FEATURES

Cargo.toml

@@ -20,6 +20,7 @@ rust-version = "1.85"
 [workspace.dependencies]
 anyhow = "1"
 async-trait = "0.1"
+reqwest = { version = "0.12", default-features = false, features = ["json"] }
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 tempfile = "3"

README.md

@@ -15,8 +15,9 @@ The editor experience stays the same.
 ## Status and compatibility
 
 Pairagen is beta software. It has been developed and tested primarily with the
-Codex CLI app-server backend. Generic CLI, stdio agent, and local model adapters
-are available, but currently receive less real-world testing than Codex.
+Codex CLI app-server backend. Persistent Claude CLI (stream-json), Ollama HTTP,
+generic CLI, and stdio agent adapters are available, but currently receive less
+real-world testing than Codex.
 
 Requirements:
 
@@ -36,6 +37,9 @@ Implemented capabilities include:
 - patch gate
 - mock backend
 - generic CLI backend
+- persistent Claude CLI backend (one stream-json process per session)
+- Ollama HTTP backend for local models (model stays loaded, JSON-forced output)
+- structured agent denial (`deny` op) rendered as a distinct card
 - deterministic token-budgeted project context with LSP hints and dependency ranking
 
 ## Installation
@@ -170,6 +174,21 @@ Cards stay anchored beside the source line and do not take focus. Use `<leader>p
 to jump to a finding or the first line of an inline draft, and `<leader>pr` to
 focus the current Pair card.
 
+When the agent can only proceed from a different file — say the fix belongs in
+a component class while a template is open — it does not fail or deny. It asks
+mid-turn: Pair shows a permission prompt ("Pair agent wants to open …"), and on
+approval the file opens and the same agent turn continues with the new buffer.
+One request, no retry. Declining (or ignoring for two minutes) surfaces a deny
+card with the location and an `[o] Open & retry` shortcut instead.
+
+By default the first card is whatever fits the prompt best: a hypothesis, a
+finding, or a clarifying choice when the prompt is ambiguous. Start the prompt
+with `/{kind}` to demand a specific card instead — `/hypothesis`, `/finding`,
+`/patch` (alias `/fix`), `/choice`, or `/summary`. For example
+`/patch guard the payload here` skips discovery and drafts a patch directly.
+Unknown words after `/` are treated as normal prompt text, so paths like
+`/tmp/project` are safe.
+
 The goal and accepted-step count stay visible on cards and editable drafts. After
 accepting a patch, Pair shows a local receipt without calling the agent again.
 `Next` explicitly continues the same goal and must return either one local patch

doc/tags

@@ -0,0 +1,22 @@
+:Pair	pairagen.txt	/*:Pair*
+:PairAgent	pairagen.txt	/*:PairAgent*
+:PairFix	pairagen.txt	/*:PairFix*
+:PairFollow	pairagen.txt	/*:PairFollow*
+:PairHide	pairagen.txt	/*:PairHide*
+:PairLog	pairagen.txt	/*:PairLog*
+:PairLogClear	pairagen.txt	/*:PairLogClear*
+:PairModel	pairagen.txt	/*:PairModel*
+:PairNext	pairagen.txt	/*:PairNext*
+:PairOther	pairagen.txt	/*:PairOther*
+:PairReply	pairagen.txt	/*:PairReply*
+:PairReset	pairagen.txt	/*:PairReset*
+:PairResume	pairagen.txt	/*:PairResume*
+:PairStop	pairagen.txt	/*:PairStop*
+:PairWhy	pairagen.txt	/*:PairWhy*
+pairagen	pairagen.txt	/*pairagen*
+pairagen-commands	pairagen.txt	/*pairagen-commands*
+pairagen-contents	pairagen.txt	/*pairagen-contents*
+pairagen-health	pairagen.txt	/*pairagen-health*
+pairagen-privacy	pairagen.txt	/*pairagen-privacy*
+pairagen-setup	pairagen.txt	/*pairagen-setup*
+pairagen.txt	pairagen.txt	/*pairagen.txt*

lua/pair/card.lua

@@ -96,8 +96,16 @@ function M.lines(card)
     M.add(lines, card.summary or card.title)
   elseif card.kind == "error" then
     M.add(lines, card.message or card.title)
+  elseif card.kind == "deny" then
+    table.insert(lines, "Agent could not proceed")
+    table.insert(lines, "")
+    M.add(lines, card.reason or card.title)
   elseif card.kind == "choice" then
     M.add(lines, card.question or card.title)
+    for index, option in ipairs(card.options or {}) do
+      table.insert(lines, "")
+      M.add(lines, string.format("%d. %s", index, option.label or option.id or ""))
+    end
   end
 
   local location = M.location(card)
@@ -247,6 +255,10 @@ function M.actions(card)
     table.insert(parts, M.hint(config.values.keymaps.go_to, "Go to line"))
   end
 
+  if card.kind == "deny" and type(card.location) == "table" then
+    table.insert(parts, M.hint("o", "Open & retry"))
+  end
+
   for _, action in ipairs(actions) do
     local name = type(action) == "table" and "apply_patch" or action
     local label = labels[name]
@@ -293,6 +305,12 @@ function M.bind(buf, card)
     require("pair").go_to()
   end, { buffer = buf, nowait = true, silent = true })
 
