pushing to the limit

.gitignore

@@ -1,5 +1,6 @@
 /target
 /doc/tags
+*.log
 loopbioticd-v*.tar.gz
 loopbioticd-v*.tar.gz.sha256
 .luarc.local.json

AGENTS.md

@@ -0,0 +1,130 @@
+# Repository instructions for agents
+
+## Required product contracts
+
+Before changing this repository, read these three files in full:
+
+1. [`doc/ui.md`](doc/ui.md) — visible surfaces, hierarchy, layout, responsive
+   behavior, and theming.
+2. [`doc/interactions.md`](doc/interactions.md) — state transitions, focus,
+   navigation, actions, review safety, and recovery.
+3. [`doc/feeling.md`](doc/feeling.md) — product character, voice, pacing, trust,
+   and the intended experience.
+
+They are separate views of one product and are mandatory context, including for
+backend or protocol work that can surface new state, progress, errors, cards, or
+actions in Neovim.
+
+## Non-negotiable UI model
+
+Loopbiotic has exactly two user-facing product Windows:
+
+- **PromptWindow** owns user intent, the visible turn mode, prompt text, and visible attached
+  context.
+- **AgentWindow** owns all agent progress, responses, and Widgets.
+
+Use the vocabulary from `doc/ui.md`: a technical Neovim float or buffer may be a
+Frame; content inside a Window may be a View or Widget. Frames, Views, pickers,
+progress states, file choosers, and Widgets do not create additional product
+Windows.
+
+Do not introduce a third product Window. AgentWindow does not render general
+next-intent actions. Its deliberate exception is patch review: the agent comment
+and `Accept` / `Reject` remain the explicit source-mutation gate. `Accept` may
+continue an already authorized Goal; `Reject` pauses it and opens PromptWindow
+without running the model. A Widget may change local representation or pending
+prompt context, but only an explicit PromptWindow submission may introduce new
+intent.
+
+There is no automatic intent-routing mode. The mode visible in PromptWindow is
+the user-selected response contract: `fix`/`propose` require Patch,
+`explain`/`review` require Finding, and `investigate` requires Hypothesis. Never
+infer or replace the selected mode from prompt wording. Every patch remains inert
+behind `Accept` / `Reject`.
+
+Every PromptWindow, including Reply, must visibly own one valid turn mode.
+`keymaps.modes` opens the local picker; submit snapshots and transmits that exact
+mode through both `session/start` and `session/reply`. Never add a mode control
+to AgentWindow or contact the backend when the picker selection changes.
+
+## Living-contract rule
+
+The three product documents must self-heal with every change.
+
+- At task start, compare the request and current implementation with all three
+  contracts. Identify which statements and invariants are affected.
+- During implementation, treat explicit user direction as the intended delta and
+  preserve unrelated invariants. Do not redesign adjacent surfaces implicitly.
+- Before finishing, reconcile all three documents against the resulting behavior,
+  even when only one appears directly affected.
+- Update affected text in the same patch as code and tests. Rewrite current-state
+  sections in place; do not add design changelog entries.
+- Normative sections describe the approved product. Current contradictions are
+  allowed only when they are explicit under `Known implementation gaps` in the
+  relevant contract. Update or remove those gaps as implementation moves.
+- Remove obsolete descriptions. Do not present a known gap as shipped canonical
+  behavior and do not leave implemented behavior undocumented.
+- When runtime code and a descriptive statement disagree, code is evidence of
+  current behavior, not automatic authority for product intent. Resolve the
+  conflict using the user's request, protected invariants, tests, and surrounding
+  design; then repair every stale side.
+- A documentation-only task still requires checking cross-document consistency.
+  A backend-only task may leave the files unchanged only after verifying that no
+  user-visible state, wording, timing, action, or trust boundary changed.
+
+## Change vocabulary
+
+Use these scope words consistently in plans and handoffs:
+
+- **Tune:** preserve the concept and adjust a property.
+- **Rebuild:** the named surface or flow may change structure.
+- **Unify:** apply an existing project pattern to another location.
+- **Explore:** compare alternatives; do not silently make one the contract.
+- **Local:** change only the named component or state.
+- **Systemic:** change the shared rule and every affected consumer.
+
+For product-facing changes, be able to state:
+
+```text
+Scope:
+Intent:
+Preserve:
+UI delta:
+Interaction delta:
+Feeling delta:
+States checked:
+Acceptance condition:
+```
+
+This is a reasoning checklist, not a required user-facing form.
+
+## Product change definition of done
+
+A product-facing change is complete only when:
+
+- implementation and tests match the normative contract, or every remaining
+  mismatch is explicitly and accurately tracked as a known implementation gap;
+- loading, success, empty, error, hidden, disabled, and narrow-layout states were
+  considered where relevant;
+- visible shortcut labels match actual bindings;
+- focus, navigation, cancellation, retry, and source-mutation boundaries remain
+  explicit;
+- every agent state reuses AgentWindow; new intent originates in PromptWindow,
+  while `Accept` only continues an already authorized Goal;
+- AgentWindow tab ownership, visible/wrapped state, off-tab behavior, and `pr`
+  restoration were preserved or deliberately updated;
+- Widget payloads are registered, versioned, schema-validated, provenance-aware,
+  and limited to allowlisted local intents;
+- Widget-selected context is visible and removable before prompt submission;
+- patch review keeps the agent comment, diff, and explicit `Accept` / `Reject`;
+  rejection performs no model turn and transfers focus through PromptWindow back
+  to AgentWindow when the prompt is closed;
+- new-file and directory proposals are workspace-bound, collision-checked, and
+  reviewed through Netrw when available without treating it as a product Window;
+- new runtime copy matches the voice in `doc/feeling.md`;
+- documentation describes the resulting current state rather than the edit that
+  produced it.
+
+Keep implementation pointers in the contracts at file/module granularity. Avoid
+line-number references and duplicated internal details that cannot be maintained
+reliably.

