@primitive.ai/prim 0.1.0-alpha.90 → 0.1.0-alpha.92

package/README.md

@@ -39,7 +39,7 @@ One command does the whole install — auth, session hooks, daemon, git hooks,
 skill, and the welcome:
 
 ```bash
-prim setup                 # add --agent codex or --agent hermes; --no-daemon explicitly opts out
+prim setup                 # add --agent cursor, codex, or hermes; --no-daemon explicitly opts out
 ```
 
 Or run the steps individually:
@@ -49,7 +49,7 @@ Or run the steps individually:
 prim auth login
 
 # 2. Wire the session hooks (decision capture + presence)
-prim claude install        # or: prim codex install / prim hermes install
+prim claude install        # or: prim cursor install / prim codex install / prim hermes install
 
 # 3. Start the companion daemon (latency + team presence)
 prim daemon start
@@ -85,6 +85,7 @@ step-by-step the agent follows — and the manual fallback — live in
 
 ```bash
 prim setup                   # Run the whole install in one shot
+prim setup --agent cursor    # Same, for Cursor IDE and Cursor CLI
 prim setup --agent codex     # Same, for OpenAI Codex
 prim setup --agent hermes    # Same, for Hermes Agent (global-only config)
 prim setup --no-daemon       # Stop it and persistently opt out of supervised delivery
@@ -102,13 +103,13 @@ prim uninstall
 
 Stops the companion daemon, removes the current repository's Prim-owned agent
 and Git-hook surfaces when run inside a Git repository, removes the user-scoped
-Claude, Codex, Hermes, and Git-hook surfaces, then deletes only schema-valid
+Claude, Cursor, Codex, Hermes, and Git-hook surfaces, then deletes only schema-valid
 Prim runtime bytes. Foreign configuration is retained; ambiguous ownership
 makes the command fail closed and retain the runtimes. If a removal races or
 fails after one runtime changes, its JSON result reports each runtime's exact
 state so rerunning can safely finish cleanup. Authentication,
 undelivered journals, repository bindings, and agent skill guidance are kept.
-Use `prim skill uninstall --agent <claude|codex|hermes> --scope <project|user>`
+Use `prim skill uninstall --agent <claude|cursor|codex|hermes> --scope <project|user>`
 when you also want to remove a known skill target.
 
 ### Auth
@@ -142,8 +143,8 @@ graph, and presence is reported. Each hook
 self-resolves the CLI at run time (PATH, then a local install, then
 `npx --yes @latest`), so it keeps working with no global install.
 
-Installs into the current project by default — the repo's `.claude/settings.json`
-/ `.codex/hooks.json`, resolved from the git root (so any subdirectory works);
+Installs into the current project by default — the repo's `.claude/settings.json`,
+`.cursor/hooks.json`, or `.codex/hooks.json`, resolved from the git root (so any subdirectory works);
 pass `--scope user` to install machine-wide. Hermes is the exception: it reads
 shell hooks only from the global `~/.hermes/config.yaml`, so `prim hermes install`
 is always user-scoped — and prim merges in place, leaving the rest of that file
@@ -152,6 +153,8 @@ is always user-scoped — and prim merges in place, leaving the rest of that fil
 ```bash
 prim claude install                # Install Claude Code hooks (project scope; uninstall / status)
 prim claude install --scope user   # Install machine-wide instead
+prim cursor install                # Install native Cursor hooks (project scope)
+prim cursor install --scope user   # Install hooks, skill, and CLI footer machine-wide
 prim codex install                 # Install OpenAI Codex hooks (project scope)
 prim hermes install                # Install Hermes Agent hooks (global ~/.hermes/config.yaml)
 ```
@@ -308,9 +311,11 @@ and offers to install into `.husky/`.
 prim statusline        # Render the team-presence statusline (reads the daemon)
 ```
 
-Claude Code has one custom status-line slot. Installation uses a staged,
-lightweight Primitive renderer when that slot is empty or already Primitive's;
-an existing custom status line is preserved and reported explicitly. Use
+Claude Code and Cursor CLI each have one custom status-line slot. Their user-scope
+installations use a staged, lightweight Primitive renderer when that slot is empty
+or already Primitive's; an existing custom status line is preserved even with
+`--force` and reported explicitly. Cursor's Primitive footer replaces Cursor's
+default footer and requires a Cursor CLI restart after installation or removal. Use
 `prim daemon status` or `prim doctor` for the same health signal in that case.
 
