0.8.0

0.8.0 (2026-10-12)

SpecsGraph lets agents take a project from spec to built code (AGENT-FLOW-PLAN). The MCP server now has 59 tools (18 read, 41 write). The new tools are build_clusters, build_plan, roadmap_manage, build_markImplemented, work_item_claimNext, work_item_claims and task_publish. spec_apply takes route; task_route takes addUnrouted, ready, items and publish; work_item_claim takes claimant; work_item_update takes claimant and done. A new project setting, Agents may publish (Admin, needs Agent autonomy), lets agents publish; a person publishes otherwise.

Added

• specsgraph-build: the end-to-end loop and the multi-agent playbook. Check (project_list flags, build_status, completion from build_clusters); a fast spec pass (spec_apply up to 500 documents per call, then task_route with addUnrouted, ready and publish, or route on the last call when small, items left to build_plan); a baseline for existing code (build_markImplemented with a note and codeRef); the plan, in one planner session (build_clusters, build_plan with milestones, repeated while remaining, roadmap_manage); the work loop per agent (work_item_claimNext with the server-given claimant, previous, a branch and pull request citing T-n.s, work_item_update with codeRef, heartbeats, done when no repository is watched, the reason of an empty answer); the multi-agent rules (work_item_claims, one item per claimant, never another session's item); and a table of refusals and what they mean.

Changed

• Engineer: no separate proposal_open step (spec_apply opens the Proposal). Agent autonomy covers task_route with addUnrouted and agent publish (task_publish). "Build status and the roadmap" covers build_markImplemented, roadmap_manage, the empty-roadmap path and the hand-over to specsgraph-build. "Implementing agreed work" takes items with work_item_claimNext and a claimant, cites T-n.s, continues from previous, and completes with done only when no repository is watched.
• Ingest: under Agent autonomy, direct writes in batches of up to 500 documents (no stage: true), 25 to 50 when staged for review, and a final phase that routes, readies and publishes with task_route (addUnrouted) or spec_apply route. proposal_get before each phase only when staging.
• Director: routes "build the app end to end" to specsgraph-build, and describes the multi-agent fan-out (one planner session, N worker sessions, server-given claimants).
• Product and brownfield: a person publishes unless the project lets agents publish (task_publish). Brownfield records existing code with build_markImplemented under Agent autonomy; both offer specsgraph-build.
• /specsgraph:setup writes the autonomy line, the agent publish line, work_item_claimNext with a claimant, T-n.s in pull request titles, work_item_claims and the end-to-end loop line. The session reminder says Main moves when a task publishes, by a person or by an agent where the project allows it.
• README: six skills, the Build step of the loop, Agents may publish in the loop, principles and FAQ.
• scripts/tools.json regenerated from the 59-tool registry.

0.7.0

SpecsGraph adds work items, a derived build status, a roadmap and two new kinds. The MCP server now has 52 tools (16 read, 36 write). The new tools are work_item_next, work_item_claim, work_item_update, build_status and roadmap_get. proposal_plan takes workItems. thread_open, thread_ask and thread_list take item and workstream: main, and thread_ask takes askUser.

Added

• Engineer: "Implementing agreed work", which takes a work item with work_item_next and work_item_claim, reads the spec at the item's read.scope, names the item in the pull request, records it with work_item_update (codeRef), heartbeats and asks with thread_ask (item, blocking) when stuck. The merge, not the agent, marks the item done.
• Engineer: "Build status and the roadmap", which reads build_status and roadmap_get before answering what is built or what comes next.
• Engineer: the Architecture kinds, decision (status, background, resolution, consequences, alternatives, applies-to) and constraint (category, statement, measure, applies-to).

Changed

• Engineer: bounded contexts carry the context map (upstreams with a pattern); use cases carry preconditions; a new integration event needs a bounded-context-id; roles carry actor-kind; threads and questions may be on Main.
• Product: features and scenarios link the use cases that realise them (use-cases); roles may be external systems.
• Brownfield: build status replaces "SpecsGraph has no implemented flag": a person marks mapped code as implemented on the Build page. The code-to-kind table adds context map, external systems, decisions and constraints.
• Ingest: sixteen kinds, the context map, decisions and constraints in the foundation, use cases on features, and a closing note on marking existing code as implemented.
• Director: sixteen kinds in six groups, and build or roadmap questions answered read only.
• /specsgraph:setup writes the work-item section the web app's Connect an agent panel shows. The session reminder mentions work items.
• README: the loop adds a Build step, the kind table has 16 kinds, and the FAQ says an agent cannot mark code done.
• scripts/tools.json regenerated from the 52-tool registry.

