hit the AI limit

.github/workflows/ci.yml

@@ -2,7 +2,7 @@ name: CI
 
 on:
   push:
-    branches: [main]
+    branches: [master]
   pull_request:
 
 permissions:
@@ -22,6 +22,16 @@ jobs:
       - run: cargo clippy --workspace --all-targets -- -D warnings
       - run: cargo test --workspace --no-fail-fast
 
+  lua-lint:
+    runs-on: ubuntu-24.04
+    steps:
+      - uses: actions/checkout@v6
+      - uses: taiki-e/install-action@v2
+        with:
+          tool: stylua,selene
+      - run: stylua --check lua/ tests/ scripts/headless-smoke.lua
+      - run: selene lua/
+
   neovim:
     runs-on: ubuntu-24.04
     steps:
@@ -30,9 +40,22 @@ jobs:
         with:
           version: stable
           neovim: true
+      - uses: dtolnay/rust-toolchain@stable
+      - uses: Swatinem/rust-cache@v2
+      - name: Run Lua unit tests
+        run: nvim --headless -u NONE -i NONE -l tests/lua/run.lua
       - name: Run headless smoke test
         env:
           XDG_STATE_HOME: ${{ runner.temp }}/state
           XDG_DATA_HOME: ${{ runner.temp }}/data
           XDG_CACHE_HOME: ${{ runner.temp }}/cache
         run: nvim --headless -u NONE -i NONE -l scripts/headless-smoke.lua
+      - name: Build loopbioticd
+        run: cargo build -p loopbioticd
+      - name: Run real daemon round-trip
+        env:
+          XDG_STATE_HOME: ${{ runner.temp }}/state
+          XDG_DATA_HOME: ${{ runner.temp }}/data
+          XDG_CACHE_HOME: ${{ runner.temp }}/cache
+          LOOPBIOTIC_SMOKE_BIN: target/debug/loopbioticd
+        run: nvim --headless -u NONE -i NONE -l scripts/headless-smoke.lua

.gitignore

@@ -1 +1,5 @@
 /target
+/doc/tags
+loopbioticd-v*.tar.gz
+loopbioticd-v*.tar.gz.sha256
+.luarc.local.json

.luarc.json

@@ -0,0 +1,13 @@
+{
+  "$schema": "https://raw.githubusercontent.com/LuaLS/vscode-lua/master/setting/schema.json",
+  "runtime": {
+    "version": "LuaJIT"
+  },
+  "workspace": {
+    "library": ["$VIMRUNTIME/lua", "${3rd}/luv/library"],
+    "checkThirdParty": false
+  },
+  "diagnostics": {
+    "globals": ["vim"]
+  }
+}

.stylua.toml

@@ -0,0 +1,6 @@
+column_width = 120
+line_endings = "Unix"
+indent_type = "Spaces"
+indent_width = 2
+quote_style = "AutoPreferDouble"
+call_parentheses = "Always"

CHANGELOG.md

@@ -6,6 +6,108 @@ The project follows [Semantic Versioning](https://semver.org/).
 