 ### Welcome
@@ -337,6 +342,7 @@ prim moves flush          # Drain the local journals to the server (also runs fr
 
 ```bash
 prim skill install --agent claude   # Install the decision-graph guide for Claude Code
+prim skill install --agent cursor   # Install the native Cursor skill
 prim skill install --agent codex    # …or write the guide into another agent's rules file
 prim skill uninstall --agent claude # Remove it
 prim skill status --agent claude    # Report whether it's installed
@@ -347,9 +353,11 @@ Teaches your agent how to work with the decision graph. For **Claude Code**
 `<repo>/.claude/skills/prim/` (or `~/.claude/skills/prim/` with `--scope user`)
 — a `.claude-plugin/plugin.json` + `SKILL.md` that auto-loads as the model-invoked
 `prim@skills-dir` skill, no marketplace step; restart Claude Code or run
-`/reload-plugins` after installing. For every other agent it writes a managed
-block into the rules file that agent reads (`--agent codex` → AGENTS.md,
-`--agent hermes` → .hermes.md, or an auto-detected .cursor/rules, …). A bare
+`/reload-plugins` after installing. For **Cursor**, it installs an owned `SKILL.md`
+at `<repo>/.cursor/skills/prim/` (or `~/.cursor/skills/prim/` with `--scope user`)
+without overwriting foreign or locally edited files. For the remaining agents it
+writes a managed block into the rules file that agent reads (`--agent codex` →
+AGENTS.md or `--agent hermes` → .hermes.md). A bare
 `prim skill install` (no `--agent`) auto-detects a rules file and writes the
 block; pass `--target <path>` for an explicit file.
 

package/SKILL.md

@@ -11,7 +11,7 @@ description: Use the prim CLI for Primitive’s decision graph. MUST INVOKE befo
 
 In an active repo, prim passively captures the **decisions** your team makes -- which library, which pattern, which config value -- into a queryable graph, and links them: a decision can depend on earlier decisions (auto-linked from shared files, or related by hand — see *Relate decisions*) and reference the files it touched. **Conflict Gates** can check a later change against that graph and surface any load-bearing decision it conflicts with. When **Enforcement** is enabled for an organization, prim can warn, pause, or block the edit until you reconcile the decision and retry. Capture in active repos runs independently.
 
-Low-level capture runs automatically in active repos through the session hooks installed by `npx --yes @primitive.ai/prim claude install` (Claude Code), `npx --yes @primitive.ai/prim codex install` (Codex), or `npx --yes @primitive.ai/prim hermes install` (Hermes). An inactive repo is not passively captured. Deliberately record higher-order forks in the road through `prim decisions create` as described below. Your job is also to **read** the graph before load-bearing edits and **answer** the occasional rationale confirmation. (Responding to Conflict Gates applies only once Enforcement is enabled — see below.)
+Low-level capture runs automatically in active repos through the session hooks installed by `npx --yes @primitive.ai/prim claude install` (Claude Code), `npx --yes @primitive.ai/prim cursor install` (Cursor), `npx --yes @primitive.ai/prim codex install` (Codex), or `npx --yes @primitive.ai/prim hermes install` (Hermes). An inactive repo is not passively captured. Deliberately record higher-order forks in the road through `prim decisions create` as described below. Your job is also to **read** the graph before load-bearing edits and **answer** the occasional rationale confirmation. (Responding to Conflict Gates applies only once Enforcement is enabled — see below.)
 
 ## Auth
 
@@ -36,10 +36,10 @@ The CLI auto-refreshes a still-valid session from the stored refresh token (proa
 
 **Conflict Gates** check each edit against the decision graph before it lands and surface any load-bearing decision it conflicts with. When **Enforcement** is enabled for an organization, a conflicting edit can be warned, blocked, or paused for confirmation until you reconcile the decision and retry. Automatic decision capture in active repos (above) runs independently.
 
-When Enforcement is enabled, before an edit (Claude Code: Edit/Write/MultiEdit; Codex: apply_patch; Hermes: write_file/patch) a PreToolUse hook scores the target file against the graph:
