Skip to content

Changelog

All notable changes to Quoriv will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[1.7.0] — 2026-05-24

Added — /init codebase onboarding

  • /init slash command — Claude-Code-style "understand this repo and write a memory file". Kicks off a canned, systematic exploration: the agent calls write_todos to plan, maps the structure with ls / glob / regex_grep, reads the important files in batches (so DeepAgents' SummarizationMiddleware keeps the context window healthy on large repos rather than trying to read everything at once), then writes a structured PROJECT.md at the repo root with Overview / Tech stack / Structure / Architecture / Conventions / Commands / Notable sections. If a PROJECT.md already exists, it reads and improves it instead of overwriting blind. Because PROJECT.md is auto-loaded into the system prompt every session via DeepAgents' MemoryMiddleware, running /init once makes every future session repo-aware.

[1.6.4] — 2026-05-24

Changed

  • Raised per-turn recursion_limit from 1000 to 10,000 (matching DeepAgents' graph ceiling), so the agent can work like Claude Code on large codebases — read/list hundreds of files, run long multi-step edit-and-verify loops, and chain sub-agent delegations without hitting the cap. Kept finite as a runaway-loop backstop rather than truly unbounded.

[1.6.3] — 2026-05-24

Fixed

  • GRAPH_RECURSION_LIMIT / "Recursion limit of 25 reached" on large-codebase tasks. A turn that reads/lists many files (e.g. "understand my codebase") died after ~12 tool calls because the per-turn RunnableConfig didn't set recursion_limit, so LangGraph fell back to its built-in default of 25. (DeepAgents' compiled graph defaults to 9999, but that default is overridden by whatever's in the config at invoke time.) _drive_turn now sets recursion_limit=1000 explicitly, so multi-step exploration and long agentic loops run to completion.

[1.6.2] — 2026-05-24

Added

  • regex_grep tool — full Python regex search across the repo. DeepAgents' built-in grep is literal-substring only; regex_grep accepts arbitrary regex (def \w+_test, ^class\s+\w+\(.*Base.*\):, etc.), an optional glob filter, case-insensitive flag, and a max_matches cap (default 200). Skips the usual noise dirs (.git, .venv, node_modules, __pycache__, dist, build, target, .mypy_cache, etc.) and binary files (UTF-8 decode error → skip). Registered in QUORIV_TOOLS so the agent picks it up automatically.

Changed

  • PROJECT_PLAN.md status section rewritten to reflect actual state. The doc was stuck at "Phase 1 Slice 1 complete"; reality is all four planned phases shipped + a Phase 5 (post-v1.0 UX polish) underway through v1.6.x. Open housekeeping (domain pick, telemetry vendor, plugin API end-to-end verification) called out explicitly.

[1.6.1] — 2026-05-24

Fixed

  • Slow paste into the input box. complete_while_typing=True fired the slash completer on every character of a pasted blob — a 1000-char paste meant 1000 completer calls + 1000 popup repaints. Now gated by a Condition that only enables completion when the buffer is short (< 80 chars) and starts with /, so pastes are instant and the slash-command typeahead still works for the case it was designed for.
  • Input box stuck at one line. The input Window was hardcoded to height=1, wrap_lines=False so long text scrolled horizontally off the right edge and pasted newlines got flattened. Changed to height=Dimension(min=1), wrap_lines=True, dont_extend_height=True and flipped the buffer to multiline=True. The frame now grows downward as you paste or type more lines, and pasted text retains its line breaks. Enter still submits (our keybinding wins over the default newline-on-Enter behaviour); Esc-Enter still inserts a literal newline.

[1.6.0] — 2026-05-24

Added — Rotating status verb for non-reasoning models

  • "Pondering…" / "Concocting…" / "Incubating…" status row while the model is generating but hasn't started streaming yet. Renders as a single italic-grey line (▸ Verb…) above the input frame, cycling through ~30 whimsical verbs every 3 seconds — Brewing, Cogitating, Concocting, Conjuring, Crystallizing, Distilling, Excogitating, Forging, Gestating, Hatching, Incubating, Marinating, Musing, Percolating, Pondering, Ruminating, Synthesizing, Tinkering, Unspooling, Weaving, Wrangling, etc.
  • Lifecycle: starts at the top of _drive_turn before the first event arrives; stops as soon as real content (text or thinking) shows up via push_chunk / push_thinking; stops on on_tool_start so the tool's own one-line label takes over; restarts on on_tool_end so the user sees activity while the model resumes; final stop on turn end.
  • Doesn't interfere with reasoning models — the ▸ Thinking… block from v1.5.8 supersedes the rotating verb the moment any actual reasoning tokens arrive. Reasoning shows for claude-opus-4-7, o3 / gpt-5.5-pro, gemini-3.1-pro, deepseek-r1 / deepseek-v4-pro, kimi-k2.6; the verb keeps everyone else (gpt-5.4, gpt-5.4-mini, claude-haiku-4-5, gemini-2.5-flash, etc.) feeling alive.

Internal

  • ChatApp.set_status / start_status / stop_status — manage the rotation. start_status spawns an asyncio task that cycles the verb; stop_status cancels and clears. Both are no-ops when there's no running event loop (test-safe).
  • New _status_container Window in the layout, gated by ConditionalContainer on _status_text is not None.
  • push_chunk and push_thinking call stop_status() themselves, so any caller that pushes content automatically suppresses the verb.

[1.5.11] — 2026-05-24

Fixed

  • Silent tool_end for write_file / edit_file. The v1.5.7 generic-tool branch was still printing content='Updated file /foo.py' name='write_file' tool_call_id='…' for write/edit calls after the diff renderer had already shown what changed. Added both to the silent-on-end allowlist alongside read_file / ls / glob / grep.

[1.5.10] — 2026-05-23

CI

  • Apply ruff format to tests/unit/test_thinking_blocks.py (added in v1.5.8). CI lint was failing ruff format --check . on that file; pure cosmetic reflow with no semantic change.

[1.5.9] — 2026-05-23

CI

  • binaries.yml build job: explicit permissions: contents: read. v1.5.8's binaries workflow died on the Windows matrix with fatal: could not read Username for 'https://github.com' — actions/checkout@v4 couldn't attach the runner's GITHUB_TOKEN because the default permission set in our repo configuration didn't include contents: read for that job. Declaring it explicitly fixes the auth.
  • binaries.yml attach-to-release: explicit token: for softprops/action-gh-release@v2, mirroring the v1.5.6 fix in release.yml.

No code changes vs v1.5.8.

[1.5.8] — 2026-05-23

Added — Thinking / reasoning surfacing

  • Streamed reasoning shows above the answer in an italic-grey ▸ Thinking… block. Works for every provider, with three different protocols handled by _chunk_blocks:
  • Structured content blocks — Anthropic Claude (extended thinking type="thinking") and OpenAI Responses API / o-series / gpt-5.x (type="reasoning"). Reasoning blocks go straight to chat_app.push_thinking; text blocks go to the regular stream.
  • additional_kwargs.reasoning_content — DeepSeek (V4, R1) and Kimi (K2.6 thinking) via the langchain-openai wrapper. We read both the reasoning_content and reasoning keys to handle minor build differences.
  • Inline <think>...</think> tags — DeepSeek-R1 in non-Responses-API mode emits reasoning as plain text wrapped in tags. A small state-machine parser (_strip_think_tags) handles tags that span chunk boundaries, including the case where <thi arrives in one chunk and nk> in the next.
  • Collapsed scrollback flush. When finalize_stream commits the answer to scrollback, the full chain-of-thought is not echoed — only a one-liner ▸ Thought for N chars (italic grey) lands above the rendered markdown answer. The thinking stays visible during streaming for live feedback; once the turn is done it stops dominating scrollback. Matches Claude Code's collapsed reasoning view.
  • ChatApp.push_thinking(text) — new method parallel to push_chunk. ChatApp.thinking_buffer exposes the in-flight buffer for tests.

