote) instead of dkg_add_task - packages/node-ui WireWorkspacePanel.tsx — user-visible modal copy rewritten the same way - scripts/import-ontology.mjs — doc-comments updated to describe dkg_query against the meta sub-graph (script body itself is still live; it's a generic seeder) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Closes the empty-test-suite gate before the CI filter flip. Without these fixtures the CI lane silently runs zero MCP tests once packages/mcp-server is removed (qa-engineer's verification-plan §0.8 block-list item). Coverage: - Assertion CRUD quintet (create, write, promote, discard, query) round-trip with @en language-tag preservation through the JSON dump. - Wave-2 P1/P3 assertion adds (import_file: filesystem reads with inferred MIME, missing-file errors; history: lifecycle JSON dump). - dkg_memory_search: 3-layer (no projectId) and 6-layer (projectId) fan-out, trust-tier dedup (VM > SWM > WM collapses lower-tier hits on the same entity URI), backend-not-ready error path, raw-peerId routing (DID-form would silently zero out WM hits), SKILL.md §6.3 6-element layer-string contract. - dkg_query schema migration: accept post-#17 (view + includeSharedMemory) shape, reject legacy `layer` enum, lock view to the canonical 3 values. - 9 wave-2 P0/P1/P2 adds: dkg_context_graph_create (slug derivation + invalid-slug rejection + canonical-naming UX note), dkg_subscribe (default + override includeSharedMemory), dkg_sub_graph_create (wrapper-idempotent), dkg_publish (URI-vs-literal auto-typing), dkg_shared_memory_publish (selection: all default, non-empty rootEntities, registerIfNeeded pre-flight), dkg_status + dkg_wallet_balances (zero-arg renders + daemon-error path). - dkg_list_context_graphs + dkg_sub_graph_list rename guards plus the SKILL.md §6 reconciliation note ("called 'projects' in the DKG node UI") in dkg_list_context_graphs. Test infrastructure: - packages/mcp-dkg/test/harness.ts — FakeServer + FakeClient in-memory stubs implementing every method mcp-dkg tools call. FakeServer.call(name, input) parses input through the registered zod schema (strict mode) so schema-rejection assertions are honest. - packages/mcp-dkg/vitest.config.ts — drop passWithNoTests; refresh the comment header to point at the new wave-3-A fixtures. 52/52 passing locally on `pnpm --filter @origintrail-official/dkg-mcp test`; `pnpm build` (tsc --build) is clean (tests live outside rootDir: src). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…r deletion Coordinated transition commit that locks the pre-deletion housekeeping in one diff so wave-4 (`packages/mcp-server/` deletion) flips cleanly. Per qa-engineer's verification-plan §0.8 risk #1, the CI filter flip must be coupled with the same change-set that flips lockfile/wiring to keep MCP test coverage on the lane (preventing the silent CI lane drop that the empty-test-suite fixture work in #23 just guarded). Five coordinated edits: 1. .github/workflows/ci.yml:404 — `--filter` target `@origintrail-official/dkg-mcp-server` → `@origintrail-official/dkg-mcp`. Lane stays green via the 52 wave-3-A fixtures landed in `7b760331`. 2. CLAUDE.md:7-18 — MCP server config block flipped to the umbrella subcommand form `{ "command": "dkg", "args": ["mcp", "serve"] }`, with surrounding prose noting that `dkg mcp serve` is the new entry point (the umbrella `dkg` CLI is on PATH after `npm install -g @origintrail-official/dkg`). 3. CLAUDE.md:183 — schema TTL pointer `packages/mcp-server/schema/ dev-paranet.ttl` → `packages/mcp-dkg/schema/dev-context-graph.ttl` (the renamed + re-homed TTL from W1#5). 4. package.json:27 — `mcp` script `node packages/mcp-server/dist/index.js` → `node packages/cli/dist/cli.js mcp serve`. Smoke-tested: `dkg mcp serve --help` exits 0 and forwards args into the `@origintrail-official/dkg-mcp` `main()` (verified via the umbrella's `passthrough` argv synthesis at `cli.ts:1790`). 5. packages/mcp-dkg/package.json:3 — version `0.1.0` → `10.0.0-rc.3`, aligning with the repo-wide RC bump on `f8b98ca3`. 6. packages/cli/test/integrations.test.ts:52 + :527 — drop `@0.1.0` pin per roadmap §9 decision 14. Published version is `0.1.0-dev.<ts>.<sha>` so a `@0.1.0` pin does not resolve; the fixture now ships `args: ['-y', '@origintrail-official/dkg-mcp']` and the assertion uses loose `.toContain` match on the args array. Comments inline-cite the roadmap decision. The brief listed line 36 alongside 52 + 527; that line is the hello-world CLI fixture's `version: '0.1.0'` field, exercised by 4 unrelated assertions (L417, L421, L450, L467) — out of scope for the mcp-dkg pin drop. Flagged via SendMessage; no edits at L36. Verification: - `pnpm --filter @origintrail-official/dkg-mcp test` → 52/52 ✓ - `pnpm --filter @origintrail-official/dkg exec vitest run integrations.test.ts` → 41/41 ✓ - `node packages/cli/dist/cli.js mcp serve --help` → exit 0 - `node packages/cli/dist/cli.js mcp serve help` (passthrough) → shows the dkg-mcp internal help dispatcher (Pre-existing unrelated failures on `slot-helpers.test.ts` — Windows EPERM on `symlink` calls without admin — predate this commit and are independent of MCP wiring.) Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…ation) Removes the legacy mcp-server package as the final step in the V10 MCP keeper consolidation. The single keeper is `packages/mcp-dkg`, which exposes the canonical 21-tool surface (6 read tools + 7 assertion-CRUD tools + 1 memory-search + 3 setup + 2 health + 2 publish) per SKILL.md §4a/§6/§7. The 9 V10-canonical tool adds, 3 renames, and 10 drops landed across waves 1-2 (#16/#17/#18); the fixture suite landed in #23 (52 tests on the keeper); the transition wiring (CI filter, CLAUDE.md, version bump, integrations.test pin drop) landed in #24. **Scope of this commit:** 1. `git rm -r packages/mcp-server/` — 10 files removed 2. `pnpm install` regenerates `pnpm-lock.yaml` (workspace count 22 → 21) 3. Repo-root `vitest.config.ts` — drops the `packages/mcp-server` project entry 4. `RELEASE_PROCESS.md` — release-version-alignment bullet flipped `packages/mcp-server/package.json` → `packages/mcp-dkg/package.json` 5. `README.md` — package list line dropped (only one MCP package now) 6. `packages/adapter-autoresearch/README.md` — quick-start prerequisites, Cursor/IDE config block, and CLI launch invocation flipped from `node packages/mcp-server/dist/index.js` to the umbrella subcommand `dkg mcp serve` (same launch path the W3-B `CLAUDE.md` flip uses) 7. `packages/mcp-dkg/src/adapters.ts` — JSDoc no longer compares the loader to a non-existent legacy package 8. `docs/onboarding/04-package-map.md` — graph node + section + table rows rewritten to point at `mcp-dkg` (with the `dkg mcp serve` UX note from CLAUDE.md), only one MCP package remains in the dependency summary table 9. `docs/experiments/openclaw-benchmark/scripts/run-exp-b.sh` — the experiment runbook's `mcpServers` JSON flipped to the umbrella CLI form 10. `docs/PHASE2_ARCHITECTURE_PLAN.md` + `docs/plans/PLAN_REALTIME_SUBSCRIPTIONS.md` — historical plan/architecture docs surgically updated. Both docs reference the legacy package as a then-current implementation site. Rather than blanket-rewriting (which would falsify the historical record), the package-name string is replaced with a "now removed; see `pre-v10-tool-drop` tag" annotation that preserves the architectural argument while satisfying the brief's "ZERO hits" requirement on `git grep`. **Verification:** - `pnpm install` → 21 workspaces, lockfile regenerated cleanly - `pnpm run build` → 19/19 turbo tasks green (was 20+ before) - `pnpm --filter @origintrail-official/dkg-adapter-autoresearch test` → 37/37 ✓ (the loader-shim consumer was repointed in W2#15 — deletion has no effect on it) - `pnpm --filter @origintrail-official/dkg-mcp test` → 52/52 ✓ - `git grep "packages/mcp-server\|@origintrail-official/dkg-mcp-server\|mcp-server-dist"` → ZERO hits across the entire repo **Recovery:** for any historical content (the dropped tool wiring, the legacy connection.ts, the dev-paranet.ttl, the wider test suite), checkout the `pre-v10-tool-drop` annotated tag. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
W3-C's `af8d9165` trimmed the legacy-loader-comparison comment
(`Compared to the legacy mcp-server loader …`) when sweeping
`packages/mcp-server` references. Per team-lead's expanded W3-B
ruling on adapter-context comments, this comment is exempt and stays
as migration context — same precedent already established for
`packages/mcp-dkg/src/tools.ts` ("Replaces the legacy
`dkg_sparql` registration"), `packages/mcp-dkg/src/tools/assertions.ts`
("than the sugared `dkg_propose_decision` / `dkg_add_task` write
tools"), and `packages/mcp-dkg/vitest.config.ts` (skip-list
explanation).
Restored shape with one path correction: the original referenced
`packages/mcp-server/src/index.ts`, which no longer exists. Anchored
the historical pointer to the `pre-v10-tool-drop` tag instead so a
future reader can recover the original loader shape without chasing
a dead path.
No build/test impact: the comment is JSDoc-only.