+When Enforcement is enabled, before an edit (Claude Code: Edit/Write/MultiEdit; Cursor: Write/Delete/Shell; Codex: apply_patch; Hermes: write_file/patch) a PreToolUse hook scores the target file against the graph:
 
 - **deny** -- the edit is blocked: it conflicts with a load-bearing prior decision. Don't fight it. Read the reason line; it names the decision id. If you genuinely intend to override that decision, run `npx --yes @primitive.ai/prim reconcile dec_<shortId>`, then retry the edit once. Otherwise choose an approach that respects the decision.
-- **warn / additional context** -- the edit proceeds, but a relevant prior decision is surfaced. Read it. Codex cannot pause mid-tool, so a would-be `ask` arrives as a **deny** carrying the reconcile directive. Hermes also has no soft-confirm tier, so a would-be `ask` arrives as a **deny** carrying the same directive: reconcile and retry, or set `PRIM_HOOK_MODE=warn` to downgrade it to context-only.
+- **warn / additional context** -- the edit proceeds, but a relevant prior decision is surfaced. Read it. Cursor, Codex, and Hermes cannot pause at this gate, so a would-be `ask` arrives as a **deny** carrying the reconcile directive: reconcile and retry, or set `PRIM_HOOK_MODE=warn` to downgrade it to context-only.
 - **"decision check skipped / not verified" or "... partial / truncated"** -- the check could not fully run. Treat constraints as UNKNOWN, not clear; never read silence as approval.
 
 When enabled, the gate sends its preflight check directly to Primitive and fail-opens on infrastructure errors (for example, a network failure or org-unbound token) -- a service problem would not block your edit. That is why an "unavailable" note matters: it is the honest signal that the check, not your edit, is what failed.
@@ -47,7 +47,7 @@ When enabled, the gate sends its preflight check directly to Primitive and fail-
 ## Read the graph before large or load-bearing edits
 
 - `npx --yes @primitive.ai/prim decisions check --files "src/a.ts,src/b.ts"` -- which active decisions reference the files you're about to touch (comma-separated paths, one `--files` value). Run it before a big change.
-- `npx --yes @primitive.ai/prim decisions recent` -- the team's recent decisions, each row badged by author and agent (`Your Claude Code` / `Your Codex` / `Your Hermes`); `--limit <n>` and `--since <dur>` narrow it. `--author "<name>"` filters to one teammate (feed name, `"First Last"`, last name, username, email, or email local-part) -- the way to answer "what has X decided?"; an unknown or ambiguous name comes back as `unavailable` with the reason, and `authorHasDecisions` in the JSON distinguishes "no feed-visible decisions" (false) from "has decisions, none in this window" (true). The page defaults to the 10 most recent, so it can hide older ones: on an author query the JSON's `windowTotal` is how many that teammate has in the window. When `windowTotal` exceeds the rows returned, don't present the page as complete -- tell the user you're showing the most recent N of `windowTotal` and offer to pull the rest, which is a re-run with `--limit <windowTotal>` (capped at 100; `windowTotalCapped` means the count is a floor rendered `N+`, and a window past 100 can't be fetched whole).