0.6.0

SpecsGraph adds Agent autonomy, a per-project setting an Admin controls (on for new projects, off for existing ones). The MCP server now has 47 tools (13 read, 34 write). The new tools are member_list, proposal_accept, task_route, task_assign, task_submitReady, task_markDone, thread_resolve, thread_answer, workstream_discardChange, workstream_resolveConflict, workstream_complete and workstream_abandon. task_open takes artefactIds and assignee, and spec_apply takes stage.

Changed

• Product, engineer and brownfield: with Agent autonomy on, spec_apply writes agreed documents straight into the workstream (stage: true keeps a review), and the skill may accept, route, assign and submit Ready for what the user agreed, in one proposal_accept when it can. Without it, those acts stay a person's, as before. Publishing stays a person's act either way.
• Ingest: under Agent autonomy, an import still stages for review (stage: true) unless the user asks for direct writes.
• /specsgraph:setup and the session reminder mention Agent autonomy.
• README: the loop, design principles and FAQ cover Agent autonomy.
• scripts/tools.json regenerated from the 47-tool registry.

0.5.0

Changed

• specsgraph-ingest reads every source a project lives in, not only the tracker. Phase 1 starts with a source inventory (tracker, code, Figma and other designs, Confluence or Notion, Drive or SharePoint, Slack or Teams, API and data specs, meeting notes). Each source records how it is reached and its read status, and an unreachable source becomes a gap and a question, never a silent skip. A table says what to take from each kind of source and which artefacts it yields. Every fact carries a short source reference. The blueprint reconciles sources with an agreed precedence and turns contradictions (ticket against design against code against chat) into questions. It also keeps a coverage matrix per source. Reading never writes to a source.
• Director routes whole-project capture from code, designs, docs or chat to specsgraph-ingest.
• README: what the skills read for an ingest, and that they only use connectors and credentials the user already has.

0.4.0

Added

• specsgraph-ingest: captures an existing project from its tracker backlog, documents and attachments (optionally its code) into one workstream, in phases. The phases are snapshot, blueprint, foundation, model per context, behaviour per context, questions, and an optional verify. Every kind's fields are filled, and every unconfirmed point is staged as a default and asked with thread_ask (options, current default, source keys, deduplicated against thread_list). A ledger and per-context briefs on disk make the import resumable and cheap to continue. It follows the MCP contracts closely: cheap reads (fields, kind, context, state, cursor, maxBytes), idempotencyKey on every write, dryRun for a new kind, optional plans per context, the thread_ask refusals, and answeredSince taken from the server's own timestamps. It also covers what real imports hit: Proposals accepted mid-run, YAML quoting, read-model key and mapping, handler triggers that need the other contexts' events first, twins across contexts.

Changed

• Director routes backlog and document imports to specsgraph-ingest.
• Brownfield hands a whole-project capture over to specsgraph-ingest.
• README: five skills, and the version badge now follows plugin.json.

0.3.0

SpecsGraph SDLC P1 is live: the MCP server has 35 tools (12 read, 23 write). task_submitReady is gone; thread_ask, proposal_plan and proposal_assign are new, and spec_apply takes plan.

Changed

• /specsgraph:setup writes the section the web app's Connect an agent panel shows: read workstream:WS-n while the spec is in flight, ask with thread_ask instead of guessing, people mark a task Ready and publish it and the landing marks it done, cite tracker keys such as KAN-43. It no longer names task_submitReady.
• Product, engineer and brownfield: stage work for a task into a plan (spec_apply with plan, proposal_plan, proposal_assign); ask the team with thread_ask (options, blocking, askTeam) and collect answers with thread_list (kind, answeredSince); change requests hold an accept until the artefact is staged again; only a person reopens a question or a change request, and a conflict thread never reopens. Closing steps say Ready is a person's act and the task moves to Done when its publication lands.
• Director: orientation checks a workstream for questions still waiting on people.
• README: plans, and no tool marks a Task Ready or answers a question.
• scripts/tools.json regenerated from the P1 tool registry.

0.2.7

Fixed

• .mcp.json names the server URL plainly. Claude Code filled in its shell-style default, but claude.ai and Claude Desktop showed it as a placeholder to replace when the plugin added its connector. The URL can no longer be overridden: SpecsGraph is cloud only.

0.2.6

Fixed

• plugin.json sets icon again: the directory listing reads the icon from that key and showed a letter without it.

0.2.5

Changed

• CHANGELOG wording only, for the directory scan.

0.2.4

Changed

• The SessionStart hook falls back to the current directory as . instead of $PWD, which the directory scanner read as a password.
• plugin.json drops icon and privacyPolicyUrl (the directory flagged them as unknown keys; it reads .claude-plugin/icon.svg and the README's privacy link on its own), and the unused assets/specsgraph-og.png is gone.