Tests

  • tests/unit/test_thinking_blocks.py (11) — structured type="thinking" (Anthropic), type="reasoning" (OpenAI Responses), additional_kwargs.reasoning_content (DeepSeek / Kimi), and the inline <think> parser including partial-tag-at-chunk-boundary cases.

Test count: 920 → 931 (+11). All gates green.

[1.5.7] — 2026-05-23

Changed — Claude-Code-style tool feedback

  • One-line human labels for every tool call. Replaced the raw → tool_name(arg=value, …) dump with a contextual one-liner — • Reading /a.py, • Listing /, • Searching for 'foo', • Running pytest -q, • Delegating to researcher: <description>, • Fetching <url>, etc. The full args / output JSON the agent passes around no longer floods scrollback.
  • Tool output is silent by default. read_file / ls / glob / grep no longer re-render their results to the chat — the agent quotes the relevant bits in its next message anyway. execute still shows shell output (compact, max 10 lines, non-zero exit codes flagged in red) because that's typically the point of the call. Custom / unknown tools fall back to a short truncated excerpt so failures stay visible.
  • Todo-list rendering for write_todos. DeepAgents' TodoListMiddleware exposes a write_todos tool whose payload is a list of {content, status} entries. Quoriv now renders it as a Plan section with checkbox markers — [x] for completed (green strikethrough), [~] for in-progress (yellow bold), [ ] for pending — so the user can follow the agent's plan as it executes long-running tasks.

Internal

  • render_tool_end now takes an optional name= so the renderer can pick a per-tool strategy. The old call-without-name path still works (treated as a generic tool) for any external callers.
  • tests/unit/core/test_events.py rewritten to assert on the human labels instead of the raw arg dump.

[1.5.6] — 2026-05-23

CI

  • Pass token: ${{ secrets.GITHUB_TOKEN }} explicitly to softprops/action-gh-release@v2. The v1.5.5 PyPI publish succeeded but the follow-up GitHub Release attachment failed with "Bad credentials" — the action's default GITHUB_TOKEN env lookup didn't pick up the token even though permissions: contents: write was set on the job. No code changes vs v1.5.5.

[1.5.5] — 2026-05-23

Changed

  • HITL approval prompt now uses the same inline picker as /login, /mode, /load. The Float-based Dialog modal is replaced by chat_app.prompt_picker with three numbered options:
  • Approve once
  • Approve and remember for this session
  • Reject

The Rich panel showing tool name + args + description still prints to scrollback above the picker (same context, same look), so the only thing that changed is the verdict UX — no more separate floating canvas, no more a / r / A letter shortcuts. Arrow keys + Enter (or digits 1-3) match every other interactive selection in the app.

Removed

  • _render_approval_body_ansi helper in prompts.py (was only used by the Dialog modal — the picker doesn't need the ANSI-rendered body).
  • Unused io.StringIO import in prompts.py (left over from the removed helper).

[1.5.4] — 2026-05-23

Fixed

  • Bottom status line now reflects the model the user just chose. The toolbar closure read the function-local model_id parameter captured at _interactive_loop entry, so it stayed pinned to whatever config.model.default resolved to (typically openai:gpt-4.1) even after /login switched to a new provider/model and printed Ready. openai:gpt-5.4. Moved model_id into the mutable state dict that the toolbar already reads from for thread + permission mode; both the startup-onboarding path and mid-session /login now write the new id there. The toolbar updates on the next redraw — no relaunch needed.

[1.5.3] — 2026-05-23

Changed

  • /mode and /load now open the Claude-Code-style inline picker on bare submit, matching the /login flow. Typing /mode<Enter> shows the four permission modes inline (above the input frame, numbered, arrow-key navigable); typing /load<Enter> shows saved sessions newest-first with their timestamps and short thread IDs. Picking with Up/Down + Enter (or digits 1-9) rewrites the command to /mode <choice> / /load <choice> and dispatches through the existing handler — same live agent rebuild + status-line update logic. Esc cancels and prints Mode unchanged. / Load cancelled..

Removed

  • CompletionsMenu argument popup for /mode and /load. The small autocomplete popover the v1.4.0 SlashCommandCompleter.argument_providers mapping fed was the "small black canvas" the user disliked; with the inline picker handling discoverable selection it's no longer needed. The SlashCommandCompleter still surfaces command-name completions for fast typing, just not argument values.

[1.5.2] — 2026-05-23

Changed

  • Claude-Code-style inline picker for /login (provider + model). Replaced the CompletionsMenu-popup approach with a real inline picker Window that renders a numbered list directly inside the chat area, above the input frame. Title + description sit at the top; selected row is prefixed with >; digit keys 1-9 jump-and-confirm; Up/Down/PageUp/PageDown navigate; Enter selects; Esc cancels. No floating box, no separate "small black canvas" — the picker consumes real layout space the same way the stream window does.

Added

  • ChatApp.prompt_picker(*, title, description, options) — new picker primitive. Uses a ConditionalContainer gated on _picker_active[0] so the picker Window only appears when active, with its own keybindings active only while it has focus.

Removed

  • _make_picker_completer helper in app.py (the per-wizard one-shot Completer that fed the now-replaced prompt_input(open_completion=True) path).

[1.5.1] — 2026-05-23

Fixed

  • Ruff lint failures on main. tests/unit/test_providers.py had two inline import dataclasses / __import__("pytest") calls that tripped PLC0415 (import should be at the top-level of a file). Moved them to the module imports.
  • Pre-existing UP038 violations in observability/trace.py and tools/ast_tools.py — converted three isinstance(x, (A, B)) tuple calls to the union-syntax isinstance(x, A | B) that newer ruff flags by default. These had silently passed CI on older ruff but started failing once ruff>=0.6.0 resolved to 0.12.x.

[1.5.0] — 2026-05-23

Added — Phase 5 Slice 4: provider onboarding wizard

  • /login wizard — three-step interactive flow inside the persistent ChatApp: pick provider → paste API key (masked) → pick model. Each picker is the same inline CompletionsMenu dropdown the v1.4.x slash commands use, so navigation is ↑/↓ + Enter with no modal Float and no terminal-size constraints. The chosen model is persisted to ~/.quoriv/config.toml and the running agent is rebuilt against the new key without restarting.
  • /setup alias — same wizard, runs mid-session. Useful for switching provider or model while keeping the conversation thread alive.
  • /logout — wipes the keychain entry for the current provider (the provider portion of model.default). The saved model.default is left in config.toml so the next /login starts from a known state.
  • Auto-onboarding at startup — if build_agent raises MissingAPIKeyError because no key is configured, run_chat no longer prints "no API key found" and exits. Instead it launches the persistent ChatApp with agent=None and runs the /login wizard as the driver's first action. After the user completes the wizard the agent is built and the session continues normally.
  • New providers, with curated model lists (verified via web search, May 2026):
  • DeepSeek (deepseek:*) — OpenAI-compatible at api.deepseek.com. Models: deepseek-v4-pro (default), deepseek-v4-flash, deepseek-v3.1, deepseek-r1.
  • Kimi / Moonshot (kimi:*) — OpenAI-compatible at api.moonshot.ai/v1. Models: kimi-k2.6 (default), kimi-k2.5.
  • xAI Grok (grok:*) — OpenAI-compatible at api.x.ai/v1. Models: grok-4.3 (default), grok-4.20-non-reasoning.
  • Refreshed model lists for existing providers:
  • OpenAI: gpt-5.5 (default), gpt-5.5-pro, gpt-5.4, gpt-5.4-mini, gpt-5.4-nano.
  • Anthropic: claude-sonnet-4-6 (default), claude-opus-4-7, claude-haiku-4-5-20251001.
  • Gemini: gemini-3.1-pro (default), gemini-3.5-flash, gemini-2.5-pro, gemini-2.5-flash, gemini-2.5-flash-lite.