- `pnpm --filter @origintrail-official/dkg-mcp build` → clean
- `pnpm --filter @origintrail-official/dkg-mcp test` → 52/52 ✓
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
W3-C's `af8d9165` over-scrubbed two frozen-in-time planning docs when sweeping the literal `packages/mcp-server` string toward the "ZERO hits" goal. Per team-lead's Option B ruling on historical fidelity, planning docs that describe the architecture as it was at plan-authoring time should NOT be rewritten to reflect a later post-consolidation reality — same posture as the migration-context JSDoc comments preserved in `packages/mcp-dkg/src/tools.ts`, `assertions.ts`, and `vitest.config.ts`. Reverts to original V9-era references: - `docs/PHASE2_ARCHITECTURE_PLAN.md:115` — restores the trio in-repo-client list (`packages/mcp-server`, `packages/mcp-dkg`, `packages/node-ui`); drops the post-consolidation parenthetical - `docs/PHASE2_ARCHITECTURE_PLAN.md:121` — restores "older `mcp-server` releases" enumeration - `docs/PHASE2_ARCHITECTURE_PLAN.md:222` — restores the original `packages/mcp-server` test-coverage bullet (with the `connection.ts` route-coverage rationale) plus the `packages/mcp-dkg` follow-up bullet - `docs/plans/PLAN_REALTIME_SUBSCRIPTIONS.md:300` — restores `packages/mcp-server/src/index.ts` as the original implementation site referenced by the §3.2 plan Final cross-repo grep state: 3 deliberate residual hits across these two files, deliberately preserved as point-in-time records. Recovery for any historical content remains via the `pre-v10-tool-drop` annotated tag. No build/test impact: - `pnpm --filter @origintrail-official/dkg-mcp test` → 52/52 ✓ Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…l ruling Reverts `0567b8a1` (which restored Group B planning-doc references to their original V9-era form on a misread of team-lead's "Option B authorized" message). Team-lead's follow-up clarified that the combined approach actually shipped in `af8d9165` + `c77c277b` — line-by-line surgical scrub with `pre-v10-tool-drop` tag-pointer annotations — is the better posture across the board: - No broken pointers for future readers to chase down. - Historical context preserved (annotations explicitly say "see `pre-v10-tool-drop` tag" so the original architectural argument is recoverable). - Clean repo-wide grep — no "but ignore these residual hits" carve-outs in any verification step. - Internally coherent with how `adapters.ts:18` was restored in `c77c277b`. This commit's tree state is identical to `c77c277b`, satisfying team-lead's "wave-3 is final at HEAD c77c277" ruling. Done as a revert (not a force-push or hard reset) so qa-engineer's in-flight #26 verification work is not disrupted by branch-history rewrites. Affected files (re-scrubbed per `af8d9165`): - `docs/PHASE2_ARCHITECTURE_PLAN.md:115` — drops the trio in-repo-client list back to the post-deletion duo + parenthetical pointer - `docs/PHASE2_ARCHITECTURE_PLAN.md:121` — restores "the historical legacy MCP package (now removed; see `pre-v10-tool-drop` tag)" form - `docs/PHASE2_ARCHITECTURE_PLAN.md:222` — restores the merged `packages/mcp-dkg` test-coverage bullet with the historical fourth-client annotation - `docs/plans/PLAN_REALTIME_SUBSCRIPTIONS.md:300` — restores `packages/mcp-dkg/src/index.ts` pointer (was the now-removed legacy package at plan-authoring time; see `pre-v10-tool-drop` tag) Cross-repo grep for `packages/mcp-server | @origintrail-official/dkg-mcp-server | mcp-server-dist`: ZERO hits. No build/test impact (docs-only). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Closes the two W3-A close-out asks from qa-engineer's verification-plan v7.4 sign-off. Both are cheap regression-resistance fixtures that don't admit a regression today (the underlying invariants already hold) but guard against future accidental drift during refactors. **Follow-up 1 — Wave-2 #18 drop-sweep** (`query-schema.test.ts`): Asserts every one of the 10 dropped tool names is absent from the fully-registered tool surface. Loads every bundle (read + assertion + memory-search + setup + health + publish) so a tool re-introduced via *any* registrar trips the assertion. Source for the dropped-name list: qa-engineer's verification-plan §0.8 fixture 4 verbatim list + pre-v10-tool-drop tag. **Follow-up 2 — read-side regex-scope guard** (`assertion-lifecycle.test.ts`): Asserts the four read-side assertion tools (`write`, `promote`, `discard`, `query`) accept rich non-slug names like "Bad Name With Spaces". The bug-class qa flagged: implementer copies `/^[a-z0-9-]+$/` from `dkg_assertion_create` to all five tools because they look symmetric — but read-side tools must operate on assertion names authored by other agents (whose names don't conform to the slug shape). v0.5 §4.16 alignment paragraph. Test count: 52 → 54. - `pnpm --filter @origintrail-official/dkg-mcp test` → 54/54 ✓ Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Two future-regression tests per verification-plan v8 §0.10.7 items 1-2: 1. **Drop-sweep** — assert that none of the 10 tool names removed in c222ddc (W2-#18) reappear in tools/list output. The bug-class-most- likely is a well-meaning re-registration in a future cycle slipping past review because the surface size shifts from 21 to 22. Caught at the suite level, not the surface-probe level. 2. **Read-side regex-scope guard** — per matrix v0.5 §4.16 alignment paragraph, the /^[a-z0-9-]+\$/ regex on the assertion `name` argument is creator-side input validation only. Read-side / lookup-side tools (write/promote/discard/query/history) MUST NOT inherit it. Test passes a non-conforming name through each read-side tool and asserts the schema does NOT reject it. Plus a positive control on dkg_assertion_create rejecting the same input — the asymmetry is visible in this file alone. Single new file: packages/mcp-dkg/test/drop-sweep.test.ts (17 tests). Suite goes from 52 → 71 tests; all pass. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
qa-engineer landed `ebbdba28` (`drop-sweep.test.ts`) covering the same two W3-A close-out follow-ups I'd implemented in `e3e8eb75` — but with strictly better coverage: 1. **Drop-sweep**: qa uses `it.each` over the 10 dropped names PLUS a surface-size assertion locked at 21. The size lock is a bonus catch that fails loudly on accidental add OR remove, which my single `for`-loop implementation lacked. 2. **Regex-scope guard**: qa covers all 5 read-side tools (`write/promote/discard/query/history` — I missed `_history`) AND adds a positive-control test confirming `dkg_assertion_create` STILL rejects `"Bad Name With Spaces"`. The asymmetry is visible in a single file rather than split across two. Reverting `e3e8eb75` so the branch carries one canonical implementation of these guards (qa's), not two. Test surface goes 71 → 69, still well above the 54-baseline that the W3-A close-out targeted. No qa coordination needed — this is straightforward deduplication after parallel-contributor races; both fixtures shipped to the same branch within minutes of each other and qa's is the better one. Verified locally: - `pnpm --filter @origintrail-official/dkg-mcp test` → 69/69 ✓ Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…W5-A)
Replace the bullet-list "For AI agents" callouts with a routing table
plus dedicated MCP and OpenClaw subsections at parallel depth, followed
by a 10-step "DKG V10 as agent memory" round-trip quickstart that
reproduces the validated Phase-1 flow end-to-end.
The MCP path lands the canonical `dkg mcp setup` + `dkg mcp serve`
verbs and the actual JSON block the umbrella CLI emits today
(`{ "command": "dkg", "args": ["mcp", "serve"] }`); the OpenClaw path
gets the same six-bullet shape (install → setup → restart → 4-line
verification → flags → troubleshooting). The CLI commands cheat-sheet
gains the two new `dkg mcp` verbs alongside `dkg openclaw setup`, and
the Setup Guides table gets an MCP row.
Per audit-and-plan v1.10:
- `agent-context` is the canonical CG (lazily auto-created)
- `dkg_assertion_create + dkg_assertion_write + dkg_memory_search`
are the names that prove the round-trip
- `dkg_publish` vs `dkg_shared_memory_publish` distinction documented
in the quickstart
- env-var fallbacks (`DKG_API`/`DKG_TOKEN`/`DKG_PROJECT`/`DKG_AGENT_URI`)
surfaced in HTTP-401 troubleshooting
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Drop the absolute-path Cursor/Claude Code wiring snippets in favour of
the umbrella CLI's `dkg mcp setup` + `dkg mcp serve` verbs (writes the
canonical `{ "command": "dkg", "args": ["mcp", "serve"] }` block into
every detected client). Document the full 21-tool surface organized
into six categories (Health/identity, Discovery, Setup, Write,
Publish, Search & query) with audit-locked descriptions lifted from
each tool's `description` field.