CHANGELOG.md

@@ -8,6 +8,28 @@ The project follows [Semantic Versioning](https://semver.org/).
 
 ### Fixed
 
+- Model output that is not parseable as a Loopbiotic op (for example JSON with
+  an unescaped quote inside a string value) is now repaired through the normal
+  retry loop: the model is asked to re-emit one strict JSON object before the
+  error card is shown. The error card with the raw output still appears after
+  three failed attempts.
+- The skills picker no longer errors on open; `q`/`<Esc>` cancel it and restore
+  the previous selection as the footer promises.
+- The PromptWindow title now names the model of the phase the current mode
+  runs: discovery modes show the discovery model instead of the patch model,
+  and an explicit model change updates the title immediately.
+- The OpenAI-compatible backend now runs each turn under the shared
+  `LOOPBIOTIC_TURN_TIMEOUT_SECS` deadline, no longer leaks cancellation state
+  when a speculative turn is aborted, caps workspace file reads, and surfaces
+  the raw model output when a local model answers in prose.
+- Empty or malformed model-preference state now supplies both patch and
+  discovery model maps instead of crashing backend setup.
+- OpenAI-compatible local models now return the same typed patch hunks as
+  Codex. Rust renders and validates the unified diff, so weak models no longer
+  have to reproduce fragile diff syntax themselves.
+- The Codex model picker now reads the authenticated app-server `model/list`
+  catalog and its default model. A discovery-only model is no longer presented
+  as the entire set of selectable patch models.
 - 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
@@ -49,10 +71,13 @@ The project follows [Semantic Versioning](https://semver.org/).
 
 ### 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.
+- The real-model A/B runner flushes each result to JSONL immediately, preserving
+  completed samples when a long local benchmark is interrupted.
+- Every Prompt and Reply carries a visible user-selected mode. Automatic intent
+  routing was removed: fix/propose require Patch, explain/review require Finding,
+  and investigate requires Hypothesis. Slash-prefixed text no longer overrides
+  that visible mode. Persistent multi-step goal execution remains explicit, and
+  every patch stays behind local Accept/Reject review.
 - 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.
@@ -69,6 +94,30 @@ The project follows [Semantic Versioning](https://semver.org/).
 
 ### Added
 
+- The LM Studio/OpenAI-compatible backend now uses streaming Responses with
+  provider-stored session chains, optional private reasoning, real cancellation,
+  and Rust-owned bounded workspace read/search/list tools. Tools are available
+  only for discovery and explicit Goals, never execute commands or mutate
+  source, and remain independent of MCP and instruction Skills.
+- A controlled real-model Project Intelligence A/B runner now compares the
+  feature-disabled baseline, marker-derived ProjectProfile, and ProjectProfile
+  plus selected Skills. It includes Angular 22, TypeScript 6, React, Nx, and
+  Rust fixtures, deterministic rubrics, token/latency/retry telemetry, and an
+  OpenAI-compatible LM Studio backend for safe sequential local-model runs.
+- PromptWindow now has a session-scoped Markdown Skills multiselect. Configured
+  files such as `AGENTS.md` autoload as locked inert instructions; optional root
+  Markdown files are explicitly selected, remain visible through Reply, and are
+  content-addressed and bounded before reaching the backend. A deterministic
+  ProjectProfile is built in Rust by marker-activated technology adapters and
+  supplies exact lockfile versions, Nx/Cargo workspace areas and dependencies,
+  commands, Compose infrastructure, and active Neovim LSP capabilities without
+  MCP or a model discovery turn. Protocol version is now 12.
+- Prompts resolve a session-pinned static Flow graph through LSP call hierarchy
+  without opening UI or changing geometry. The explorer is an explicit `F`
+  toggle with lazy caller/callee expansion, exact call-site/reference
+  navigation and responsive split/single-pane layouts. Callstack answers carry
+  a structured `flow_path` and render it directly in the card UI. The bounded
+  normalized graph is sent to every backend without agent-side rediscovery.
 - 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
@@ -87,7 +136,7 @@ The project follows [Semantic Versioning](https://semver.org/).
   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.
+  tags).
 - 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