New modules / APIs

  • src/quoriv/providers.py — ProviderEntry dataclass + PROVIDERS registry + get_provider() / provider_choices() / model_choices() accessors. The single source of truth for "what does the dropdown show."
  • src/quoriv/models/deepseek.py, kimi.py, grok.py — thin ChatOpenAI wrappers (each ~25 LOC) bound to the provider's OpenAI-compatible endpoint via base_url=. Registered in quoriv.models.factory._PROVIDERS.
  • quoriv.config.loader.save_default_model(model_id) — persists [model] default = "<id>" to ~/.quoriv/config.toml, preserving the rest of the file. Hand-rolled minimal TOML emitter avoids adding a tomli-w dependency.
  • ChatApp.prompt_input(completer=None, password=False, open_completion=False) — extended with three new kwargs:
  • completer= swaps the buffer's completer for the duration of the prompt (restored on return) — used by the picker steps.
  • password=True flips a ConditionalProcessor/PasswordProcessor pair on the input control so keystrokes render as * while the underlying text is committed verbatim.
  • open_completion=True schedules Buffer.start_completion(select_first=True) on the next event-loop tick so the dropdown is visible without the user typing first.

Design rationale

  • Static curated list, not models.list() at runtime. The provider API's models.list() returns 60+ IDs for OpenAI alone, mostly dated snapshots like gpt-5-2026-01-08 that users don't want to choose from. A short hand-picked list ships in providers.py and gets refreshed per release. Reasoning is documented in the module docstring.
  • OpenAI-compatible third-parties via base_url=. DeepSeek, Kimi, and xAI all serve the OpenAI Chat Completions API shape, so each provider is a ~25-line ChatOpenAI wrapper instead of a fresh SDK integration. Their keychain entries / env vars are separate (DEEPSEEK_API_KEY, MOONSHOT_API_KEY, XAI_API_KEY) so users can mix and match.
  • Onboarding runs inside the persistent ChatApp, not as a pre-launch CLI prompt. This gives identical UX whether the wizard fires at startup (no key configured) or mid-session via /login — same dropdowns, same key bindings, same scrollback.
  • Async dispatch path for /login / /setup / /logout. The sync _handle_slash dispatcher can't await the wizard, so the driver intercepts these three commands before slash dispatch and runs the async onboarding helper directly.

Tests

  • tests/unit/test_providers.py (7) — registry order preserved, every provider has at least one model, curated defaults spot-checked, unknown lookups return None/[], ProviderEntry is frozen.

Test count: 913 → 920 (+7). All gates green.

[1.4.2] — 2026-05-23

Fixed

  • Inline autocomplete popup actually appears now. The v1.4.0 completer was emitting completions but the layout had no CompletionsMenu widget to render them — the popup never showed. Added a persistent Float(content=CompletionsMenu(max_height=10), xcursor=True, ycursor=True) to the ChatApp's FloatContainer so suggestions render anchored to the buffer cursor, popping upward into terminal scrollback when there isn't room below.
  • No more "Window too small..." on bare /mode. Removed the post-submit select_option_modal dispatch for /mode and /load. The picker rendered as an empty black rectangle / overflow error in the inline Application's tiny reserved area; with the inline completion menu now working, discoverable selection happens through the autocomplete popup instead. Typing /mode (space) opens the menu; arrow keys + Enter pick + submit in one keystroke. Bare /mode<Enter> still falls through to _handle_mode which prints the mode list to scrollback.

select_option_modal is retained on ChatApp as a primitive for future modal interactions but is no longer wired to any slash command.

[1.4.1] — 2026-05-23

Fixed

  • /mode picker no longer renders as "Window too small..." in narrow terminals. The previous Float positioning forced left=2, right=2 which made the picker claim a minimum width of (terminal_cols - 4) — when the mode option labels are wider than that, prompt_toolkit aborts and shows the overflow message. Dropped the right= constraint so the Float sizes to its content, and turned on wrap_lines=True on the body window so long labels fold instead of overflowing. Height dimension was relaxed (min=len(options), no max) so wrapped lines have room to expand.

[1.4.0] — 2026-05-23

Added — Inline argument typeahead (Claude-Code-style)

  • SlashCommandCompleter now supports an argument_providers mapping — pass {cmd: () -> [(value, meta), ...]} and the completer will surface argument completions when the user has typed <cmd> (command + space). The popup is the same inline complete-while-typing autocomplete the user already sees for the command name itself, so the flow is: type /mo → menu shows /mode → Enter → buffer becomes /mode and the menu reopens with read-only / ask / auto / yolo → Down + Enter → buffer becomes /mode auto and the command submits in the same Enter press.
  • ChatApp Enter binding applies-and-submits highlighted completions. When the autocomplete menu has a completion selected via arrow keys, pressing Enter applies it to the buffer first. Completions that end with a trailing space (command names with an arg provider, e.g. /mode) leave the cursor in the buffer so the user can flow into the argument phase; completions without a trailing space (concrete arguments like auto) apply and submit in a single Enter press — no double-tap required.
  • _interactive_loop now wires /mode and /load argument providers into SlashCommandCompleter. The mode provider returns the four ALLOWED_MODES with their descriptions; the load provider returns saved sessions newest-first with timestamp + short thread id metadata.

Changed

  • Command-name completions emit a trailing space only when the command has a registered arg provider. /help, /exit, /clear etc. complete to the bare command so Enter submits immediately; /mode, /load, /save complete to <cmd> (with space) to flow into argument completion.

Tests

  • 8 new tests in tests/unit/ui/test_slash_completer.py covering: arg-provider commands get the trailing space, non-arg commands don't, /mode surfaces all mode values, prefix narrows, meta from provider, case-insensitive matching, unknown commands silent, provider called fresh per completion request.

Test count: 905 → 913 (+8). All gates green.

[1.3.1] — 2026-05-23

Changed

  • Option picker now anchors above the input frame instead of opening as a centred Dialog with a dim backdrop. Switched the select_option_modal overlay from prompt_toolkit.widgets.Dialog (with with_background=True) to a Frame placed in a Float at bottom=4, left=2, right=2 — chat stays visible behind the popup, the popup hugs the input box, and the visual feels like Claude Code's inline dropdown instead of a centred modal. Keys (↑ / ↓ / Enter / Esc), focus handling, and the cancel-returns-None contract are unchanged.