 ## [Unreleased]
 
+### Fixed
+
+- Navigating to a card or draft no longer throws "Cursor position outside
+  buffer" when the target line does not exist yet — for example a patch hunk
+  that appends to the end of a short file, or an agent-supplied location past
+  the end of the buffer. Cursor targets are now clamped to the real buffer.
+- Backend turns now run under a configurable deadline
+  (`LOOPBIOTIC_TURN_TIMEOUT_SECS`, default 600, `0` disables). A wedged agent
+  CLI is killed and respawned on the next turn instead of hanging the session
+  forever; the Ollama HTTP client uses the same deadline.
+- Stopping or switching the backend now fails in-flight requests instead of
+  dropping their callbacks, so the thinking spinner can no longer get stuck.
+- RPC sends to an already-exited backend are dropped and logged instead of
+  throwing inside scheduled callbacks.
+- Managed `loopbioticd` downloads now have connect and transfer time limits,
+  and only the expected binary is extracted from the release archive.
+- Retry exhaustion and goal-batch shape mismatches in the session harness
+  degrade to error cards instead of panicking the daemon.
+- Goal hunks that no longer apply after the buffer changes now offer a
+  one-keypress retry that regenerates only the stale local step.
+- Rejecting a draft now stops locally with explicit Retry/Edit/Stop actions
+  instead of immediately spending another model turn on a replacement.
+- Resuming an already-visible action card now moves focus into its window,
+  and patch action cards register every configured shortcut they display
+  (accept, reject, retry, why, and go-to) in addition to their single-key aliases.
+- Working cards tolerate protocol `null` metadata instead of failing while
+  rendering their token footer. Sending a message while a draft is open now
+  restores the source and abandons the preview before conversation starts;
+  tab navigation never starts from a float, and background-tab action floats
+  are closed only after their tab becomes current, avoiding a Neovim tabline
+  use-after-free.
+- Mechanical model diff wrappers are normalized before validation: CRLF line
+  endings, markdown fences, matching git headers, and unambiguous `./`, `a/`,
+  or `b/` path prefixes. Prose, rename/copy metadata, unmatched fences, and
+  headers naming another file are rejected instead of being silently dropped.
+- Cursor-local editor errors remain explicit model context even when their
+  source line is already in the primary excerpt. Codex receives the diagnostic
+  text on discovery and patch turns, so a distant warning or deprecation no
+  longer displaces the error beside the cursor.
+- Push CI runs again: the workflow triggered on a nonexistent `main` branch.
+
+### Changed
+
+- Auto sessions are conversational-first. Their first response and normal
+  replies cannot return patches or completion summaries; persistent goal
+  execution starts only from the explicit `Goal` action. Sending a message
+  pauses an active goal and answers conversationally.
+- Goal work is limited to one file, one coherent hunk, and 32 changed lines
+  per turn. Only explicit goals may speculate on the next patch; ordinary
+  speculation is read-only post-accept conversation.
+- `initialize` validates the client protocol version when one is supplied and
+  returns a structured error (`-32001`) on mismatch instead of failing later
+  with cryptic errors.
+- The project context scan no longer blocks async worker threads.
+- Backend prompts put static contracts and append-only session history before
+  volatile action and editor context data, preserving the longest reusable
+  provider-cache prefix across turns and sliced goal continuations.
+- Internal restructuring: `engine.rs`, `codex_app.rs`, and the context crate
+  are split into focused modules; duplicated backend and Lua helpers are
+  shared; Lua state reset is defined next to the state it resets.
+
+### Added
+
+- Conversation turns have a 10-second visible-response budget and work turns
+  a 20-second budget. Slow turns yield a focusable `Working` card, continue in
+  the background, and can be interrupted through the real Codex
+  `turn/interrupt` API or by terminating persistent CLI processes. Completion
+  arrives through `agent/turn_ready`; slow-turn timing is logged locally and
+  injected once as compact feedback on the next model turn.
+- Accepting a non-goal patch automatically surfaces a read-only conversational
+  next card, with no intermediate “local step applied” summary. That card is
+  prefetched during review by default; rejecting remains a local decision and
+  never regenerates code.
+- Backend preflight in the prompt window (failures surface before typing;
+  composed prompts survive failed starts), repeated-error escalation with
+  actionable guidance, and a client-side error boundary that preserves the
+  session when a UI callback fails.
+- The `backend/warmup` handshake now reports an explicit identity: the active
+  backend, the concrete model the next turn will use (configured, or resolved
+  from the backend — the Claude CLI announces it at process start, Ollama
+  always knows it), and the models the backend can enumerate (Ollama's local
+  tags). Protocol version is now 10.
+- The prompt window title names the active agent and resolved model (never
+  "default"), refreshing as soon as warmup resolves it, and `Ctrl-l`
+  (`keymaps.models`) opens a model picker fed by the backend-enumerated
+  models, an optional per-agent `models` list, and the last reported model.
+  Selections persist per agent exactly like `:LoopbioticModel`.
+- Identity is phase-aware: the reported model is always the patch-drafting
+  one, and a differing discovery model is shown separately instead of
+  masquerading as the active model. The shipped Codex agent uses
+  `gpt-5.4-mini` at low effort for conversation; Claude uses
+  `discovery_model = "haiku"`. The claude picker offers the CLI aliases
+  `sonnet`, `opus`, and `haiku` since the CLI has no model-listing API.
+- Lua tooling (`stylua`, `selene`, LuaLS config) enforced in CI, headless Lua
+  unit tests for the patch engine and session state, `loopbioticd` JSON-RPC
+  integration tests, session state-machine transition tests, and a real
+  daemon round-trip smoke test.
+- Agent-attempt telemetry includes a closed `violation_class` for contract
+  retries and rejected cards, allowing context mismatches, malformed diffs,
+  wrong files, missing fields, kind mismatches, duplicate steps, and
+  incoherent goal batches to be aggregated without logging patch content.
+
 ## [0.3.2] - 2026-07-14
 