Folds in:
- Three-command install path (`npm install -g @origintrail-official/dkg
&& dkg init && dkg start && dkg mcp setup`)
- Manual config block for environments where `dkg mcp setup` can't run
- Monorepo / contributor `pnpm exec tsx` form for from-source work
- Canonical round-trip recipe (links the repo-root quickstart)
- `dkg_publish` vs `dkg_shared_memory_publish` distinction per
SKILL.md §4a; both ship ungated to match the OpenClaw adapter
- `dkg_query` two-axis schema (`view` + `includeSharedMemory`)
- `dkg_memory_search` hit shape (`contextGraphId`, `layer`,
`trustWeight`)
- env-var fallbacks (`DKG_API`/`DKG_TOKEN`/`DKG_PROJECT`/
`DKG_AGENT_URI`) for HTTP-401 troubleshooting
- Package layout pointing at the actual six tool registration files
- Historical-recovery note for the 10 dropped V9 tools (annotated tag
`pre-v10-tool-drop`, design archive at
`agent-docs/dkg-v10-mcp-consolidation/v9-design-archive.md`)
- Schema reference renamed `dev-context-graph.ttl` (was
`dev-paranet.ttl`)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…and UX parity Per user explicit ask (#30): MCP setup must be a two-command flow matching `dkg openclaw setup` exactly. Previously it took 4 commands: npm install -g @origintrail-official/dkg dkg init dkg start dkg mcp setup Now it's two: npm install -g @origintrail-official/dkg dkg mcp setup **Implementation strategy:** zero re-implementation of init/start logic — `dkg mcp setup` re-uses the same primitives `dkg openclaw setup` does (`loadNetworkConfig`, `writeDkgConfig`, `startDaemon`, `readWalletsWithRetry`, `logManualFundingInstructions` from `@origintrail-official/dkg-adapter-openclaw`; `requestFaucetFunding` from `@origintrail-official/dkg-core`). Behaviour stays byte-aligned across the two setup verbs (network defaults, config-merge semantics, daemon-readiness probe, faucet retry/back-off, manual-curl fallback). **Step order** (each step idempotent and skippable): 1. Init `~/.dkg/config.json` if absent (falls through silently when it already exists — preserves the existing idempotent re-run posture) 2. Start the daemon via `startDaemon(effectivePort)` — no-op when a healthy daemon is already reachable on the configured port 3. Optionally fund the node's wallets via testnet faucet 4. Detect MCP-aware clients, classify state (registered / stale / not-registered), and write the canonical `{ command: "dkg", args: ["mcp", "serve"] }` block 5. Optional verification probe against `/api/status` **Flag posture** mirrors `dkg openclaw setup`: --port <n> Override API port (default 9200) --name <s> Override agent name (used only on first init) --no-start Skip daemon start (configure only) --no-fund Skip wallet funding via testnet faucet --no-verify Skip post-setup verification probe --dry-run Preview steps without writing or starting anything --force Refresh every detected client regardless of state --print-only Emit canonical JSON only; skip every other step --yes Auto-confirm registrations (default) **Files:** - `packages/adapter-openclaw/src/index.ts` — re-export `loadNetworkConfig`, `writeDkgConfig`, `startDaemon`, `readWalletsWithRetry`, `logManualFundingInstructions` from the package barrel so non-OpenClaw consumers can re-use them without deep-imports. - `packages/cli/src/mcp-setup.ts` — rewrite to the bundled flow. Existing client-detection / classification / write logic preserved (well-tested + already idempotent); init + daemon-start + faucet added in front of it. Helper-injected `McpSetupActionDeps` so the action is unit-testable without touching the real filesystem or spawning the daemon. - `packages/cli/src/cli.ts:1797+` — register the new flags on `dkg mcp setup` and dynamic-import the deps from `@origintrail-official/dkg-adapter-openclaw` + `@origintrail-official/dkg-core`. Same import surface (and same package-resolution failure mode) as `dkg openclaw setup`. - `packages/cli/test/mcp-setup.test.ts` — 9 new tests: clean-machine bundled flow, idempotent skip on existing config, --no-start, --no-fund, --dry-run, --print-only short-circuit, out-of-range --port, --port + --name override threading, faucet-failure manual-fallback. All stub the deps surface. **Verification:** - `pnpm --filter @origintrail-official/dkg build` → clean - `pnpm --filter @origintrail-official/dkg-adapter-openclaw build` → clean - `pnpm --filter @origintrail-official/dkg-mcp test` → 69/69 ✓ (no MCP regressions) - `pnpm --filter @origintrail-official/dkg exec vitest run mcp-setup.test.ts integrations.test.ts openclaw-setup-cli-args.test.ts` → 54/54 ✓ - `dkg mcp setup --help` → renders all 9 new flags - `dkg mcp setup --print-only` → emits `{ command: "dkg", args: ["mcp", "serve"] }` docs-lead now updates the README's MCP section to the 2-command flow; qa-engineer re-verifies the docs walkthrough against the new behavior. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
W6-pre (`e78d8027`) bundled `dkg init` + `dkg start` + `dkg mcp setup` into a single `dkg mcp setup` invocation. Update repo-root README to reflect the new shape: - Routing-table label: "three commands" → "two commands" - Quickstart code block: drop `dkg init` and `dkg start` lines; the bundled flow now runs all 5 steps (init → start → fund → register → verify) end-to-end. - Inline 5-step prose enumerating what `dkg mcp setup` does, plus the full step-skip flag matrix (`--no-start`, `--no-fund`, `--no-verify`, `--dry-run`, `--force`, `--print-only`). - 10-step round-trip outline collapses steps 1-3 (install / init / start) into "install" + "set up", dropping one numbered step. `dkg mcp setup` short-circuits when its work is already done, so re-runs stay safe. Auth and config still live in `~/.dkg/auth.token` and `~/.dkg/config.yaml`; the JSON block written into the client config is unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Mirrors the repo-root README rewrite in `1c4ff9ef`. The bundled `dkg mcp setup` flow from W6-pre (`e78d8027`) now runs init, daemon-start, faucet funding, MCP-client registration, and verification in a single invocation. Update the Install section to reflect this: - Two-command quickstart (`npm install -g @origintrail-official/dkg` + `dkg mcp setup`) replacing the previous four-command sequence. - Inline 5-step enumeration of what `dkg mcp setup` does in order, plus the step-skip flag matrix (`--no-start`, `--no-fund`, `--no-verify`, `--dry-run`, `--force`, `--print-only`, `--port <n>`, `--name <s>`). - Note that the bundled flow re-uses the same primitives `dkg openclaw setup` does so the two verbs stay byte-aligned on network defaults, daemon-readiness probes, faucet retry/back-off, and manual-curl fallback. The 21-tool surface, view semantics, capture-hook docs, troubleshooting matrix, and package layout sections are unchanged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…ivity (F1) qa-review-round-1 F1: the W2 #17 schema migration replaced the single-axis `layer: 'wm'|'swm'|'union'|'vm'` enum with the two-axis `view + includeSharedMemory` shape, but the migration was applied to `dkg_query` only. `dkg_get_entity` (`tools.ts:219`) and `dkg_list_activity` (`tools.ts:304`) still exposed the legacy `layer` enum on their public input schemas — even though the internal handlers already mapped to the canonical scope object. Concrete inconsistency a caller hits: an agent learns `view: 'verified-memory'` from `dkg_query`, then calls `dkg_get_entity({ uri, layer: 'verified-memory' })` — zod rejects with `Invalid enum value`. Forces a per-tool API context-switch. This commit: - Migrates both tools' public input schemas to `view + includeSharedMemory`, matching `dkg_query`. - Preserves the historical `layer: 'union'` default (WM∪SWM) when callers supply neither `view` nor `includeSharedMemory`. The internal scope-object mapping is unchanged on the wire. - Updates the "no triples found" rendering on `dkg_get_entity` to print the canonical `view=...` label instead of `layer=...`. - Adds a sweep test in `query-schema.test.ts` asserting NO public tool exposes the legacy `layer` field — same bug-class guard as the drop-sweep block in `drop-sweep.test.ts`. Plus four per-tool acceptance/rejection cases (`view: 'verified-memory'` accepted, `layer: 'union' / 'wm'` rejected) and a default-shape guard confirming the V9-era `WM∪SWM` default still routes correctly. No daemon-side change required — wire-shape was already two-axis (`DkgClient.query` accepts both `view` and `includeSharedMemory` as separate fields per `client.ts:133-183`); this is a public-tool-surface alignment only. Test count: 69 → 75. - `pnpm --filter @origintrail-official/dkg-mcp build` → clean - `pnpm --filter @origintrail-official/dkg-mcp test` → 75/75 ✓ Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…reate (F2)
qa-review-round-1 F2: `dkg_context_graph_create` wrapped
`client.createContextGraph(...)` but the client method's return type
was `{ created, uri }` with no idempotency flag — the daemon's HTTP
409 on duplicate id was caught at the tool layer as a generic
"Failed to create" error, so a caller couldn't programmatically
distinguish "duplicate" from "real failure". Forced the awkward
workaround in the tool description: "Call `dkg_list_context_graphs`
first to see if one with this name already exists" — an extra
HTTP round-trip on the agent for information the create call should
have returned.
The sister tool `dkg_assertion_create` does this right:
`assertions.ts:87-89` reads `result.alreadyExists` and emits a
distinct "already exists" message. The two tools had inconsistent
handling of the same daemon-side idempotency contract; this commit
unifies them.
This commit:
- `packages/mcp-dkg/src/client.ts` — `createContextGraph` now
catches the daemon's HTTP 409 + `already exists | duplicate |
conflict` body shape and returns `{ created, uri, alreadyExists:
true }` instead of throwing. New return type
`{ created: string; uri: string; alreadyExists: boolean }` mirrors
`createAssertion`'s pattern at `client.ts:331-354`.
- `packages/mcp-dkg/src/tools/setup.ts` — `dkg_context_graph_create`
reads `result.alreadyExists` and renders a distinct "Context
graph 'X' already exists" message vs the "Created context graph
'X'" path. Description rewritten to drop the
`dkg_list_context_graphs` workaround sentence and explicitly
document the idempotency contract.
- `packages/mcp-dkg/test/harness.ts` — `FakeClient.createContextGraph`
mirrors the real client's idempotency contract (returns
`alreadyExists: true` for duplicate ids, `false` otherwise).
- `packages/mcp-dkg/test/setup-publish-health.test.ts` — two new
tests: first-create vs second-create message asymmetry, and a
description-text guard asserting the workaround sentence is gone
+ the idempotency contract is documented.
Test count: 75 → 77.
- `pnpm --filter @origintrail-official/dkg-mcp build` → clean
- `pnpm --filter @origintrail-official/dkg-mcp test` → 77/77 ✓
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
qa-review-round-1 F6: `mcp-setup.ts:283-322` only reconciled
`effectivePort` against the persisted config inside the `else`
branch (when `writeDkgConfig` ran). When the existing-config skip-
write branch was taken (`configExists && !opts.name && !opts.port`),
`effectivePort` stayed at the CLI default 9200 even when
`~/.dkg/config.json` had a different `apiPort`.
Concrete reproducer (from qa-review-round-1.md F6): a user who
previously ran `dkg openclaw setup --port 9300` (so the persisted
config has `apiPort: 9300`) and then runs `dkg mcp setup` with no
flags would:
1. Hit the existing-config branch — print "Node config exists;
leaving untouched" — but `effectivePort` still 9200.
2. Start the daemon on 9200 instead of 9300.
3. Verification probe hits the wrong port.
4. Registered MCP entry implicitly references the wrong daemon.
Fix: lift the read-back logic into a helper
(`reconcileFromPersistedConfig`) that runs unconditionally on every
branch where the persisted config is on disk — both the skip-write
branch (existing-config) and the write-then-read branch (writeDkgConfig
just ran). The helper is also tolerant of corrupt JSON / missing
fields (the previous behaviour silently used pre-merge values; same
posture preserved).
This commit:
- `packages/cli/src/mcp-setup.ts` — extract the read-back into
`reconcileFromPersistedConfig`; call it from both branches.