[1.3.0] — 2026-05-23

Added — Phase 5 Slice 3: arrow-key option picker

  • ChatApp.select_option_modal(title, options, current=None) -> str | None — Float-based dropdown picker overlaid on the persistent Application. Keys: ↑ / ↓ (also Home / End / PageUp / PageDown) move the highlight, Enter confirms, Esc / Ctrl+C cancel. Returns the chosen value or None on cancel. The currently active option (e.g. the current permission mode) is pre-highlighted when passed via current=.
  • /mode with no argument opens the mode picker instead of just printing the list. The user arrow-keys to read-only / ask / auto / yolo, hits Enter, and the agent rebuilds against the same checkpointer — same flow as /mode auto typed directly, just discoverable. Cancelling leaves the mode unchanged.
  • /load with no argument opens the saved-session picker when at least one session is saved. Sessions are listed newest-first with name + timestamp + short thread id. Cancelling skips the load. When no sessions exist the original "no saved sessions" hint still prints.
  • Direct argument forms (/mode auto, /load <name>) still work without opening a picker — the new flow is purely additive.

Changed

  • Global Enter / Ctrl+C / Esc-Enter bindings now guard against firing while a modal Float has focus. Without the guard, pressing Enter inside a picker would also submit the (empty) input buffer behind it; pressing Ctrl+C inside a picker would tear the whole Application down.
  • _driver calls _expand_slash_with_picker(...) before slash dispatch — a small helper that detects bare /mode / /load and rewrites the command to its picker-resolved form. The dispatcher (_handle_slash) itself is unchanged.

[1.2.2] — 2026-05-23

Fixed

  • Apply ruff format to the four files added/modified in 1.2.0 / 1.2.1 (src/quoriv/app.py, src/quoriv/ui/chat_app.py, src/quoriv/ui/prompts.py, tests/unit/ui/test_chat_app.py). CI lint was failing ruff format --check .; pure cosmetic reflow with no semantic change.

[1.2.1] — 2026-05-23

Fixed

  • Echo user input into scrollback after submit. With the v1.2.0 persistent Application the input buffer reset on Enter, so the message the user just sent disappeared from screen before the assistant's reply streamed in — leaving the transcript looking like a one-sided monologue. The driver now prints > {message} to scrollback right after prompt_input() returns.
  • Checkpointer crash: 'JsonPlusSerializer' object has no attribute 'dumps'. langgraph-checkpoint-sqlite 2.x calls self.jsonplus_serde.dumps(...) for metadata, but the langgraph-checkpoint 4.x line (transitively pulled by langgraph ≥1.2) only ships dumps_typed / loads_typed. Bumped the dependency to langgraph-checkpoint-sqlite>=3.0,<4.0 — the 3.x line uses json.dumps for metadata and serde.dumps_typed for state, which is compatible with langgraph-checkpoint 4.x.

[1.2.0] — 2026-05-23

Added — Phase 5 Slice 2: persistent prompt_toolkit Application

  • quoriv.ui.chat_app.ChatApp — the chat session is now hosted by a single, persistent prompt_toolkit.Application that lives for the entire quoriv chat invocation, instead of spinning a fresh Application up per input turn. Layout: stream window + bordered input frame + status toolbar, with a FloatContainer ready to host modal overlays. Exposes:
  • await app.run(driver) — starts the Application and runs the chat-driver coroutine as a background task; cleans up on either the driver's return or app.exit().
  • await app.prompt_input() — Future-backed input wait. Enter resolves it with the buffer contents; Ctrl+C / Ctrl+D raise EOFError (matches PromptSession semantics); Esc-Enter inserts a newline for multi-line composition.
  • push_chunk(text) / await finalize_stream() — feed agent stream tokens into the live FormattedTextControl (markdown re-rendered to ANSI on every redraw via Rich) and flush the finished response into terminal scrollback through run_in_terminal. No flicker, no duplicate text.
  • console — Rich Console whose file adapter forwards everything into the Application's scrollback via app.print_text. Welcome banners, slash command help, tool diffs, errors — all the existing console.print(...) call sites land above the persistent layout without modification.
  • await prompt_approval_modal(body_text=...) — installs a modal Dialog Float over the stream + input area, focuses it, and resolves a Future when the user presses a (approve) / r (reject) / A (approve-always) / Esc (cancel).
  • Float-based approval modal — quoriv.ui.prompts.prompt_approval now overlays a modal dialog on the persistent ChatApp instead of opening a separate PromptSession. The Rich-rendered "approval required" panel is captured to ANSI via Rich and embedded as the dialog body. Result: the approval prompt no longer steals terminal ownership from the chat UI — the input frame and toolbar stay visible while the user decides.
  • prompt_toolkit Window-based stream renderer — quoriv.ui.stream.StreamRenderer is now a thin adapter over the ChatApp's stream window. push(text) updates the in-window markdown buffer; await finalize_async() flushes the rendered text into scrollback and clears the buffer atomically inside run_in_terminal. Rich Live is gone — no more "Rich finishes → prompt_toolkit takes over" hand-off between agent output and the next user prompt.

Changed

  • quoriv.app._interactive_loop builds one ChatApp for the whole session and drives it via a _driver coroutine spawned inside chat_app.run(...). Mode / thread / agent state migrate into mutable dicts captured by the driver so live /mode and /clear keep working.
  • quoriv.app._drive_turn, _stream_events, _collect_decisions, and _handle_slash accept an optional chat_app: ChatApp | None keyword and route streaming + approval through it when supplied. Headless code paths (the e2e stub test) keep working by falling back to direct console.print(chunk, end="") of streamed tokens.
  • /exit now calls chat_app.exit() so Application.run_async returns cleanly; /clear calls chat_app.clear_transcript() (resets the in-flight stream window) instead of emitting a CSI 2J that would have scrolled into the persistent scrollback.
  • quoriv.ui.__init__ re-exports ChatApp; quoriv.ui.chat_input re-exports it too so the user's mental model of "the chat_input Application" maps to the new class.

Removed

  • The PromptSession-based interactive approval path inside quoriv.ui.prompts.prompt_approval. Callers without a chat_app now auto-reject with an explanatory message — auto_deny (read-only mode) is unaffected.
  • Rich Live use inside StreamRenderer.

Tests

  • tests/unit/ui/test_chat_app.py (new, 10) — construction surfaces, FloatContainer present, layout window counts with/without toolbar, keybindings cover Enter / Ctrl-C / Ctrl-D / Esc-Enter, push_chunk accumulates, clear_transcript resets, lazy console proxy, pre-run Rich print lands in stream buffer.
  • tests/unit/ui/test_stream.py rewritten for the ChatApp wiring — 12 cases covering buffer accumulation, sync/async finalize, ChatApp forwarding, idle-finalize short-circuit.
  • tests/unit/test_app_decisions.py — fake_prompt stub picks up the new chat_app=None kwarg.

Test count: 892 → 905 (+13). All gates green.

[1.1.0] — 2026-05-23

Added — Phase 5 Slice 1: chat UX polish + bordered input box

  • New welcome banner (quoriv.ui.banner.render_welcome_banner) — Rich Panel with ASCII Quoriv logo, centered tagline, key/value session context (model / mode / cwd / memory files), a two-column slash command grid (collapses to one column on narrow terminals), and the /help + /exit hint as the safety net.
  • Bordered input box (quoriv.ui.chat_input.build_chat_app + prompt_boxed) — replaces the inline PromptSession.prompt_async call with a real prompt_toolkit.Application whose layout puts the input BufferControl inside a Frame widget. The visible result is the bordered "box" users see in Claude Code, Aider, and similar TUIs.
  • Application is constructed per-turn (not persistent) — between turns Rich streams agent output normally because only one of (Rich, prompt_toolkit) owns the terminal at any moment. Persistent Application with Rich Live piped through prompt_toolkit is queued for a future slice.
  • Key bindings: Enter submits, Ctrl-C / Ctrl-D raise EOFError (matches PromptSession semantics so the existing chat loop's exit path keeps working), Esc-Enter (Alt+Enter on most terminals) inserts a newline for multi-line composition.
  • output / input kwargs let tests inject DummyOutput + create_pipe_input() so the Application constructs without a real TTY.
  • Slash-command autocomplete (quoriv.ui.slash_completer.SlashCommandCompleter) — prompt_toolkit Completer that fires only when the buffer starts with / and the user hasn't moved past the command word. Description shows in the popup's meta column. Bails on Tell me about /tmp/foo-style freeform so chat prompts never see a slash popup. Case-insensitive prefix match.
  • Shared input history across turns — _interactive_loop now creates a single InMemoryHistory and reuses it, so up-arrow recall works the way users expect.