 ### Added

Cargo.toml

@@ -24,6 +24,5 @@ reqwest = { version = "0.12", default-features = false, features = ["json"] }
 serde = { version = "1", features = ["derive"] }
 serde_json = "1"
 tempfile = "3"
-thiserror = "2"
 tokio = { version = "1", features = ["io-util", "macros", "process", "rt-multi-thread", "sync", "time"] }
 uuid = { version = "1", features = ["v4"] }

README.md

@@ -40,7 +40,7 @@ Implemented capabilities include:
 - raw, cached, and non-cached session token usage plus a local error log
 - JSON-RPC over stdio
 - Rust session harness
-- continuous goal state machine with local hunk-by-hunk review
+- conversational-first state machine with explicit goals and local hunk review
 - patch gate
 - mock backend
 - generic CLI backend
@@ -159,22 +159,51 @@ restores it on the next Neovim start. A model explicitly configured in
 `setup()` always takes precedence. `:LoopbioticModel default` clears the stored model
 and returns that agent to its own default.
 
+The prompt window title always names the active agent and the concrete model
+the next turn will use, e.g. `codex / gpt-5.4-mini`. Without a configured
+model it shows the model the backend announces during warmup (or reported
+after the last turn), and `model?` until one is known — it never shows
+`default`. The title always names the patch-drafting model; when an agent
+runs discovery on a different model (the shipped Codex agent uses
+`gpt-5.4-mini` at low effort; Claude pins `discovery_model = "haiku"`), that
+is shown separately, e.g.
+`claude-fable-5 · discovery haiku`. Press `<C-l>` (`keymaps.models`) inside
+the prompt to pick a model from every known candidate: the configured model,
+the models the backend enumerates (Ollama's local tags; claude offers its
+stable CLI aliases `sonnet`, `opus`, `haiku`), an optional `models` list on
+the agent definition, and the model reported by the last turn. Picking sets
+the patch model; `discovery_model` stays as configured. The picked model
+persists per agent exactly like `:LoopbioticModel`; the prompt window and its
+typed text stay open.
+
+```lua
+require("loopbiotic").setup({
+  agents = {
+    ["local"] = {
+      kind = "ollama",
+      host = "http://127.0.0.1:11434",
+      models = { "qwen2.5-coder:7b", "llama3.1:8b" }, -- extra picker candidates
+    },
+  },
+  keymaps = {
+    models = "<C-l>", -- model picker inside the prompt window
+  },
+})
+```
+
 ## Flow
 
 ```text
 <leader>a
 Prompt
-Persistent agent goal
-Agent inspects the project and prepares the complete multi-file change
-One local editable hunk
-Edit the inline draft
-Accept, Reject, edit, message, or ask Why
-Why → explanation → Back to the same pending draft
-Accept → advance to the next queued hunk without calling the agent
-Reject → agent automatically reworks the hunk
-Next file opens automatically inside the workspace without calling the agent
-Next editable hunk
-Repeat until completed-goal summary and local diagnostics check
+Conversational answer, finding, or question — never an implicit patch
+Follow up or reply back-to-back
+Draft → one local editable hunk
+Accept → automatic read-only next card, without an intermediate summary
+Reject → stop locally; Retry only when explicitly requested
+Goal → explicitly authorize a sequence of small reviewed hunks
+While a goal hunk is reviewed, only its next small hunk may be prepared
+Repeat until one final goal summary and local diagnostics check
 ```
 
 Cards stay anchored clear of the source line and do not take focus. Use `<leader>pg`
@@ -185,13 +214,19 @@ and draft explanations stay compact by default; press `z` while the card is
 focused to expand or collapse their full text (`keymaps.details` changes this
 key).
 