- `packages/cli/test/mcp-setup.test.ts` — two new tests covering
the post-fix behaviour: (a) existing config with `apiPort: 9300`
+ no flags → `startDaemon` receives 9300, not 9200; (b) same
shape with `--no-start` runs without throwing on the read-back
branch (companion structural test).
Test count for cli/test/mcp-setup.test.ts: 9 → 11.
- `pnpm --filter @origintrail-official/dkg build` → clean
- `pnpm --filter @origintrail-official/dkg exec vitest run mcp-setup.test.ts` → 11/11 ✓
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… daemon-409 flag (F12)
qa-review-round-2 F12: `dkg_shared_memory_publish`'s
`registerIfNeeded` branch tolerated the already-registered case via
`message.includes('already registered')` — a locale-fragile
substring match against an unstructured daemon error message.
Three failure modes the substring approach exposed:
1. If the daemon's wording ever changes ("CG X is already registered
on-chain" → "X is already on-chain" / "duplicate registration"),
the substring check fails: a real already-registered case
surfaces as a hard "Failed to register context graph: ..." error
AND blocks the publish step.
2. Conversely, if a future error message contained the substring
"already registered" for an unrelated reason, the tool would
silently swallow a real failure.
3. Localisation drift in error messages (English-only assumption).
Fix: push the typed signal into the client, mirroring the F2 pattern
on `createContextGraph`. The daemon's `/api/context-graph/register`
already returns HTTP 409 on the already-registered case
(`packages/cli/src/daemon/routes/context-graph.ts:520-524`), so the
client can branch on `err.status === 409` (a typed boolean signal)
and surface `alreadyRegistered: true` as a first-class field in
the return type. The tool then branches on the typed flag — no
substring text match anywhere in the path.
This commit:
- `packages/mcp-dkg/src/client.ts` — `registerContextGraph` now
catches `DkgHttpError` with `status === 409` and returns
`{ registered, alreadyRegistered: true }` rather than throwing.
New return type `{ registered, onChainId?, txHash?, hint?,
alreadyRegistered: boolean }` mirrors `createAssertion` and
`createContextGraph` shapes.
- `packages/mcp-dkg/src/tools/publish.ts` —
`dkg_shared_memory_publish`'s registerIfNeeded branch reads
`result.alreadyRegistered` and only sets `registration` (the
success-summary record) when it was newly-registered. Non-409
failures now propagate as tool errors without the substring
filter — the swallow-on-substring failure mode is gone.
- `packages/mcp-dkg/test/harness.ts` — `FakeClient.registerContextGraph`
returns the new shape with `alreadyRegistered: false` by default;
callers can override per-test for the alreadyRegistered case.
- `packages/mcp-dkg/test/setup-publish-health.test.ts` — updated
the existing `registerIfNeeded: true` test to the new shape.
Added two F12-locked tests: typed-flag tolerance (publish proceeds,
success summary doesn't claim "Registered on-chain"), and
non-409-failure propagation (a thrown rpc-unreachable error
surfaces as a tool error).
Test count: 77 → 79.
- `pnpm --filter @origintrail-official/dkg-mcp build` → clean
- `pnpm --filter @origintrail-official/dkg-mcp test` → 79/79 ✓
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
qa-review-round-1 F3 + qa-review-round-2 F13: `dkg_publish` and
`dkg_shared_memory_publish` ship ungated (matrix v0.6 user-locked,
parity with OpenClaw). The original F3 ask was a loud pre-publish
chain-warning; user explicitly opted for option (a) WITHOUT
warning prose ("we don't do that anywhere at the moment, can add
later if needed"). This commit collapses F3 entirely into F13's
chain-provenance echo so callers can verify post-hoc which chain
the publish landed on.
The daemon's `/api/shared-memory/publish` response does NOT include
`chainId` in the JSON body — the value is threaded through the
tracker only (`packages/cli/src/daemon/routes/memory.ts:483-488`).
The cleanest path is reading from the existing
`/api/wallets/balances` endpoint (which already exposes `chainId`
as a first-class field) once the publish succeeds.
This commit:
- `packages/mcp-dkg/src/tools/publish.ts` — adds a
`resolveChainId(client)` helper that reads the daemon's
configured chainId via `getWalletBalances()` and returns `null`
on probe failure (best-effort, non-fatal). Both `dkg_publish`
and `dkg_shared_memory_publish` summaries now append a
`Chain: <chainId>` line when resolution succeeds.
- `dkg_shared_memory_publish` — also echoes the resolved
`accessPolicy` in the registration line when `registerIfNeeded`
ran (e.g. `Registered on-chain: chain:cg (accessPolicy=1)`).
Caller can verify the daemon committed the value they
requested without a separate read-back call.
- `packages/mcp-dkg/test/setup-publish-health.test.ts` — five new
F3+F13 tests: chainId echoed on `dkg_publish`, chainId echoed
on `dkg_shared_memory_publish`, graceful omission when the
wallet-balances probe fails (publish itself stands), accessPolicy
echoed on the registration line, and a
no-warning-prose guard locking the user's explicit posture (no
"spends gas" / "verify chainId" / generic "warning" copy in
either response).
Test count: 79 → 84.
- `pnpm --filter @origintrail-official/dkg-mcp build` → clean
- `pnpm --filter @origintrail-official/dkg-mcp test` → 84/84 ✓
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… (F11)
qa-review-round-2 F11 + user steer ("investigate so we are sure"):
the canonical wire form for `accessPolicy` is the numeric `0|1`
per the daemon, not the string `'open'|'private'`. Confirmed by
reading the two daemon-side handlers:
- `packages/cli/src/daemon/routes/context-graph.ts:455` —
`accessPolicy: typeof accessPolicy === 'number' ? accessPolicy : undefined`
in the `POST /api/context-graph/create` handler. Strings are
silently dropped to undefined.
- `packages/cli/src/daemon/routes/context-graph.ts:509-510` —
`if (accessPolicy !== undefined && (accessPolicy !== 0 && accessPolicy !== 1))`
in the `POST /api/context-graph/register` handler returns
HTTP 400 with the explicit error
`'"accessPolicy" must be 0 (open) or 1 (private)'`. Strings
are rejected at the wire boundary.
The mcp-dkg implementation already uses the canonical numeric
form (`packages/mcp-dkg/src/tools/publish.ts:192-197` —
`z.union([z.literal(0), z.literal(1)])`). The divergent side was
the parity-matrix v0.8 §4.10 line 291 spec doc, which had
`z.enum(['open', 'private'])` — a string-form recommendation
that would have failed at the daemon if implemented. The matrix
doc lives under `agent-docs/` (gitignored team-workspace artefact);
parity-analyst owns the corresponding spec-doc realignment in
their next matrix-version cycle.
This commit's contribution: four schema-shape lock tests in
`setup-publish-health.test.ts` so a future re-alignment to the
matrix's pre-correction recommendation can't regress silently.
- `accessPolicy: 0` accepted (open)
- `accessPolicy: 1` accepted (private)
- `accessPolicy: 'open'` rejected at the schema layer
- `accessPolicy: 2` rejected (out-of-range numeric)
The string-rejection test mirrors the daemon's strict guard at
the wire boundary so the mcp-dkg public surface fails fast on
the same shapes the daemon would 400 on.
Test count: 84 → 88.
- `pnpm --filter @origintrail-official/dkg-mcp test` → 88/88 ✓
parity-analyst hand-off: matrix v0.8 §4.10 line 291 needs to flip
from `z.enum(['open', 'private'])` to
`z.union([z.literal(0), z.literal(1)])` to match implementation
+ daemon canonical. No code change needed on their end.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…eachability (F14)
qa-review-round-2 F14: pre-fix outer guard at `mcp-setup.ts:334`
was
if (shouldFund && !dryRun && shouldStart) { ... }
else if (!shouldFund) { console.log('Skipping wallet funding (--no-fund)'); }
Two failure modes:
1. **Silent omission on `--no-start`**: when the user passed
`--no-start` with default fund=true, `shouldStart` was false →
funding skipped → BUT the `else if (!shouldFund)` branch never
fired (fund was still true). User saw NO funding-step output at
all — neither attempted nor skipped.
2. **Blocked legitimate re-run**: daemon already running from a
prior invocation, faucet was down on first run, user re-runs
to retry funding. They had to remember to NOT pass
`--no-start` for funding to attempt at all.
Fix per qa's recommendation: decouple the funding decision from
`shouldStart`. New flow:
- `--no-fund` → existing explicit-skip log (intact, no behaviour
change)
- `--dry-run` → existing dry-run log (intact, no behaviour change)
- Otherwise → probe daemon reachability at `/api/status` on
`effectivePort` (same endpoint `startDaemon` uses for readiness
+ the existing verify-step probe). If reachable, proceed with
funding regardless of which invocation started the daemon. If
not reachable, log explicit "Skipping wallet funding (daemon
not reachable on port X)" — replaces the silent omission.
This commit:
- `packages/cli/src/mcp-setup.ts:334-388` — refactor the outer
guard to the new three-branch shape (--no-fund / --dry-run /
probe + faucet). Inner faucet logic (URL resolution, wallet
read-back, retry/back-off, manual-curl fallback) is unchanged.
- `packages/cli/test/mcp-setup.test.ts` — rewrite the existing
`--no-start` test to reflect the new contract (faucet skipped
via the explicit "daemon not reachable" log line, not via the
pre-fix conflation). Add two new F14-locked tests:
- `--no-start` with daemon-already-reachable → funding
proceeds (the legitimate-re-run case the bug blocked)
- `--no-fund` + `--no-start` → explicit-skip log fires; the
unreachable log MUST NOT fire (short-circuit precedence)
Tests use `vi.spyOn(globalThis, 'fetch')` to deterministically
exercise the probe path.
Docs: no README/CLAUDE.md edits needed — the existing prose
neither documented nor implied the bug-state, so the new
behaviour matches the documented contract.
Test count for cli/test/mcp-setup.test.ts: 11 → 13.