Changed

  • quoriv.app._render_welcome delegates to the richer banner renderer in quoriv.ui.banner. Memory files and the slash-command grid are surfaced inside the panel so users see available commands without typing /help first.
  • quoriv.app._interactive_loop swaps PromptSession.prompt_async(HTML(">")) for prompt_boxed(...) with the shared completer + history + bottom_toolbar. Removed now-unused PromptSession and HTML imports.
  • TestWelcomePanel tests in test_app_slash.py updated for the new Rich Table layout — legacy "Memory:" colon-syntax dropped; matches the bare "Memory" label in the new grid.

Deferred to a future slice

  • Modal approval dialogs — Float-managed overlay replacing the existing inline prompt_approval() flow. Requires the Application to be persistent.
  • Scrolling output Window above the bordered input — Rich Live streaming piped into a prompt_toolkit Window widget. Requires a StreamRenderer rewrite and persistent Application. Both are tightly coupled (all-or-nothing) and warrant a dedicated session.

Tests

  • 27 new tests added in tests/unit/ui/:
  • test_banner.py (7) — smoke, version / model / mode / cwd surfaced, memory row omitted when empty / shown when present, slash grid rendered, /help hint always present, None cwd handled.
  • test_slash_completer.py (10) — empty buffer silent, non-slash silent, bare / lists everything, prefix filters, exact match self-completes, case-insensitive, description shown as meta, no-match empty, post-arg silent, freeform with embedded /path silent.
  • test_chat_input.py (10) — (app, buffer) returned, layout has more than one Window (Frame contributes border windows), bottom_toolbar adds exactly one extra Window, completer / history wired to Buffer, default in-memory history, keybinding registry contains Enter / Ctrl-C / Ctrl-D / Esc-Enter, frame_title kwarg accepted, prompt_boxed is async with the documented kwargs.

Test count: 865 → 892 (+27). All 4 gates green.

[1.0.1] — 2026-05-19

Fixed

  • Dependency drift on fresh installs. pyproject.toml originally pinned deepagents>=0.0.1 and the LangChain ecosystem with open upper bounds. A fresh pip install quoriv==1.0.0 therefore resolved deepagents 0.7+, which moved LocalShellBackend out of deepagents.backends, and quoriv chat crashed at import time with ImportError: cannot import name 'LocalShellBackend' from 'deepagents.backends'.
  • Tightened all agent-runtime pins to the version windows v1.0 was actually tested against:
  • deepagents>=0.6.1,<0.7.0
  • langchain>=1.0,<2.0, langchain-core>=1.0,<2.0, langgraph>=1.0,<2.0, langgraph-checkpoint-sqlite>=2,<3, langchain-openai>=1.0,<2.0
  • Provider extras (anthropic, gemini, ollama) and the all-providers bundle now carry matching major-version caps.
  • pydantic>=2.7,<3.0, pydantic-settings>=2.3,<3.0, typer>=0.12,<1.0 — same rationale, lower blast radius.
  • No code changes; this is purely a dependency-constraint hotfix. quoriv chat boots cleanly on pip install quoriv==1.0.1 from a fresh interpreter.

[1.0.0] — 2026-05-19

First stable release. The project crossed feature-complete at the end of Phase 4: full chat / TUI loop, six model providers (OpenAI, Anthropic, Gemini, Ollama, vLLM, OpenRouter), permission modes with HITL, AST + git + tests + web tools, MCP + Python plugins, hooks, replay viewer, eval harness with runner + CLI, opt-in telemetry with HTTP backend, tag-triggered PyPI release pipeline (OIDC trusted publishing), cross-platform PyInstaller binaries, and an mkdocs-material documentation site.

Added

Phase 0, Day 1 — Foundation scaffold

  • pyproject.toml with full Phase 1 dependencies and optional groups (ast, mcp, anthropic, gemini, ollama, dev, docs)
  • Apache 2.0 license
  • README with project overview, architecture, and usage examples
  • CONTRIBUTING.md with development workflow and architectural rules
  • SECURITY.md with disclosure policy
  • .gitignore covering Python, IDE, and Quoriv-specific artifacts
  • .pre-commit-config.yaml with ruff, mypy, and standard hooks
  • GitHub Actions CI: test.yml (Windows / macOS / Linux × Python 3.11 / 3.12), lint.yml (ruff + mypy)
  • Issue templates: bug report, feature request

Phase 0, Day 2 — Config layer + folder skeleton

  • src/quoriv/ folder skeleton with subpackages for core, models, tools, permissions, plugins, plugins.mcp, ui, config, observability, repo
  • src/quoriv/py.typed (PEP 561 marker — Quoriv ships types)
  • quoriv.config.schema — Pydantic v2 models (QuorivConfig, ModelConfig, PermissionsConfig, UIConfig, ToolsConfig) with extra="forbid" strictness and Literal types for enums (PermissionMode, Theme)
  • quoriv.config.loader — TOML loader with global (~/.quoriv/config.toml) + project (.quoriv/config.toml walked up from cwd) merge; global_config_path, project_config_path, load_config, _deep_merge
  • config.example.toml — annotated example at repo root
  • tests/conftest.py with fake_home fixture
  • 36 unit tests for config schema + loader

Phase 0, Day 3 — API keys + model factory

  • quoriv.config.keychain — keyring-backed API key storage with env-var fallback precedence (PROVIDER_ENV_VARS, set_api_key, get_api_key, delete_api_key, list_known_providers)
  • quoriv.models.base — ModelSpec (parses "provider:name" with first-colon split for Ollama tags), ModelCapabilities, MissingAPIKeyError
  • quoriv.models.factory — get_model("provider:name") with lazy provider loading via importlib; UnknownProviderError; list_providers
  • quoriv.models.openai — OpenAI provider via langchain-openai resolving keys through keychain
  • fake_keyring test fixture (in-memory keychain + env-var isolation)
  • Tests for keychain (12), base (14), factory (7), openai (5) — 74 tests total, all passing

Phase 0, Day 4 — CLI + chat loop (no DeepAgents yet)

  • quoriv.__main__ — makes python -m quoriv work the same as the quoriv console script
  • quoriv.cli — Typer app with commands:
  • quoriv chat [--model] [--mode] — start an interactive session
  • quoriv doctor — health-check Rich table (Python, configured models, permission mode, API key status per provider)
  • quoriv version — print the installed version
  • quoriv config show — print the loaded merged configuration as JSON
  • quoriv config set <provider> — prompt for API key (hidden input) and store in OS keychain
  • quoriv config list-providers — table of known providers, env-var names, and whether a key is configured
  • quoriv.app — async chat loop using rich.Console + prompt_toolkit.PromptSession; streams responses via LangChain model.astream(messages); slash commands /help, /clear, /exit, /quit; graceful Ctrl+C handling; helpful prompt when an API key is missing
  • Tests for the CLI commands using typer.testing.CliRunner (interactive chat deferred to Phase 1 integration tests)