+  if card.kind == "deny" and type(card.location) == "table" then
+    vim.keymap.set("n", "o", function()
+      require("pair").open_and_retry()
+    end, { buffer = buf, nowait = true, silent = true })
+  end
+
   for _, action in ipairs(actions) do
     local name = type(action) == "table" and "apply" or action
     local label = labels[name]
@@ -327,6 +345,7 @@ function M.highlight(buf, lines, card)
       or line:match("^At  ") and "PairMuted"
       or line:match("tokens$") and "PairMuted"
       or card.kind == "error" and "DiagnosticError"
+      or card.kind == "deny" and line == "Agent could not proceed" and "DiagnosticError"
     if group then
       vim.api.nvim_buf_add_highlight(buf, -1, group, index - 1, 0, -1)
     end

lua/pair/config.lua

@@ -6,6 +6,9 @@ M.values = {
     args = {},
     mode = "auto",
     agent = "mock",
+    -- Speculative prefetch of the likely next card: "fix" requests the patch
+    -- in the background while you read a discovery card; "off" disables it.
+    prefetch = "fix",
   },
   distribution = {
     repository = "DorianDevp/pairagen",
@@ -36,19 +39,24 @@ M.values = {
       args = { "dev", "stdio-agent" },
     },
     claude = {
-      kind = "generic",
+      kind = "claude_app",
       command = "claude",
       args = {},
+      -- Discovery cards (hypothesis/finding/choice) run on a faster model
+      -- with a capped thinking budget; patch drafting keeps the main model
+      -- with adaptive thinking. Set either to nil to use the CLI default.
+      discovery_model = "haiku",
+      discovery_thinking = 1024,
     },
     aider = {
       kind = "generic",
       command = "aider",
       args = {},
     },
     ["local"] = {
-      kind = "generic",
-      command = "ollama",
-      args = { "run", "qwen2.5-coder:7b" },
+      kind = "ollama",
+      model = "qwen2.5-coder:7b",
+      host = "http://127.0.0.1:11434",
     },
   },
   keymaps = {
@@ -285,7 +293,13 @@ end
 
 function M.backend_env()
   local _, agent = M.agent_config()
+  local env = M.agent_env(agent)
+  env.PAIR_PREFETCH = M.values.backend.prefetch or "fix"
+
+  return env
+end
 
+function M.agent_env(agent)
   if agent.kind == "mock" then
     return {
       PAIR_BACKEND = "mock",
@@ -316,6 +330,31 @@ function M.backend_env()
     }
   end
 
+  if agent.kind == "claude_app" then
+    local args = vim.deepcopy(agent.args or {})
+
+    return {
+      PAIR_BACKEND = "claude_app",
+      PAIR_CLAUDE_COMMAND = agent.command,
+      PAIR_CLAUDE_ARGS = table.concat(args, " "),
+      PAIR_CLAUDE_ARGS_JSON = vim.json.encode(args),
+      PAIR_CLAUDE_MODEL = agent.model or "",
+      PAIR_CLAUDE_DISCOVERY_MODEL = agent.discovery_model or "",
+      PAIR_CLAUDE_DISCOVERY_THINKING = agent.discovery_thinking
+          and tostring(agent.discovery_thinking)
+        or "",
+    }
+  end
+
+  if agent.kind == "ollama" then
+    return {
+      PAIR_BACKEND = "ollama",
+      PAIR_OLLAMA_MODEL = agent.model or "",
+      PAIR_OLLAMA_HOST = agent.host or "",
+      PAIR_OLLAMA_KEEP_ALIVE = agent.keep_alive or "",
+    }
+  end
+
   local args = M.agent_args(agent)
 
   return {

lua/pair/init.lua

@@ -16,6 +16,26 @@ rpc.on("agent/progress", function(progress)
   thinking.progress(progress)
 end)
 
+-- Mid-turn permission request: the agent can only continue once a different
+-- file is open. On approval the same agent turn resumes with fresh context —
+-- no card is shown and no extra request is spent.
+rpc.on_request("editor/open_location", function(params, respond)
+  local location = params.location or {}
+  local file = location.file or "?"
+  local question = string.format(
+    "Pair agent wants to open %s:%s\n%s",
+    vim.fn.fnamemodify(file, ":~:."),
+    location.line or 1,
+    params.reason or ""
+  )
+
+  if vim.fn.confirm(question, "&Open\n&Deny", 1, "Question") == 1 and navigation.open_location(location) then
+    respond({ granted = true, context = context.session() })
+  else
+    respond({ granted = false })
+  end
+end)
+
 function M.setup(opts)
   config.setup(opts)
   require("pair.commands").setup()
@@ -81,6 +101,7 @@ function M.start(text, mode, source)
     state.goal = message.result.goal or state.goal
     state.token_usage = message.result.token_usage
     state.turn_token_usage = message.result.turn_token_usage
+    state.backend_model = message.result.model or state.backend_model
     state.context_report = message.result.context_report
     log.event("context_optimization", message.result.context_report or {})
     log.event("agent_attempts", message.result.attempts or {})
@@ -148,6 +169,7 @@ function M.action(action)
 