-For a goal patch, Loopbiotic validates every returned file against its live editor
-buffer, queues the complete batch, and opens each location only when its hunk is
-ready for review. Navigation and acceptance are local operations and do not
-start another model turn. Automatic navigation is restricted to the current
-workspace; edits are still inert drafts until the user accepts each hunk. If
-the agent could not inspect a required file, `open_location` remains a fallback
-that supplies its buffer in a subsequent turn.
+Conversational turns have a 10-second interaction deadline and work turns have
+a 20-second deadline. Crossing it yields a focusable `Working` card instead of
+holding Neovim; the agent continues in the background, and `Cancel` interrupts
+the real backend turn. Slow-turn timing is recorded in the local trace and a
+compact, content-free instruction is injected into the next turn so the agent
+prioritizes an earlier useful response.
+
+For an explicit goal, each backend turn may return at most one file, one
+coherent hunk, and 32 changed lines, plus a plan of the remaining coherent
+steps. The next hunk may be prepared while the current one is reviewed.
+Accepting normally surfaces it immediately; rejecting cancels it and never
+generates a replacement without `Retry`. Automatic navigation stays inside the
+workspace, and every edit remains an inert draft until accepted.
 
 Loopbiotic moves directly to the evidence for a location-bearing card and to the
 first non-blank character of the first added line for a draft, including drafts
@@ -207,27 +242,22 @@ with `/{kind}` to demand a specific card instead — `/hypothesis`, `/finding`,
 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. In
-the default `auto` mode the agent owns the complete process: it inspects the
-project and prepares the complete change across the required files in one turn.
-Loopbiotic then presents that batch hunk by hunk; `Accept` advances locally without
-another model call, and accepting the final complete hunk closes the goal
-locally. `Reject`, `Retry`, a file the agent could not inspect, or an explicit
-question can return control to the agent. User control is the hunk gate, not a
-repeated discovery/assess/draft ceremony. Asking `Why`
-opens a side conversation about the pending hunk and then returns to that exact
-draft without advancing or replacing it. Explicit
-`/{kind}` prompts and investigate/explain/review modes remain available for
-one-card workflows.
+The goal and accepted-step count stay visible on cards and editable drafts, but
+`auto` is conversational: it never starts goal execution or returns a patch on
+the first turn. Use `Goal` (`<leader>pG` or `:LoopbioticGoal`) to authorize
+persistent execution, and send a normal message at any time to pause it and get
+a conversational answer. Asking `Why` explains the pending hunk and returns to
+that exact draft without advancing or replacing it.
 
 When the goal completes, Loopbiotic automatically checks error-level diagnostics in
 the changed, loaded buffers after a short delay. `Check` repeats the same local
 operation without spending model tokens, saving buffers, or running shell test
 commands.
 
-Speculative patch prefetch is off by default because an unused draft still costs
-a full model turn. Set `backend.prefetch = "fix"` only when that latency/cost
-tradeoff is intentional.
+Ordinary speculation is read-only: by default
+`backend.prefetch = "read_only"` prepares only the conversational card that
+follows an accepted non-goal draft. Set it to `"off"` to disable this. Patch
+speculation is allowed only inside an explicitly active goal.
 
 Cards show raw, cached, and non-cached turn and session usage against
 `backend.token_budget` (50,000 raw tokens by default).
@@ -327,7 +357,7 @@ Choosing `Fix` on a card first moves the source context to the card's
 The patch agent therefore receives the recommended consumer or template instead
 of the file where discovery happened.
 
-The future optional classical-ML ranking design is documented in [`ml.md`](ml.md).
+The future optional classical-ML ranking design is documented in [`doc/ml.md`](doc/ml.md).
 The current implementation does not train or run an ML model.
 
 ## Commands
@@ -336,6 +366,8 @@ The current implementation does not train or run an ML model.
 :Loopbiotic
 :LoopbioticReply
 :LoopbioticFix
+:LoopbioticGoal
+:LoopbioticCancel
 :LoopbioticWhy
 :LoopbioticFollow
 :LoopbioticOther