Phase 0, Day 5 — DeepAgents wired

  • quoriv.core.agent.build_agent — constructs a session-scoped CompiledStateGraph via deepagents.create_deep_agent, with LocalShellBackend(root_dir=cwd) for real file ops + shell, an in-memory MemorySaver checkpointer for multi-turn state, and always-on PATH_PROTECTION rules denying writes to .env*, .git/**, .ssh/**, and secrets/**
  • quoriv.core.PATH_PROTECTION — tuple of FilesystemPermission deny rules that no permission mode can disable (security invariant)
  • quoriv.core.events — Rich-rendering helpers for LangGraph events: render_token, render_tool_start, render_tool_end, _format_args
  • quoriv.app rewritten to drive the DeepAgent:
  • Replaced direct model.astream(messages) with agent.astream_events({"messages": [HumanMessage]}, version="v2")
  • Per-session thread_id keys the checkpointer — /clear rotates to a new thread for a fresh conversation
  • Event-kind dispatch for on_chat_model_stream, on_tool_start, on_tool_end
  • quoriv chat gains a --cwd option to target a specific repo root
  • New tests:
  • tests/unit/core/test_agent.py — PATH_PROTECTION shape (5 rules covering env/git/ssh/secrets), build_agent raises MissingAPIKeyError without keys, returns a graph exposing astream_events/ainvoke/invoke when keys are present, honors model_override
  • tests/unit/core/test_events.py — render_token, render_tool_start, render_tool_end (short/long/None/multiline cases), _format_args (key=value, truncation per-value and per-call, non-dict fallback, empty dict)

With Day 5 wired, quoriv chat now has the full DeepAgents built-in toolset available out of the gate: write_todos, ls, read_file, write_file, edit_file, glob, grep, execute, and task (for sub-agent delegation).

Phase 1 Slice 1 — Permission modes wired

  • quoriv.permissions.modes — new module exporting PermissionMode (Literal["read-only", "ask", "auto", "yolo"]), WRITE_TOOLS ({"write_file", "edit_file"}), SHELL_TOOLS ({"execute"}), interrupt_on_for_mode(mode), and is_read_only(mode)
  • quoriv.permissions.paths — canonical home for PATH_PROTECTION (moved out of core/agent.py, where it was a transient placeholder)
  • quoriv.permissions.__init__ — re-exports the public surface from both submodules
  • quoriv.core.agent.build_agent now accepts a mode: PermissionMode = "ask" parameter and applies the compiled interrupt_on= dict via interrupt_on_for_mode(mode). Modes compile as:
  • yolo → {} (no prompts)
  • auto → {"execute": True} (prompt only before shell)
  • ask, read-only → {"write_file": True, "edit_file": True, "execute": True} (prompt before every write or shell call). Hard write denial in read-only is enforced at the approval-prompt UI (Slice 2).
  • quoriv.app.run_chat validates the mode string against ALLOWED_MODES and passes it through to build_agent
  • 24 new tests:
  • tests/unit/permissions/test_modes.py — 16 tests covering all 4 modes, tool-set membership and disjointness, dict freshness, is_read_only
  • tests/unit/permissions/test_paths.py — 8 tests covering rule shape, env/git/ssh/secrets coverage, POSIX rooting, immutability
  • tests/unit/core/test_agent.py updated: TestPathProtection removed (moved with the data to test_paths.py), TestBuildAgentModes added (parametrized build for all 4 modes)

Changed

  • PATH_PROTECTION is no longer re-exported from quoriv.core — import it from quoriv.permissions instead.

Test count: 108 → 130 (+22). All ruff / ruff format / mypy strict / pytest gates green.

Phase 1 Slice 2 — Approval prompt UI for HITL pauses

  • quoriv.ui.prompts — new module with prompt_approval(), ApprovalDecision (frozen dataclass), parse_choice(), READ_ONLY_DENIAL_MESSAGE constant, DecisionType Literal
  • Renders a yellow Rich Panel showing tool name, JSON-formatted args, and middleware description; prompts via prompt_toolkit for a/r (aliases: approve/reject/y/yes/n/no/deny)
  • auto_deny=True (used in read-only mode) renders the panel then auto-rejects with an explanatory message back to the agent
  • quoriv.app._stream_agent refactored to _drive_turn: loop that streams events, calls agent.aget_state() to detect pending HITL interrupts, prompts the user for each ActionRequest, and resumes with Command(resume={"decisions": [...]})
  • 28 new tests for parse_choice (approve/reject aliases, invalid input), ApprovalDecision (defaults, frozen behavior), prompt_approval (auto_deny path), _render_approval_panel, _format_args

Test count: 130 → 158 (+28). All gates green.

Phase 1 Slice 3 — Markdown streaming + edit_file diff renderer

  • quoriv.ui.stream — new module with StreamRenderer: Rich Live + Markdown wrapper that accumulates streamed tokens and live-renders them with markdown semantics (bold, code blocks, lists, syntax highlighting). Properties: is_streaming, buffer. Methods: push(text), finalize() -> str. Safe to call finalize on idle state.
  • quoriv.ui.diff — new module with compute_diff() (pure function returning difflib.unified_diff text) and render_edit_diff() (renders the diff with Rich Syntax(theme="ansi_dark") and an edit_file header line). Handles no-changes and missing-file-path cases.
  • quoriv.app._stream_events rewritten:
  • Each call now owns a StreamRenderer instance (lifecycle managed by try/finally)
  • on_chat_model_stream → renderer.push(text) (replaces raw render_token)
  • on_chat_model_end and on_tool_start → renderer.finalize() to close the Live cleanly
  • on_tool_start for edit_file → render_edit_diff() (colored unified diff) instead of generic header
  • All other tools fall through to the existing render_tool_start / render_tool_end
  • 16 new tests: test_stream.py (initial state, empty-push noop, accumulation, finalize semantics, restart after finalize); test_diff.py (identical strings → empty diff, change → unified diff, file path in headers, context lines respected, addition/removal only, render handles no-changes and missing path)

Test count: 158 → 174 (+16). All gates green. Source files: 25 → 27.

Phase 1 Slice 1b — PathProtectionMiddleware (custom guard)

  • quoriv.permissions.guard — new module with PathProtectionMiddleware, a langchain.agents.middleware.AgentMiddleware subclass that enforces PATH_PROTECTION deny rules at the middleware layer. Runs in after_model (and aafter_model), scans the latest AIMessage.tool_calls for path-bearing tools, and replaces denied calls with synthetic error ToolMessage objects so the agent observes the rejection on its next turn — a hard denial that bypasses HITL.
  • _TOOL_OPERATION map covers DeepAgents' built-in filesystem tools (ls, read_file, glob, grep → read; write_file, edit_file → write). Tools outside the map are treated as path-irrelevant and pass through.
  • _check_denial uses wcmatch.glob.globmatch with BRACE | GLOBSTAR flags — same semantics as DeepAgents' own FilesystemMiddleware, so deny patterns match identically whether DeepAgents adopts native sandbox-compatible permissions= later or not.
  • _extract_path reads file_path or path from the tool call's args dict (the two argument names DeepAgents' file tools use).
  • quoriv.permissions.__init__ re-exports PathProtectionMiddleware from guard.
  • quoriv.core.agent.build_agent now wires middleware=[PathProtectionMiddleware(list(PATH_PROTECTION))] into create_deep_agent. The custom guard layer is required because DeepAgents 0.6.1 raises NotImplementedError when permissions= is combined with a SandboxBackendProtocol backend (which LocalShellBackend is) — and we need LocalShellBackend for real shell execution.
  • 26 new tests in tests/unit/permissions/test_guard.py covering: rule passthrough, allow-rule precedence, every denied tool name (write/edit/read/ls/glob/grep), path arg variants (file_path vs path), glob patterns (*.env, .git/**, secrets/**), unrelated tools passing through, no-AIMessage / no-tool-calls early exits, multiple tool calls in one message (some denied, some kept), aafter_model delegating to sync, immutable rules view, and integration with PATH_PROTECTION itself.

