@pcircle/memesh 4.8.3 → 4.9.0

package/.claude-plugin/marketplace.json

@@ -8,7 +8,7 @@
       "name": "memesh",
       "source": "./",
       "description": "MeMesh \u2014 agentic memory for coding agents. Captured from the agent's real work via hooks, recalled when it acts. One SQLite file, zero cloud required.",
-      "version": "4.8.3",
+      "version": "4.9.0",
       "author": {
         "name": "PCIRCLE AI"
       },

package/.claude-plugin/plugin.json

@@ -4,7 +4,8 @@
   "author": {
     "name": "PCIRCLE AI"
   },
-  "version": "4.8.3",
+  "version": "4.9.0",
+  "mcpServers": "./.claude-plugin/mcp.json",
   "homepage": "https://github.com/PCIRCLE-AI/memesh",
   "repository": "https://github.com/PCIRCLE-AI/memesh",
   "license": "MIT",

package/.codex-plugin/mcp.json

@@ -1,7 +1,9 @@
 {
-  "memesh": {
-    "command": "node",
-    "args": ["./dist/mcp/server.js"],
-    "cwd": "."
+  "mcpServers": {
+    "memesh": {
+      "command": "node",
+      "args": ["./dist/mcp/server.js"],
+      "cwd": "."
+    }
   }
 }

package/.codex-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "memesh",
   "description": "MeMesh — agentic memory for coding agents.",
-  "version": "4.8.3",
+  "version": "4.9.0",
   "mcpServers": "./.codex-plugin/mcp.json"
 }

package/AGENTS.md

@@ -23,7 +23,18 @@ host is recallable from all of them. Not installed yet? Follow
      field out if it was not said.
 3. **"What do you remember?"** — call `briefing` and relay its content. Do
    not answer from your own conversation context.
-4. **When you need another agent** — to hand off, to ask, to report back —
+4. **When memory could be condensed from a calendar cluster or a recent Claude Code session** —
+   when the host supports interactive prompts, offer concise choices in the
+   user's conversation language, such as **Dispatch agent task**, **Later**, or
+   **Don't suggest again this session**.
+   The last choice suppresses only this session's prompt; it does not create a
+   durable opt-out. Dispatch means
+   using `work_package` in this already-running agent session; it does not mean
+   the Dashboard can start or wake an agent. Prepare either one calendar-selected
+   digest package or, when the MCP client supplies one unambiguous matching
+   workspace root, one package from its newest Claude Code session's visible
+   turns. Submit one bounded result for human review or defer without writing.
+5. **When you need another agent** — to hand off, to ask, to report back —
    first use `message discover` with the exact project when you do not already
    know the recipient. It lists only live registrations and their routing IDs,
    host kind, declared model/current work, generation, and lease; missing
@@ -35,10 +46,11 @@ host is recallable from all of them. Not installed yet? Follow
    `recipient`; poll first, then fetch each returned `message_id`. Fetching
    does not acknowledge.
 
-## All 11 MCP tools
+## All 12 MCP tools
 
 | Tool | Purpose |
 |---|---|
+| `work_package` | Prepare one bounded untrusted digest (calendar-selected) or transcript package from the newest Claude Code session under the client's single matching MCP workspace root; submit one strictly validated result for pending human review or defer without durable change. Submission retains bounded redacted source turns for comparison; agents cannot apply or reject, and hashes identify freshness and workspace scope rather than authentication. |
 | `remember` | Store knowledge as an entity with observations, tags, and relations |
 | `recall` | Search stored knowledge (words are OR-ed, ranked by relevance); empty query lists recent |
 | `forget` | Archive an entity (soft-delete), or remove one observation via the `observation` parameter |
@@ -78,9 +90,8 @@ host is recallable from all of them. Not installed yet? Follow
   before concluding anything.
 - Every recall answer carries a `retrieval` block that says how it was
   produced — read it instead of guessing: `truncated: true` means the
-  window filled and more may exist; `degraded: true` means semantic search
-  is configured but could not run, so you are seeing keyword-only results
-  right now (worth telling the user, and `memesh doctor` explains why).
+  window filled and more may exist. Recall is local FTS5 search; do not
+  describe results as semantic, vector-ranked, or model-generated.
 
 ## What Claude Code already does — do not double-write
 
@@ -91,7 +102,10 @@ Under Claude Code with the MeMesh plugin, hooks capture automatically:
 - **PreToolUse (Edit|Write)** surfaces memories related to the file being
   edited.
 - **PostToolUse (Bash)** records git commits with diff stats.
-- **Stop** captures session knowledge and turns failures into lessons.
+- **PostToolUse (ExitPlanMode|AskUserQuestion)** reminds you to `remember` a
+  decision just made — once per tool per session. It only reminds; unlike
+  the hooks above, it writes nothing to the graph itself.
+- **Stop** captures bounded session evidence, including observed error/fix signals.
 - **PreCompact** saves important knowledge before history is compressed.
 - **UserPromptSubmit** detects "remember this" intent in the prompt.
 - **PreToolUse (Bash)** fires accepted lesson-guards: a fenced warning
@@ -100,25 +114,32 @@ Under Claude Code with the MeMesh plugin, hooks capture automatically:
   you do.
 