     state.token_usage = message.result.token_usage
     state.turn_token_usage = message.result.turn_token_usage
+    state.backend_model = message.result.model or state.backend_model
     state.context_report = message.result.context_report
     log.event("context_optimization", message.result.context_report or {})
     log.event("agent_attempts", message.result.attempts or {})
@@ -212,6 +234,7 @@ function M.reply(text)
 
     state.token_usage = message.result.token_usage
     state.turn_token_usage = message.result.turn_token_usage
+    state.backend_model = message.result.model or state.backend_model
     state.context_report = message.result.context_report
     log.event("context_optimization", message.result.context_report or {})
     log.event("agent_attempts", message.result.attempts or {})
@@ -246,6 +269,25 @@ function M.resume()
   ui.notify("No Pair card to restore", vim.log.levels.WARN)
 end
 
+-- One-key continuation for a deny card that names a location: jump there, so
+-- the next context capture sees that buffer, then retry the denied step.
+function M.open_and_retry()
+  local active_card = state.card
+  if not (active_card and active_card.kind == "deny" and type(active_card.location) == "table") then
+    ui.notify("No location on this card", vim.log.levels.WARN)
+    return
+  end
+
+  if not navigation.open_location(active_card.location) then
+    ui.notify("Could not open " .. tostring(active_card.location.file), vim.log.levels.ERROR)
+    return
+  end
+
+  ui.close(state.card_win)
+  state.card_win = nil
+  M.action("retry")
+end
+
 function M.go_to()
   if state.card and state.card.kind == "patch" and require("pair.diff").focus_change() then
     return
@@ -313,6 +355,7 @@ function M.reset()
   state.diff_first_row = nil
   state.token_usage = nil
   state.turn_token_usage = nil
+  state.backend_model = nil
   state.context_report = nil
   state.workspace_hints = nil
   state.completion_notified_card = nil

lua/pair/prompt.lua

@@ -7,9 +7,13 @@ local M = {}
 function M.open(mode)
   local source = require("pair.context").capture()
 
+  -- Let the backend pay its startup cost (CLI boot, process spawn) while the
+  -- user is still typing the prompt.
+  require("pair.rpc").request("backend/warmup", {}, function() end)
+
   M.open_for({
     title = M.title("Prompt"),
-    footer = " Ctrl-s submit  Esc normal  q close ",
+    footer = " /kind forces card type  Ctrl-s submit  Esc normal  q close ",
     submit = function(text)
       require("pair").start(text, mode, source)
     end,
@@ -29,6 +33,11 @@ end
 function M.title(kind)
   local agent = config.agent()
   local model = config.model()
+  if not model or model == "" then
+    -- No model configured: show what the backend reported it is actually
+    -- using, once a turn has completed.
+    model = state.backend_model
+  end
   local active = model and model ~= "" and (agent .. " / " .. model) or (agent .. " / default")
 
   return string.format(" Pair %s · %s ", kind, active)

lua/pair/rpc.lua

@@ -211,6 +211,24 @@ function M.on(method, callback)
   M.notifications[method] = callback
 end
 
+-- Handlers for requests initiated by paird (method + id). The handler
+-- receives (params, respond); it must call respond(result) exactly once.
+function M.on_request(method, callback)
+  M.requests = M.requests or {}
+  M.requests[method] = callback
+end
+
+function M.respond(id, result)
+  local payload = vim.json.encode({
+    jsonrpc = "2.0",
+    id = id,
+    result = result,
+  })
+
+  log.event("rpc_server_response", { id = id })
+  vim.fn.chansend(M.job, payload .. "\n")
+end
+
 function M.on_data(data)
   if not data or #data == 0 then
     return
@@ -249,6 +267,31 @@ function M.handle(line)
     return
   end
 
+  if message.method and message.id ~= nil then
+    local handler = M.requests and M.requests[message.method]
+    local id = message.id
+
+    if not handler then
+      vim.fn.chansend(
+        M.job,
+        vim.json.encode({
+          jsonrpc = "2.0",
+          id = id,
+          error = { code = -32601, message = "unknown editor request " .. message.method },
+        }) .. "\n"
+      )
+      return
+    end
+
+    vim.schedule(function()
+      handler(message.params or {}, function(result)
+        M.respond(id, result)
+      end)
+    end)
+
+    return
+  end
+
   if message.id then
     local callback = M.pending[message.id]
     M.pending[message.id] = nil

lua/pair/version.lua

@@ -1,4 +1,4 @@
 return {
   plugin = "0.1.0",
-  protocol = 5,
+  protocol = 6,
 }

rust/crates/pair_backends/Cargo.toml

@@ -12,6 +12,7 @@ publish = false
 anyhow.workspace = true
 async-trait.workspace = true
 pair_protocol = { path = "../pair_protocol" }
+reqwest.workspace = true
 serde.workspace = true
 serde_json.workspace = true
 tokio.workspace = true