Test count: 174 → 200 (+26). All gates green.

Phase 1 Slice 4 (minimal) — Python find_symbol tool

  • quoriv.tools.ast_tools — new module with find_symbol, a @tool-decorated callable that walks *.py files under a path and returns every definition matching a target name. Returns a list of records: {file, lineno, col_offset, kind, name, parent}.
  • Symbol kinds: function, async_function, class, variable (module/class-level Name = ... assignments). Methods recurse one level into class bodies and report parent=<ClassName>.
  • Implementation uses the stdlib ast module only — no tree-sitter yet. Skips .venv, venv, __pycache__, .git, build, dist directories so third-party code doesn't pollute results. Silently skips files that fail to parse or decode.
  • Accepts either a directory or a single file path. Nonexistent paths return [].
  • quoriv.tools.__init__ — QUORIV_TOOLS = [find_symbol]; quoriv.core.agent registers it via tools=list(QUORIV_TOOLS) in create_deep_agent. The DeepAgents built-ins (ls, read_file, write_file, edit_file, glob, grep, execute, task, write_todos) remain — find_symbol is purely additive.
  • 12 new tests in tests/unit/tools/test_ast_tools.py covering: function / async function / class / method-with-parent / module-level variable / no-match / subdirectory recursion / .venv + __pycache__ skip / syntactically broken file skip / nonexistent path / single-file path / BaseTool registration.

Slice 4b (tree-sitter expansion for ~30 languages, go_to_definition, find_references) is deferred.

Test count: 200 → 212 (+12). All gates green.

Phase 1 Slice 5 — Git tools (read-only)

  • quoriv.tools.git — new module with four plain @tool callables shelling out to git via subprocess.run (shell=False, list args, cwd-bound, UTF-8 decoded with errors="replace"):
  • git_status(cwd=".") — returns {branch, ahead, behind, is_clean, files} (each file: {path, index, worktree, old_path?}). Detached HEAD reports branch=None. Renames keep old_path.
  • git_diff(path=None, staged=False, revision_range=None, cwd=".") — returns {diff, is_empty}. Combines path scoping with working-tree, staged (--cached), or revision-range diffs.
  • git_log(limit=20, path=None, cwd=".") — returns {entries, count} where each entry is {sha, short_sha, author, email, date, subject}. Uses \x1f field separator + ISO dates for unambiguous parsing.
  • git_blame(file, line_start=None, line_end=None, cwd=".") — returns {file, entries} with {sha, author, date, lineno, content} per line. -L start,end (or single line) scopes the blame.
  • Uniform failure shape across all four tools: {"error": "<message>"} for non-zero git exit, not-a-repo errors, missing-file errors, invalid args (e.g., limit < 1), and the git-not-on-PATH case (FileNotFoundError → "git executable not found on PATH (...)").
  • Parser helpers _parse_status_porcelain and _parse_branch_line are exported for direct unit tests. Branch line covers ## main, ## main...origin/main, [ahead N], [behind M], [ahead N, behind M], and ## HEAD (no branch).
  • Write operations (git add / git commit / git stash / ...) are intentionally not in this slice — they land later behind interrupt_on= so HITL prompts before mutating the working tree.
  • quoriv.tools.__init__ — QUORIV_TOOLS now exposes [find_symbol, git_status, git_diff, git_log, git_blame]; __all__ re-exports each tool by name.
  • 42 new tests in tests/unit/tools/test_git.py:
  • TestParseBranchLine (6) — every branch-line variant
  • TestParseStatusPorcelain (6) — modified / staged / untracked / rename / too-short / malformed
  • TestGitStatus (6) — clean repo, untracked file, modified-then-staged, not-a-repo, FileNotFoundError for missing git binary, default-cwd via monkeypatch.chdir
  • TestGitDiff (7) — no-changes, working-tree, staged, revision range, path scoping, not-a-repo, bad revision
  • TestGitLog (7) — reverse-chronological ordering, entry-field shape, limit, path filter, invalid limit, empty repo, not-a-repo
  • TestGitBlame (5) — full-file, line range, single line, missing file, not-a-repo
  • TestToolRegistration (5) — each tool is a BaseTool with the right .name and is present in QUORIV_TOOLS
  • Test infrastructure: a small in-file helper trio (_git, _init_repo, _commit) builds deterministic repos by pinning GIT_AUTHOR_* / GIT_COMMITTER_* env vars and commit.gpgsign=false per repo so tests are stable across hosts.

Test count: 212 → 254 (+42). All gates green.

Phase 1 Slice 7 — SQLite session persistence + /save / /load / /resume

  • langgraph-checkpoint-sqlite>=2.0.0 added to runtime dependencies (resolves to 3.1.0). aiosqlite>=0.20.0 was already in place.
  • quoriv.core.persistence — new module with:
  • Path helpers: quoriv_dir(cwd), db_path(cwd), registry_path(cwd), ensure_quoriv_dir(cwd). The agent's SQLite checkpointer lives at <cwd>/.quoriv/sessions.db and the named-session sidecar at <cwd>/.quoriv/sessions.json (per-project, mirroring how .git/ is per-repo).
  • NamedSession frozen dataclass: {name, thread_id, saved_at} (ISO-8601 UTC timestamp).
  • SessionRegistry — file-backed name → thread_id mapping. Loaded on construction, written eagerly on every mutation. Malformed / missing files reset to an empty registry rather than raising — the underlying SQLite DB is the real source of truth, so a corrupted name index is a recoverable convenience-layer issue. Public API: for_cwd(cwd), save(name, thread_id, *, now=None), load(name), list_named(), most_recent(), remove(name), path.
  • quoriv.core.__init__ re-exports SessionRegistry, NamedSession, db_path, ensure_quoriv_dir, quoriv_dir, registry_path.
  • quoriv.app.run_chat rewritten to manage the saver lifecycle:
  • Resolves cwd to an absolute Path, calls ensure_quoriv_dir, opens AsyncSqliteSaver.from_conn_string(str(sessions_db)) via async with, passes the saver to build_agent(..., checkpointer=saver). The prompt-loop body moved into _interactive_loop(console, agent, registry, mode) so the lifecycle stays explicit.
  • New slash commands (with handler helpers _handle_save, _handle_load, _handle_resume, _print_saved_sessions):
  • /save [name] — anchor the current thread under name (default: first 8 chars of the thread id). Overwrites any prior entry under that name.
  • /load — list saved sessions (most-recent first), or /load <name> to switch the active thread.
  • /resume — switch to the most-recently-saved thread (by saved_at).
  • _handle_slash signature now takes a SessionRegistry; legacy commands (/help, /clear, /exit, /quit) still work and surface the new entries through /help. SLASH_COMMANDS extended accordingly.
  • 44 new tests:
  • tests/unit/core/test_persistence.py (26) — path helpers, construction (empty / no-file-on-init), save returns/persists/round-trips/overwrites/timestamps, load known/unknown, list_named ordering, most_recent by saved_at, remove existing/unknown/persists, malformed-file recovery (bad JSON, non-dict root, missing sessions key, non-list value, dropped-field entries).
  • tests/unit/test_app_slash.py (18) — SLASH_COMMANDS table, /save with-name / default-name / reports / empty-thread-id / overwrites, /load known / unknown / empty-list / populated-list, /resume most-recent / empty-registry, legacy /exit / /quit / /clear / /help / unknown-command paths.