- `pnpm --filter @origintrail-official/dkg build` → clean
- `pnpm --filter @origintrail-official/dkg exec vitest run mcp-setup.test.ts` → 13/13 ✓
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
… posture (F27) qa-review-round-5 F27: `packages/mcp-dkg/test/harness.ts:62-68` claimed to mirror the production MCP SDK but ran zod with `.strict()` mode at parse — production SDK silently DROPS unknown keys, it does not reject them. Three tests in `query-schema.test.ts` were F27-bug-class: they asserted that legacy `layer: 'union' | 'wm' | 'swm' | 'vm'` is *rejected* on tools whose schemas no longer declare it. Those tests passed in the harness (because of `.strict()`) but would not have passed against the real SDK — production would parse cleanly with the unknown key dropped, then run the handler with V10 default scope. The bug-class is "harness false confidence": tests describing a harness artefact rather than the production surface. Specifically the F1 sweep tests (post-#17 + post-F1 schema migration) at query-schema.test.ts:56-69, :155-162, :176-183 each used the `rejects.toThrow()` shape pinned to the harness's strict mode. This commit: - `packages/mcp-dkg/test/harness.ts:68` — drop the `.strict()` modifier from the inline zod object schema. Add a JSDoc note explaining the production-MCP-SDK posture so future contributors don't reintroduce strict mode by accident. - `packages/mcp-dkg/test/query-schema.test.ts` — rewrite the three F27-bug-class tests to assert production behaviour: * `dkg_query` legacy `layer: 'wm|swm|union|vm'` → silent drop at parse, handler runs with default scope (no `view`, no `includeSharedMemory`). Asserts the legacy key does NOT leak through to the wire. * `dkg_get_entity` legacy `layer: 'union'` → silent drop, the handler's no-`view` branch supplies the V9-era WM∪SWM default (`includeSharedMemory: true`) on the wire — the legacy semantic is preserved by accident-of-default, not by an explicit `union` mapping. * `dkg_list_activity` legacy `layer: 'wm'` → twin to the `dkg_get_entity` case. Crucially: `layer: 'wm'` does NOT route to a WM-only query (the explicit-strict semantic pre-F1 had); it drops entirely and uses the WM∪SWM default. This documents a real V9→V10 behaviour shift the pre-F27 harness was hiding. The other 9 `rejects.toThrow()` tests qa's audit flagged as safe — declared-field validation (regex on assertion-name slugs, enum on view, required fields) is not affected by strict mode and behaves identically against the real SDK. Untouched. Test count: 88 → 88 (3 rewrites, no count change). - `pnpm --filter @origintrail-official/dkg-mcp test` → 88/88 ✓ Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
…etch timeout (W7-2 R6)
qa-engineer's R6 polish-triage list (qa-review-round-6.md). Single
commit batching the 9 review-cycle accumulation-polish items so
the diff is reviewer-friendly. None of these are correctness
gates; they're docs/comments/defensive hardening that close the
last review-finding bucket before merge.
Items in commit order:
- **F8 + F15** (`packages/mcp-dkg/src/tools.ts:160-168`) — drop the
"to get the WM∪SWM union the legacy `dkg_sparql` exposed as
`layer: \"union\"`" historical-baggage sentence from `dkg_query`'s
description. Replaced with the cleaner forward-only phrasing
"Set `includeSharedMemory: true` alongside `view: \"working-memory\"`
to query WM ∪ SWM in one call." Same edit applied to
`packages/mcp-dkg/README.md:156` so the README's tool-table
matches the description an LLM actually sees over MCP.
qa-review-round-1 F8 + qa-review-round-2 F15.
- **F18** (`packages/mcp-dkg/src/cli/index.ts:1-22`) — JSDoc
rewrite. Drop the V9-era "Phase 8 day 2" / "Phase 8 day 3"
milestone language; describe the current operator-facing
surface (`join`, `status`, `help`) and re-route distribution
prose through the umbrella `dkg mcp serve` form (production
path) with the `dkg-mcp` direct-bin retained for monorepo dev
+ direct npm-install consumers. qa-review-round-3 F18.
- **F20** (`packages/mcp-dkg/src/sparql.ts:5`) — single-word edit:
"v9 ontology" → "V10 ontology". Drift survivor from the
W2-D R4 sweep. qa-review-round-4 F20.
- **F22** (`packages/mcp-dkg/src/sparql.ts:54-61`) — defensive
hardening: `escapeSparqlLiteral` now strips null bytes BEFORE
the other escapes. SPARQL engines (oxigraph et al.) reject
null bytes outright today, so this is no observable behaviour
change in normal use — but it guards against a future engine
swap or any truncate-at-null-byte parsing path that would
silently change query semantics. qa-review-round-4 F22.
- **F4** (`packages/mcp-dkg/src/tools/assertions.ts:196-200`) —
comment clarification on `entities ?? []`. The pre-F4 comment
implied the public API allowed the empty-array sentinel; the
rewritten comment makes it explicit that the tool blocks
empty-array from callers and that the daemon's empty-array
sentinel is hidden inside the wrapper. qa-review-round-1 F4.
- **F7** (`packages/cli/src/mcp-setup.ts:173-180`) — `classify()`
now treats both `undefined` (key absent) and `null` (key
present but explicitly nulled) as `not-registered`. Pre-F7 a
`{ dkg: null }` entry classified as `stale`, which made the
operator log claim there was a current value to refresh — the
registration outcome under `--force` was identical, but the
stderr log was confusing. qa-review-round-1 F7.
- **F23** (`packages/mcp-dkg/src/tools.ts:222-228, 320-326`) —
description-string clarification on `dkg_get_entity` +
`dkg_list_activity` `view` fields. The semantic asymmetry
between explicit `view: 'working-memory'` (strict WM unless
`includeSharedMemory: true`) and omitted `view` (default WM∪SWM)
is now explicit in both schemas: explicit selection is STRICT,
omitting `view` gets the V9-era `layer: 'union'` shape. Per
team-lead's MINOR ruling: keep the opt-in-to-strict semantics,
document it in the schema description. qa-review-round-5 F23.
- **F25** (`packages/cli/src/mcp-setup.ts:316-336`) — F6's
reconcile-from-persisted-config call lifted out of the
per-branch conditionals into a single up-front read before the
branch decision. Pre-F25 the dry-run preview used the
CLI-default `apiPort` (9200) even when the persisted config
had a different port — the preview lied. Side benefit: collapses
the two duplicate reconcile calls into one. qa-review-round-5 F25.
- **F26** (`packages/cli/src/mcp-setup.ts:393, 501`) — both
`/api/status` probes (F14's daemon-reachability probe before
funding, plus the verification probe at the end of setup)
now bound the fetch with `AbortSignal.timeout(2000)`. A
partially-up daemon (port bound but unresponsive) would have
hung setup indefinitely pre-F26. Two-line defensive fix at
each site; no behaviour change in the happy path.
qa-review-round-5 F26.
Existing test suites cover the changed surfaces — F25's
behaviour shift on dry-run preview, F26's timeout fallback
(treated as "not reachable" same as connection-refused), F7's
null-handling, F23's description-only edit, and the rest are
docs/comments. No new tests required; existing tests at 88+54
(mcp-dkg + cli) hold.
- `pnpm --filter @origintrail-official/dkg-mcp build` → clean
- `pnpm --filter @origintrail-official/dkg build` → clean
- `pnpm --filter @origintrail-official/dkg-mcp test` → 88/88 ✓
- `pnpm --filter @origintrail-official/dkg exec vitest run mcp-setup.test.ts integrations.test.ts` → 54/54 ✓
F24 (`resolveChainId` extra HTTP roundtrip) tracked as post-merge
follow-up per qa+team-lead — daemon-side improvement opportunity,
not a client fix.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The W3-B transition commit's rebase against latest main produced a conflict in package.json (main added 'check:npm-metadata' alongside the legacy 'mcp' script we were rewriting). The Edit-tool resolution landed but the conflict-marker block was inadvertently committed in the same step. This commit removes the stale markers and keeps the intended merged form: the new mcp script form + main's check:npm-metadata addition. No behaviour change beyond restoring valid JSON. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Jurij89
force-pushed
the
chore/v10-mcp-consolidation
branch
from
fc2b375 to
1190795
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
packages/mcp-dkg(@origintrail-official/dkg-mcp) is retained as the V10 keeper; legacypackages/mcp-server(the V9 dev-coordination MCP) is deleted in this PR. CI lane, rootCLAUDE.md, and rootpackage.jsonrealign to the keeper.adapter-autoresearch's loader contract repoints to@origintrail-official/dkg-mcp/adaptersso the deletion does not strand its plugin pipeline.dkg_assertion_create + dkg_assertion_write + dkg_assertion_promote + dkg_assertion_discard + dkg_assertion_query(the four-step lifecycle plus its introspection). The canonical recall path isdkg_memory_search(multi-layer fan-out withcontextGraphId+layer+trustWeightfirst-class on every hit) anddkg_query(the renamed-from-dkg_sparqltwo-axis SPARQL surface —view: working-memory|shared-working-memory|verified-memory+includeSharedMemory: boolean). The publish surface isdkg_publish(fresh-quads → SWM → VM helper) anddkg_shared_memory_publish(canonical Step-4 finalizer for the stepwise flow), both ungated to mirror the OpenClaw adapter exactly perSKILL.md §4a. Ten V9-era and coding-project tools (annotations, sugared decision/task writes, deprecateddkg_search/dkg_sparql, etc.) drop entirely; the annotated git tagpre-v10-tool-droppreserves them.dkg mcp setup+dkg mcp serveship on the umbrella CLI — true 2-command UX parity withdkg openclaw setup.setupis a bundled, idempotent flow: in order, it (1) initializes~/.dkg/config.jsonif absent, (2) starts the DKG daemon as a background process if not already running, (3) funds the node's wallets via the testnet faucet, (4) registers the MCP server with each detected client (Cursor, Claude Code) by writing a single canonical entry —{ "command": "dkg", "args": ["mcp", "serve"] }— undermcpServers.dkg, and (5) verifies daemon health. Every step short-circuits when its work is already done, so re-runs are safe. The bundled flow re-uses the same primitivesdkg openclaw setupdoes (loadNetworkConfig,writeDkgConfig,startDaemon,readWalletsWithRetry,requestFaucetFunding), keeping the two verbs byte-aligned on network defaults, daemon-readiness probes, faucet retry/back-off, and manual-curl fallback.serveis the runtime invoked by the client per the JSON block written bysetup— never run by the user. End-user install collapses to two commands:npm install -g @origintrail-official/dkg && dkg mcp setup. Repo-root README's Quick Start unifies all on-ramps (MCP, OpenClaw, ElizaOS, Hermes, standalone) into parallel-depth recipes;packages/mcp-dkg/README.mdbecomes the canonical 21-tool reference.Related work
mcp-dkgfrom late March through April 2026. The keeper verdict cites this PR's commit trail as primary evidence ofmcp-dkg's active V10 development.write → DKG graph → memory search → agent reply) that this PR'sdkg_assertion_create + dkg_assertion_write + dkg_memory_searchpath reproduces verbatim from the MCP surface. qa-engineer's regression harness ataf8d9165reran the OpenClaw round-trip against the consolidated keeper and stayed green.pre-v10-tool-dropat HEAD7a2b4090— preserves the 10 dropped tool source files for any consumer needing the V9 surface. Recovery isgit show pre-v10-tool-drop:packages/mcp-dkg/src/tools/<file>.agent-docs/dkg-v10-mcp-consolidation/v9-design-archive.md— design intent and reintroduction-pointers for the four substantive V9 patterns this PR retires (per-domain ontology starters, sugared decision/task writes, agent-emitted annotations, auto-promote chat hooks).agent-docs/dkg-v10-mcp-consolidation/— full audit / matrix / verification trail:v10-keeper-verdict.md(mcp-lead, Task Game coordinator gossip hardening (PR #29 follow-up) #1),parity-matrix.md(parity-analyst, Task Sync drift between peers due to incomplete paranet scope and weak reconciliation #2 → v0.8),docs-audit-and-plan.md(docs-lead, Task the context graph explodes even after 4-5 rounds in the game #3 → v1.10),verification-plan.md(qa-engineer, Task the consensus is hardcoded #4 → v10),tool-inventory-audit.md(audit v1.1),roadmap.md(synthesis).Files changed (by area)
The 24-commit history flows wave-1 → wave-2 → wave-3 → wave-5; the groupings below regroup by logical scope so reviewers can read the changes without committing to the chronological order.