 So under Claude Code: skip step 1 of the loop (the topology is already
-injected), and do not `remember` commits or session summaries by hand. Manual
-calls are for what hooks cannot see — decisions and their rationale, lessons
-worth keeping, and user-stated task state.
+injected), and do not `remember` commits or session summaries by hand. A
+decision made via `ExitPlanMode`/`AskUserQuestion` gets a reminder, not a
+write — act on it with a real `remember` call. Manual calls are still how
+decisions, their rationale, lessons worth keeping, and user-stated task state
+actually get stored.
 
 On a host with no hooks (Gemini CLI, Cursor, an MCP-only setup, …) the loop is
 fully manual, and it is worth running.
 
-Codex CLI can be either. Wired as an MCP server it has no hooks, like the
-above. Installed as a plugin (`codex plugin add memesh@pcircle-memesh`) it
-reads the same `hooks/hooks.json` manifest Claude Code does and runs the same
-hook scripts, so the topology is injected for you there too. If you are unsure
-which one you are in, `memesh doctor` names it: the "Hooks wired into Claude
-Code" row says which plugin runtime it found.
+Codex CLI can be either. Wired only as an MCP server it has no hooks, so call
+`briefing` yourself. Installed as a plugin
+(`codex plugin add memesh@pcircle-memesh`) it wires the MCP server and the
+separate SessionStart companion. On macOS or Linux, an eligible ordinary Codex
+CLI startup or resume registers that exact active thread for native delivery.
+The companion does not run Claude Code's eight capture/recall hooks or prove
+that topology was injected. Do not assume Codex Desktop or an unattached task
+registered; confirm the exact live session with `message discover`.
+`memesh doctor` reports installation and local integration health, but it does
+not prove host acceptance or model-visible delivery.
 
 ## Working on this repository (contributing agents)
 
 The sections above are for agents whose user installed memesh. If you are an
-agent making changes to this codebase, [CLAUDE.md](CLAUDE.md) is the entry
-point; the policy it carries, compressed:
+agent making changes to this codebase, read [CLAUDE.md](CLAUDE.md) as the
+contributor entry point and apply the working and delegation requirements
+below with it:
 
 - **Blast radius picks the process.** One module, no security surface, no
   cross-surface behaviour → implement, run the affected tests + typecheck,
@@ -132,5 +153,35 @@ point; the policy it carries, compressed:
 - **Delegation is disjoint.** Two writers never touch one file; file-editing
   subagents work in isolated worktrees and never commit or push — the
   orchestrator reads every diff before it lands.
+- **Delegate the smallest verifiable slice.** Give each agent one coherent
+  outcome that can be implemented and falsified independently. Do not hand a
+  broad implementation or cleanup to one worker when it can be split at
+  file-disjoint boundaries. Read-only analysis may be divided further by
+  symbol, test, UI, documentation, or runtime question. A formal independent
+  reviewer still receives the same complete diff; specialist scope narrows
+  the questions, never the changed files. Delegation must reduce uncertainty
+  or work, not merely add an agent.
+- **Use an explicit task card.** Every delegated task states: observable
+  outcome; owned paths and forbidden paths; frozen repository input (branch,
+  commit and, when relevant, tree or artifact digest); required invariants and
+  failure paths; exact checks and readbacks; non-goals and mutation authority;
+  and the required findings/evidence return format. Unknowns must be reported,
+  not guessed. A task without a falsifiable acceptance check is not ready to
+  delegate.
+- **Treat DeepSeek as a bounded team worker.** Prefer the local DeepSeek worker
+  for small source inventories, candidate patches, test ideas, failure-log
+  analysis, mechanical classifications, and documentation comparisons when
+  its result can be checked cheaply. Give it only the minimum secret-free
+  context and least-privilege tools. It never receives credentials, MeMesh
+  database access, external mutation authority, architecture ownership,
+  formal independent-review authority, or completion authority. Its output is
+  an untrusted candidate until the orchestrator verifies it against the frozen
+  source and executable evidence.
+- **Keep roles and evidence separate.** The orchestrator owns architecture,
+  integration, claim coverage, authorization boundaries, and the final status.
+  Implementers return a candidate, never `PASS`. A fresh read-only reviewer
+  who authored none of the changed paths replays every claim on the exact
+  immutable candidate. Any source change invalidates earlier test and review
+  evidence for that candidate.
 - **Internal notes stay local.** Plans and scratch analyses are never
   committed and never appear in commit messages or release notes.