+- `npx --yes @primitive.ai/prim decisions recent` -- the team's recent decisions, each row badged by author and agent (`Your Claude Code` / `Your Cursor` / `Your Codex` / `Your Hermes`); `--limit <n>` and `--since <dur>` narrow it. `--author "<name>"` filters to one teammate (feed name, `"First Last"`, last name, username, email, or email local-part) -- the way to answer "what has X decided?"; an unknown or ambiguous name comes back as `unavailable` with the reason, and `authorHasDecisions` in the JSON distinguishes "no feed-visible decisions" (false) from "has decisions, none in this window" (true). The page defaults to the 10 most recent, so it can hide older ones: on an author query the JSON's `windowTotal` is how many that teammate has in the window. When `windowTotal` exceeds the rows returned, don't present the page as complete -- tell the user you're showing the most recent N of `windowTotal` and offer to pull the rest, which is a re-run with `--limit <windowTotal>` (capped at 100; `windowTotalCapped` means the count is a floor rendered `N+`, and a window past 100 can't be fetched whole).
 - `npx --yes @primitive.ai/prim decisions show <idOrShortId>` and `npx --yes @primitive.ai/prim decisions cascade <idOrShortId>` -- full detail, and the downstream blast radius a change would disturb.
 
 Before presenting decision reads, use the current conversation, task, and available memory to understand what matters to the requester and how they want it delivered. Do not dump the API's chronological rows unchanged. Group related decisions around goals or workstreams, lead with decisions that affect the requester's current work, and surface conflicts, supersessions, or consequential tradeoffs before background activity. Explain why an item is relevant when the connection is supported; do not invent relevance or rationale. Preserve the response's `unavailable`, truncation, and count semantics when reshaping it.
@@ -237,7 +237,7 @@ npx --yes @primitive.ai/prim hooks uninstall
 
 Under `CI=1` (or with `--non-interactive`), `hooks install` fails fast in a Husky repo unless `--yes` or `--target` is set; the error names both escapes. `hooks uninstall` only removes the `.git/hooks` copies — if a hook was installed into `.husky/`, remove the prim block from that file manually. To suppress the hooks for one commit, use `git commit --no-verify`.
 
-These git hooks are separate from the **session hooks** (`claude install` / `codex install` / `hermes install`) that drive in-session capture (and Conflict Gates, when enabled).
+These git hooks are separate from the **session hooks** (`claude install` / `cursor install` / `codex install` / `hermes install`) that drive in-session capture (and Conflict Gates, when enabled).
 
 ## Output formats
 
@@ -259,7 +259,7 @@ Examples:
 - **An "unavailable" / "not verified" decision check is not an all-clear.** Treat constraints as UNKNOWN and proceed deliberately; never read the silence as approval — the same holds for Conflict Gates once enabled.
 - **When Enforcement is enabled, a `deny` means a real prior decision conflicts.** Reconcile only when you genuinely intend to override it; otherwise pick an approach that respects it.
 - **Reconcile bypasses are single-use and short-lived.** One bypass clears your *next* edit to the governed file; it is not a standing override.
-- **Capture of coding activity is automatic only in active repos; never inject moves by hand.** If an active repo's decisions aren't showing up, check that the session hooks are installed (`claude status` / `codex status` / `hermes status`) and the daemon is running. Deliberately authoring a higher-order fork—either at the user's request or under the conservative proactive policy above—is a separate, supported path through `decisions create`; in an inactive repo, honor the per-invocation approval boundary above.
+- **Capture of coding activity is automatic only in active repos; never inject moves by hand.** If an active repo's decisions aren't showing up, check that the session hooks are installed (`claude status` / `cursor status` / `codex status` / `hermes status`) and the daemon is running. Deliberately authoring a higher-order fork—either at the user's request or under the conservative proactive policy above—is a separate, supported path through `decisions create`; in an inactive repo, honor the per-invocation approval boundary above.
 - **Don't fabricate rationale on a confirmation.** If you don't know why a decision was made, say so rather than guessing.
 
 ## After each task