Keeper additions — new tools and CLI subcommands
The bulk of the new functionality. Lands the P0 agent-memory surface, the publish helpers, the curation primitives, and the umbrella CLI's MCP verbs.
packages/mcp-dkg/src/tools/assertions.ts(new, +440) — the sevendkg_assertion_*tools:create,write,promote,discard,query,import_file,history. Mirrors the OpenClaw adapter'sDkgNodePlugin.tswrite surface byte-for-byte.packages/mcp-dkg/src/tools/memory-search.ts(new, +388) —dkg_memory_searchclean lift fromDkgMemorySearchManager. Trust-weighted multi-layer fan-out withcontextGraphId+layer(the combined SKILL §6.3 string) +trustWeightfirst-class on every hit (per commit0e607526).packages/mcp-dkg/src/tools/publish.ts(new, +265) —dkg_publish+dkg_shared_memory_publish. Both ungated, both calling the samePOST /api/shared-memory/{write,publish}daemon endpoints; the difference is in input shape (fresh quads vs existing SWM with optionalrootEntitiesfilter and opt-inregisterIfNeededfor on-chain CG registration).packages/mcp-dkg/src/tools/setup.ts(new, +201) —dkg_context_graph_create,dkg_subscribe,dkg_sub_graph_create. SKILL.md Quickstart Step-1 / Step-3 surfaces, plus the sub-graph staging primitive.dkg_sub_graph_createis idempotent (ensureSubGraph) — a deliberate divergence from the adapter's strict semantics, documented in matrix v0.8 §4.18.packages/mcp-dkg/src/tools/health.ts(new, +92) —dkg_status,dkg_wallet_balances. Trivial wrappers overGET /api/statusandGET /api/wallets/balances; mirror the adapter's diagnostic surfaces.packages/mcp-dkg/src/adapters.ts(new, +77) —DKG_ADAPTERSplugin-loader shim ported from the legacymcp-server(signature widened to(server, client, config)— cleaner than the legacy(server, getClient, contextGraphId?)). Required to keepadapter-autoresearchworking post-deletion.packages/cli/src/mcp-setup.ts(new, +447 — wave-1 baseline extended in W6-pree78d8027) —dkg mcp setupimplementation. Bundled, idempotent flow: (1) initializes~/.dkg/config.jsonif absent, (2) starts the DKG daemon as a background process if not already running, (3) funds the node's wallets via the testnet faucet, (4) detects Cursor / Claude Code by config-file presence and classifies each client asregistered/stale/not-registered, then writes the canonical{ command: "dkg", args: ["mcp", "serve"] }entry undermcpServers.dkg, (5) verifies daemon health. Re-uses primitives fromdkg openclaw setup(loadNetworkConfig,writeDkgConfig,startDaemon,readWalletsWithRetry,requestFaucetFunding) so the two verbs stay byte-aligned. Step-skip flags:--no-start,--no-fund,--no-verify,--dry-run,--force(stale-entry refresh),--print-only(emit JSON for manual paste). First-init overrides:--port <n>,--name <s>.packages/cli/src/cli.ts(+236) — wires updkg mcp serveanddkg mcp setupas commander.js subcommands, including all eight setup-flag registrations.dkg mcp serveimports@origintrail-official/dkg-mcpand invokesmain(['node', 'dkg-mcp', ...passthrough])so argv threads correctly through to the inner CLI dispatcher (dkg mcp serve join <invite>works end-to-end without falling through to stdio MCP).packages/adapter-openclaw/src/index.ts(+11) — barrel re-exportsloadNetworkConfig,writeDkgConfig,startDaemon,readWalletsWithRetry,logManualFundingInstructionsso thedkg mcp setupbundled flow can re-use them without reaching into the adapter's internals. Keeps the twosetupverbs byte-aligned by construction.packages/mcp-dkg/src/index.ts(+69 / -…) — registers all six tool families on boot;main()is now exported and accepts argv-as-parameter so the umbrella wrapper atcli.tscan pass through cleanly.packages/mcp-dkg/src/client.ts(+287) — addscreateAssertion,writeAssertion,promoteAssertion,discardAssertion,queryAssertion,importAssertionFile,getAssertionHistory,subscribe,createContextGraph,ensureSubGraph,publishQuads,publishSharedMemory,registerContextGraph,getStatus,getWalletBalances,getAgentIdentity. The HTTP wrapper grew alongside the tools.Keeper renames & schema migrations
Public-tool-surface alignment with
SKILL.mdand the adapter — three renames + one two-axis schema migration.dkg_sparql→dkg_query(commite0ca5e6a). The daemon endpoint isPOST /api/query; both surfaces now use the canonical name. Two-axis schema introduced in the same wave: legacylayer: 'wm'|'swm'|'union'|'vm'enum is replaced byview: 'working-memory'|'shared-working-memory'|'verified-memory'+includeSharedMemory: boolean. The legacy'union'mode (which conflated two orthogonal axes) becomesview: 'working-memory'+includeSharedMemory: true. Daemon-side wire shape was already two-axis (DkgClient.queryaccepted both as separate fields perclient.ts:133-183); this is a public-tool-surface alignment only — no daemon change.dkg_list_projects→dkg_list_context_graphs(commite0ca5e6a). Description picks up the audit v1.1 verbatim-locked(also called 'projects' in the DKG node UI)reconciliation note (SKILL.md §6 user-vs-internal terminology).dkg_list_subgraphs→dkg_sub_graph_list(commite0ca5e6a). Hyphenation alignment with the adapter (dkg_sub_graph_*family).2d69cef1,81859985,a966fa99) — four tools'descriptionstrings re-aligned to audit v1.1 verbatim text.dkg_sub_graph_createsemantics oscillated through2a803940/a966fa99(parity-analyst self-correction → final lock as idempotent, divergent from the adapter's strict semantics; matrix v0.8 §4.18 documents the divergence as deliberate UX consistency over byte-for-byte adapter parity).Keeper drops — 10 tools retired
Commit
c222ddcf(refactor(mcp-dkg): drop 10 V9-era and coding-project tools). The dropped surface, by category:dkg_get_ontologyand seeding scripts (no V10 endpoint backing it; SKILL.md §6 deferred to a future per-CG ontology workflow).dkg_propose_decision,dkg_add_task,dkg_comment. These deferred to rawdkg_assertion_*writes; the sugar can come back as templated wrappers when collaborative-write semantics matter.dkg_annotate_turnand the deferred-rendezvousforSessionmode. Reintroduction-pointer is thepre-v10-tool-droptag.capture-chat.mjsreduced from 678 lines to a focused chat-turn-only capture (commit7a1be910); the auto-promote semantics moved into per-CG config flags as the reintroduction path.dkg_search(free-text) anddkg_sparql(raw SPARQL) — the former is replaced by the trust-weighteddkg_memory_search; the latter renamed todkg_query.Recovery:
git show pre-v10-tool-drop:packages/mcp-dkg/src/tools/<file>. Design intent for each drop, plus reintroduction-pointers, inv9-design-archive.md.Starter-ontology agent-guide stubs (R2 sweep)
Commit
c25da46c(chore(mcp-dkg): R2 sweep starter ontologies for V9-era tool refs). The agent-guide markdown files underpackages/mcp-dkg/templates/ontologies/{book-research,coding-project,narrative-writing,pkm,scientific-research}/previously embedded V9 tool names (dkg_propose_decision,dkg_search, etc.). The sweep replaces those references with the V10 surface (dkg_assertion_*family +dkg_memory_search+dkg_query) so agents reading the guides on first use see canonical names. The TTL files themselves carry no tool refs and are unchanged.Capture-chat reduction (R3 sweep)
Commit
7a1be910(chore(mcp-dkg+cursor): R3 reduce capture-chat hook to V10 chat-turn capture only). The hook script atpackages/mcp-dkg/hooks/capture-chat.mjsis reduced from a multi-mode capture-and-promote machine to a focused chat-turn-only capture (678 → ~150 lines). The auto-promote and per-domain-tag features moved into the dropped-tools archive; the hook now writeschat:Turntriples and stops, deferring promotion to the canonicaldkg_assertion_promotepath.Root
AGENTS.mdremovalCommit
e128bb7f(chore(repo): R4 sweep V9 cruft). The repo-rootAGENTS.mdwas a V9-era doc pointing at the legacymcp-serverand the dropped capture-chat semantics. Deleted entirely; canonical agent-facing docs are nowpackages/mcp-dkg/README.md(tool surface) and the daemon-servedSKILL.mdat/.well-known/skill.md(API contract).Transition — CI / CLAUDE.md / package metadata / test pins
Commit
abc1a35a(chore(v10-mcp): align CI + CLAUDE.md + package metadata for mcp-server deletion)..github/workflows/ci.yml— filter changed from@origintrail-official/dkg-mcp-serverto@origintrail-official/dkg-mcpso the keeper's test lane doesn't silently drop on deletion.CLAUDE.mdlines 14 + 183 — MCP server config block flipped to point at the keeper; ontology TTL pointer flipped to the re-homedpackages/mcp-dkg/schema/dev-context-graph.ttl(renamed fromdev-paranet.ttlin the same move; line 193'sparanetId?signature comment normalised tocontextGraphId?).package.json—mcpscript removed (it pointed at the legacy package).packages/cli/test/integrations.test.ts— install-recipe pin updated from["-y", "@origintrail-official/dkg-mcp@0.1.0"]to the unpinned form (the published version is dev-tagged and moves; pinning to a literal@0.1.0would not resolve).packages/mcp-dkg/package.json— version bumped from0.1.0to10.0.0-rc.3for cross-package alignment with the rest of the V10 fleet.Legacy
mcp-serverdeletion + lockfileCommit
af8d9165(feat(v10-mcp): delete legacy packages/mcp-server (V10 keeper consolidation)). Removes the entirepackages/mcp-server/directory:src/(3 files),test/(3 files),schema/(re-homed earlier indev-paranet.ttl→dev-context-graph.ttl),package.json,README.md,tsconfig.json,vitest.config.ts.pnpm-lock.yamlregenerates accordingly (-19 lines net).adapter-autoresearch's loader-contract repoint (commitf6359d45, lands in the same PR) is the dependency that gates this deletion.Adapter-autoresearch repoint
Commit
f6359d45(feat(adapter-autoresearch): repoint to mcp-dkg loader contract). The single non-keeper consumer of the legacymcp-serverflips its loader-contract import:packages/adapter-autoresearch/src/tools.ts(+116 / -…) — call-site widens to the new(server, client, config)signature.packages/adapter-autoresearch/src/types.ts(+41 / -…) — type signatures updated.packages/adapter-autoresearch/test/tools.test.ts(+373 / -…) — test scaffolding rewritten against the new shape.packages/adapter-autoresearch/README.md(+10 / -…) — install snippet updated.Fixture suite (mcp-dkg 52 → 69; cli +9 new)
Commits
7b760331(W3-A wave-1+wave-2 fixtures),ebbdba28(W5-Q regression guards), ande78d8027(W6-pre setup-flow tests).packages/mcp-dkg/test/harness.ts(new, +307) — shared test harness for stdio MCP boots.packages/mcp-dkg/test/assertion-lifecycle.test.ts(new, +203) — covers the create + write + promote + discard + query round-trip.packages/mcp-dkg/test/memory-search.test.ts(new, +130) — covers the trust-weighted fan-out and thecontextGraphId+layer+trustWeighthit shape.packages/mcp-dkg/test/query-schema.test.ts(new, +159) — covers the two-axisview+includeSharedMemoryschema migration.packages/mcp-dkg/test/setup-publish-health.test.ts(new, +231) — coversdkg_context_graph_create,dkg_subscribe,dkg_sub_graph_create,dkg_publish,dkg_shared_memory_publish,dkg_status,dkg_wallet_balances.packages/mcp-dkg/test/drop-sweep.test.ts(new, +139) — guards against accidental reintroduction of the 10 retired V9 tools (drop-sweep) and the V9-era SPARQL prefixes (read-side regex). Replaces the original commite3e8eb75's shape after64880ce0reverted in favour of the cleaner version.packages/cli/test/mcp-setup.test.ts(new, +219 — W6-pree78d8027) — 9 tests covering the bundleddkg mcp setupflow under stubbed deps: clean-machine first-run, existing-config skip, all five flag-skip permutations (--no-start,--no-fund,--no-verify,--dry-run,--print-only),--port/--namefirst-init overrides, and the faucet-failure manual-curl fallback. Lives inpackages/cli/test/(a different vitest project frompackages/mcp-dkg/test/).capture-hook.test.ts(-263),normalise-slug.test.ts(-125),starter-ontologies.test.ts(-118),uri-helpers.test.ts(-52). All four were tied to dropped surfaces; resurrecting viapre-v10-tool-dropwould resurrect dead code paths.README rewrites + docs co-changes
Commits
0b21f09a(W5-A repo-root) andf40002c8(W5-A mcp-dkg).README.md(+131 / -18) — Quick Start unified into a routing table + parallel-depth on-ramp recipes for MCP and OpenClaw, plus an inline 10-step "DKG V10 as agent memory" round-trip quickstart that reproduces the Phase-1 baseline. CLI cheat-sheet gainsdkg mcp setup/dkg mcp serve. Setup-Guides table gets the MCP row.packages/mcp-dkg/README.md(+160 / -128) — full canonical V10 form. Drops absolute-path snippets in favour ofdkg mcp setup. Documents all 21 tools by category (Health/identity, Discovery, Setup, Write, Publish, Search & query) using audit v1.1 verbatim descriptions. Adds thedkg_publishvsdkg_shared_memory_publishdistinction, the two-axisdkg_queryschema, thedkg_memory_searchhit shape, a package-layout table, and a historical-recovery note pointing atpre-v10-tool-dropandv9-design-archive.md.docs/onboarding/04-package-map.md— six legacymcp-servermentions rewritten to point at the keeper (lines 23, 50, 182, 183, 244, 267).docs/TWO-LAPTOP-DEMO.md(-170) — V9-era demo doc retired alongside the dropped surfaces.packages/mcp-dkg/docs/RECONCILIATION.md(-103) — V9-era reconciliation doc retired.packages/adapter-autoresearch/README.md(+10 / -…) — loader-contract block updated.Plan-doc historical record
Commits
0567b8a1and12a90a12restore historical-record references in the planning artifacts underagent-docs/dkg-v10-mcp-consolidation/per team-lead's wave-3-final ruling. No production-code impact.Regression guards
The
drop-sweep.test.tsandread-side regexchecks in commitebbdba28are the cheap pre-merge guardsqa-engineerproposed in W5-Q: they fail-fast if anyone reintroduces a dropped tool name or a V9 SPARQL prefix in the mcp-dkg source.Test plan
Co-authored by
qa-engineerfromverification-plan.md v10.1§0.10 (W3-D post-deletion verification againstaf8d9165) + §0.11 (W5-Q regression guards + docs-as-tutorial walkthrough against0b21f09a + f40002c8, re-verified against1c4ff9ef + 8e154a15after W6-pree78d8027collapsed steps 2+3+5 into a singledkg mcp setup).Pre-merge gates — all green
pnpm run build→ 19/19 turbo tasks.pnpm --filter @origintrail-official/dkg-mcp run buildclean. The umbrella CLI's prebuild builds adapter-openclaw + mcp-dkg in sequence; cli's own tsc step succeeds. Closes block-list item Refactor autoupdater #6 (CLI deep-import regression guard post-packages/mcp-server/deletion).dkg mcp servestdio. Spawned the umbrella CLI'sdkg mcp servesubcommand againstaf8d9165, exchangedinitialize+tools/listover JSON-RPC. Response: exactly 21 tools by name (sorted):dkg_assertion_create,dkg_assertion_discard,dkg_assertion_history,dkg_assertion_import_file,dkg_assertion_promote,dkg_assertion_query,dkg_assertion_write,dkg_context_graph_create,dkg_get_agent,dkg_get_entity,dkg_list_activity,dkg_list_context_graphs,dkg_memory_search,dkg_publish,dkg_query,dkg_shared_memory_publish,dkg_status,dkg_sub_graph_create,dkg_sub_graph_list,dkg_subscribe,dkg_wallet_balances. None of the 10 dropped names appear; all 3 renamed canonical names appear; all 9 audit-locked adds appear (incl.dkg_assertion_historyopportunistically shipped from the audit's P3 deferred bucket).pnpm --filter @origintrail-official/dkg-mcp test. Distribution:assertion-lifecycle.test.ts:11+setup-publish-health.test.ts:19+memory-search.test.ts:8+query-schema.test.ts:14+drop-sweep.test.ts:17. Pre-W2 suite was 52 (W3-A baseline7b760331); +17 from W5-Q'sdrop-sweep.test.ts(ebbdba28). Thee3e8eb75follow-up was reverted by64880ce0in favour of the cleaner standalone implementation, so the +19 increment indocs-lead's pre-fill is now+17.pnpm --filter @origintrail-official/dkg-adapter-autoresearch test. Loader contract re-pointed inf6359d45from the legacymcp-servershape to the keeper's(server, client, config)signature; tests pass against the new contract.packages/cli/test/mcp-setup.test.ts9/9 passing (W6-pree78d8027).pnpm --filter @origintrail-official/dkg test. Covers the bundleddkg mcp setupflow under stubbed deps: clean-machine first-run, existing-config skip, all five flag-skip permutations (--no-start,--no-fund,--no-verify,--dry-run,--print-only),--port/--namefirst-init overrides, faucet-failure manual-curl fallback. Companion to the install-flow simulation gate below.pnpm --filter @origintrail-official/dkg-adapter-openclaw testshows 808/810 (the 2 R7 failures noted below are unrelatedsetup.test.ts:344autoUpdate config-shape regressions; the round-trip's contract tests atmemory-integration.test.ts,dkg-memory.test.ts,memory-search-tool.test.tsall pass). The deletion-gate harness atagent-docs/dkg-v10-mcp-consolidation/scripts/run-openclaw-roundtrip.shwas the canonical pre-deletion gate per §0.10.3; live e2e via the harness is the final pre-merge release-readiness check.git grep -nE 'packages/mcp-server|@origintrail-official/dkg-mcp-server|dev-paranet\.ttl' -- ':!agent-docs/dkg-v10-mcp-consolidation/**'→ ZERO. Plus all 10 dropped names + 3 renamed-from names absent acrosspackages/**,docs/**, repo-root: only 4 intentional migration-context comments survive, all confirmed exempt per mcp-lead's Stale dist when installing adapter from local path (I-021) #21 cover note (vitest.config.ts:10,src/tools.ts:141,src/tools.ts:167,src/tools/assertions.ts:7— historical pointers, not stale code references). The singledkg_sparqlmention inpackages/mcp-dkg/README.md:148is intentional historical-context migration documentation explainingdkg_query'sview: 'union'replacement; same exempt pattern.dkg mcp setup --print-onlybyte-identical to documented JSON. Locally:node packages/cli/dist/cli.js mcp setup --print-onlyemits{ "mcpServers": { "dkg": { "command": "dkg", "args": ["mcp", "serve"] } } }. Byte-identical match to repo-rootREADME.md:92-101,packages/mcp-dkg/README.mdsetup snippet, and repo-rootCLAUDE.md:14. The full Docker-isolated install viapnpm pack+ clean container is deferred to release-readiness (see below).drop-sweep.test.ts(ebbdba28):it.eachover the 10 dropped names (dkg_review_manifest,dkg_annotate_turn,dkg_get_ontology,dkg_get_chat,dkg_set_session_privacy,dkg_request_vm_publish,dkg_search,dkg_propose_decision,dkg_add_task,dkg_comment) asserts each is absent; surface-size lock at 21 fires on accidental adds OR removes.it.eachover the 5 read-side tools (dkg_assertion_write,_promote,_discard,_query,_history) asserts each acceptsname: "Bad Name With Spaces"per matrix v0.5 §4.16; positive-control test confirmsdkg_assertion_createSTILL rejects the same input (asymmetry visible in one file).0b21f09a + f40002c8literally — every CLI command exists (dkg init,dkg start,dkg stop,dkg status,dkg auth show,dkg logs,dkg mcp setup,dkg mcp serve,dkg openclaw setupwith all 4 documented flags--no-fund / --no-start / --no-verify / --dry-run,dkg integration {list,info,install}with--tier / --allow-community / --no-verify-provenanceflags,dkg context-graph create,dkg shared-memory write/publish,dkg assertion import-file); every JSON output matches the README; every cross-link target resolves (packages/adapter-elizaos/README.md,packages/adapter-hermes/README.md,packages/adapter-openclaw/README.md,packages/mcp-dkg/README.md,docs/setup/SETUP_OPENCLAW.md,docs/setup/SETUP_CUSTOM.md,docs/setup/JOIN_TESTNET.md);mcp-dkg/README.mdenumerates exactly 21 tool table rows; first-run verification (README:105) green via the same stdio probe —tools/listincludesdkg_assertion_create,dkg_assertion_write,dkg_memory_search; env-var fallback claim matchespackages/mcp-dkg/src/config.ts:11-16exactly for all four (DKG_API,DKG_TOKEN,DKG_PROJECT,DKG_AGENT_URI). Re-verified against1c4ff9ef + 8e154a15after W6-pree78d8027bundleddkg init+dkg startintodkg mcp setup(now a 2-command flow): all 7 README:110-promised flags exist on the bundled command (--port,--name,--no-start,--no-fund,--no-verify,--dry-run,--force);--print-onlyJSON output unchanged byte-identical; help text matches the README's "one-shot bundled flow" description. Zero divergences flagged on either pass. docs-lead's rewrite is copy-pasteable end-to-end.Pre-existing failures noted (NOT introduced by this PR)
packages/adapter-openclaw/test/setup.test.ts:344— 2/810 failures in thewriteDkgConfigautoUpdate-config-shape tests. Confirmed reproducible ataf8d9165post-deletion AND onmainpre-PR; unrelated to MCP consolidation. Per team-lead's 2026-05-04 sequencing call: surface separately as a post-merge concern. Verification-plan v10 §0.4 / §0.9.3 / §0.10.3 record the chain of confirmations.Release-readiness gates — explicitly deferred to post-PR-ready / pre-merge
npm install -g @origintrail-official/dkgagainst npmjs.org — depends on a publish-authorization the user has not granted. The package's version is10.0.0-rc.3perabc1a35a(cross-package alignment); not yet tagged for release.rm -rf ~/.dkg ~/.cursor/mcp.json ~/.claude/mcp_servers.json, then global install +dkg mcp setup+dkg mcp servestdio +tools/list21-tool assertion. Replaces the originally-planned Docker-based clean-room test: the DKG daemon itself never runs in Docker (Docker references in the repo are limited to optional Blazegraph deployment for devnet and third-party community-integration runtimes), so a Docker test would have masked workstation-specific risks (PATH issues, file permissions, npm global config) that real users hit. Local simulation already passed all gates per §0.10.5 / §0.11.2; the workstation pass is the final pre-merge release-readiness check.dkg_assertion_create + dkg_assertion_write + dkg_memory_searchend-to-end. Manual verification step for the release cycle, gated on the livenpm install -gworking.QA sign-off
qa-engineersigns off the Test-plan section based on §0.10 + §0.11 verdicts. The 10 pre-merge gates above are green ataf8d9165 + ebbdba28 + e78d8027. The 3 release-readiness gates are deferred and documented; they do not block PR-open. Block-list items #1 (OpenClaw round-trip) and #6 (CLI deep-import) — the two non-negotiables from v0 — are both satisfied.Migration notes
For any
present-but-divergenttool, the BEFORE/AFTER notes below cover what consumers see across the consolidation.dkg_queryschema migration (wasdkg_sparql)Before: single-axis
layer: 'wm' | 'swm' | 'union' | 'vm'enum.After: two-axis
view: 'working-memory' | 'shared-working-memory' | 'verified-memory'+ orthogonalincludeSharedMemory: boolean. The legacy'union'mode (which conflated two axes) is nowview: 'working-memory'+includeSharedMemory: true.Why:
'union'was an enum-conflation that hid the orthogonality. The new shape lets callers express "private working memory plus the team-shared layer" without needing a special mode for it. Daemon-side wire shape already accepted both as separate fields (DkgClient.queryatclient.ts:133-183); this is a public-tool-surface alignment only — no daemon change required.Action for consumers: rename calls; map
layer: 'union'toview: 'working-memory'+includeSharedMemory: true. Tool name also flips fromdkg_sparqltodkg_query.dkg_list_context_graphs(wasdkg_list_projects)Before:
dkg_list_projects— name diverged from the daemon's/api/context-graph/listendpoint.After:
dkg_list_context_graphs. Description carries the(also called 'projects' in the DKG node UI)reconciliation note so users see both terms.Action for consumers: rename. Output shape unchanged.
dkg_sub_graph_list(wasdkg_list_subgraphs)Before:
dkg_list_subgraphs— hyphenation diverged from the adapter'sdkg_sub_graph_*family.After:
dkg_sub_graph_list. Sub-graph-CRUD family (dkg_sub_graph_create,dkg_sub_graph_list) is now consistently hyphenated.Action for consumers: rename. Output shape unchanged.
dkg_memory_searchhit-shape additionsBefore: raw bindings array per memory layer.
After: each hit carries
contextGraphId(the CG the hit was found in),layer(a combined SKILL §6.3 string of the formworking-memory:project-id/shared-working-memory:project-id/verified-memory:project-id), andtrustWeight(a numeric score; higher-trust layers collapse lower-trust hits for the same entity URI). Three new first-class fields.Why: lets consumers (and downstream agents) reason about where a memory comes from and how much to trust it without a second SPARQL round-trip.
Action for consumers: existing keyword/limit input arguments unchanged; if your code reads only the bindings array it still works, but you'll likely want to surface the new fields to your agent prompt or UI.
Dropped tools —
pre-v10-tool-droprecovery pathThe 10 retired tools listed under "Keeper drops" are not present at any name in the V10 surface. For consumers needing the V9 surface:
git show pre-v10-tool-drop:packages/mcp-dkg/src/tools/<file>returns the file at HEAD before deletion.agent-docs/dkg-v10-mcp-consolidation/v9-design-archive.mdcovers the four substantive V9 patterns we retired and the conditions under which each could come back (per-CG ontology workflow lands first → ontology starters; collaborative-write semantics matter again → sugared decision/task writes; etc.).dkg_assertion_create + dkg_assertion_writecovers thedkg_propose_decision/dkg_add_taskuse cases at the cost of one extra call) or pin to thepre-v10-tool-droptag for a frozen V9 surface.Setup flow (was 4 commands + absolute paths; now 2 commands)
Before: four-command sequence with absolute paths in the wiring snippet.
After: two-command bundled flow per W6-pre (
e78d8027).dkg mcp setupis now an idempotent installer that runs init, daemon-start, faucet funding, MCP-client registration, and verification in a single invocation. Every step short-circuits when its work is already done, so re-running is safe.The wiring block written into each detected client's config is
{ "command": "dkg", "args": ["mcp", "serve"] }— copy-pasteable verbatim from the README; no tokens or URLs in the JSON (auth lives in~/.dkg/auth.token, config in~/.dkg/config.yaml, env-var fallbacksDKG_API/DKG_TOKEN/DKG_PROJECT/DKG_AGENT_URIfor environments where the init step can't run).Action for consumers: run
dkg mcp setuponce after upgrading. The flow detects existing config / daemon / clients and skips work that's already been done.--forceto refresh stale client entries without prompting;--no-fundfor CI;--print-onlyto emit the JSON for manual paste when no client is detected.🤖 Generated with Claude Code