Test count: 254 → 298 (+44). All gates green.

Phase 1 Slice 5b — Git write tools (HITL-gated)

  • quoriv.tools.git extended with three write callables sharing the same subprocess.run (shell=False, list args, cwd-bound) and dict[str, Any] return shape as the read tools:
  • git_add(paths=None, cwd=".") — stages specific paths or all changes (git add -A). Returns {"staged_files": list[str]} from git diff --cached --name-only after the add. Empty paths list is treated as "add all".
  • git_commit(message, cwd=".") — creates a commit from the current index. Returns {"sha", "short_sha", "subject", "branch": str | None} parsed from git rev-parse / git log -1 rather than the locale-sensitive git commit output. Empty message rejected locally with a structured error.
  • git_stash(message=None, include_untracked=False, cwd=".") — pushes the working tree onto the stash. Returns {"stashed": bool, "message": str | None} — stashed=False when git printed "No local changes to save".
  • All three respect local git config: no --no-gpg-sign, no --no-verify. Tests configure commit.gpgsign=false per fixture repo so signing-required hosts do not block the suite.
  • quoriv.permissions.modes.GIT_WRITE_TOOLS = frozenset({"git_add", "git_commit", "git_stash"}) — Quoriv-specific git tools that mutate repo state, gated alongside WRITE_TOOLS in ask / read-only modes. auto mode lets them run silently (like write_file); yolo lets everything through.
  • interrupt_on_for_mode("ask") and interrupt_on_for_mode("read-only") now include the new tool names; auto and yolo deliberately do not.
  • quoriv.permissions.__init__ re-exports GIT_WRITE_TOOLS. quoriv.tools.__init__ extends QUORIV_TOOLS to include the three new tools and re-exports them.
  • 23 new tests:
  • tests/unit/permissions/test_modes.py (5) — GIT_WRITE_TOOLS membership shape, pairwise-disjoint with WRITE_TOOLS and SHELL_TOOLS, auto does NOT gate them, ask does, read-only does.
  • tests/unit/tools/test_git.py (18): TestGitAdd (6) — add-all, specific paths, empty-paths defaults, nonexistent path errors, clean-repo returns empty, not-a-repo errors. TestGitCommit (5) — staged commit, nothing-staged errors, empty-message rejected locally, subject is first line, not-a-repo errors. TestGitStash (5) — with changes, with no changes (stashed=False), with message, -u includes untracked, not-a-repo errors. TestToolRegistration parametrizes over all 7 git tools now.

Test count: 298 → 321 (+23). All gates green.

Phase 1 Slice 6 — Language-aware run_tests tool

  • quoriv.tools.tests — new module with one @tool callable run_tests(framework=None, path=None, cwd=".") that auto-detects the project's test framework from marker files and runs the suite via subprocess.run (shell=False, list args, cwd-bound).
  • Detection (in order — first match wins, so a polyglot repo with both pyproject.toml and package.json defaults to Python, matching Quoriv's own layout):
  • pyproject.toml / pytest.ini / setup.cfg → pytest
  • package.json → npm test
  • Cargo.toml → cargo test
  • go.mod → go test ./...
  • Command construction (_build_command): each framework gets its idiomatic invocation. pytest -q, npm test --silent, cargo test, go test ./.... Path scoping uses the framework's native convention — positional arg for pytest, after -- for npm/cargo, replaces ./... for go.
  • Returns {framework, command, exit_code, passed, stdout, stderr} on success. The structured shape lets the LLM check passed: bool without parsing free-form output. On failure (no detection, cwd missing, unrecognized override, runner binary not on PATH) returns {"error": "..."} plus the attempted framework / command when known, so the agent can surface what was tried.
  • quoriv.tools.__init__ registers run_tests in QUORIV_TOOLS and re-exports it. run_tests deliberately stays outside GIT_WRITE_TOOLS — it executes a runner locally without mutating repo state; the session's existing shell-execution gate applies via DeepAgents' execute if the underlying runner shells out further.
  • 29 new tests in tests/unit/tools/test_runner.py:
  • TestDetectFramework (8) — each marker file maps to the right framework, empty dir returns None, polyglot tie-break favors Python.
  • TestBuildCommand (9) — default + with-path for each framework, unknown framework raises ValueError.
  • TestRunTests (10) — passes / failure-sets-passed-false, framework override (works even with no marker files), path scoping for pytest, no-framework-detected error, unknown-framework override error, nonexistent cwd error, runner-binary-missing error (with framework + command echo), subprocess called with resolved absolute cwd, shell= never set (subprocess defaults to shell=False).
  • TestToolRegistration (2) — run_tests is a BaseTool with the right name and is present in QUORIV_TOOLS.

Slice 6b (parsed test-count summary from each runner's output) is deferred.

Test count: 321 → 350 (+29). All gates green.

Phase 1 Slice 8 — Status line + introspection slash commands

  • quoriv.app — persistent bottom_toolbar wired into the PromptSession. New pure helper _build_status_line(model_id, mode, cwd, thread_id) returns the formatted bar:

    | mode= | | thread=

Bottom-bar callable closes over the loop's thread_id so /clear / /load / /resume rotations are reflected on the next prompt without any extra plumbing. - Four new read-only slash commands wired through _handle_slash (now accepting keyword-only model_id / cwd / mode with safe defaults so legacy call sites and prior-slice tests still type-check): - /tools — lists DeepAgents built-ins (write_todos, ls, read_file, write_file, edit_file, glob, grep, execute, task) under one heading and QUORIV_TOOLS (find_symbol, git_*, run_tests) under another, each with a one-line description. - /memory — shows status of ~/.quoriv/memory.md (global) and <cwd>/PROJECT.md (project): which exist + byte size, with a hint when neither is present. These are the files DeepAgents' MemoryMiddleware will load once build_agent wires memory=[...] in a later slice. - /mode — prints the active permission mode, its description, the current interrupt_on_for_mode(mode) tool list (so the user sees exactly which tools will prompt), and the full available-modes table with the current one marked. Includes a hint that live-switching needs quoriv chat --mode <name> for now. - /cost — explicit stub pointing at Slice 9 (token tracking arrives with the local JSON trace log). - _handle_slash dispatch extended with the four new commands. The function gained keyword-only model_id / cwd / mode parameters with defaults; _interactive_loop now also takes model_id / cwd and threads them into both the toolbar closure and the slash dispatcher. - SLASH_COMMANDS table extended so /help lists every new entry. - Two new module-level tables provide the source-of-truth descriptions: _DEEPAGENTS_BUILTIN_TOOLS (9 entries) and _MODE_DESCRIPTIONS (4 entries, keyed by PermissionMode Literal). - 10 new tests in tests/unit/test_app_slash.py: - TestSlice8SlashCommandsListed (1) — all four new commands appear in SLASH_COMMANDS. - TestToolsCommand (1) — output names representative built-ins (write_todos) and Quoriv tools (git_status, run_tests) under their respective headings. - TestMemoryCommand (2) — empty-cwd reports "No memory files found"; PROJECT.md is detected and its byte count is shown. - TestModeCommand (3) — ask mode lists every gated tool (including the git writes from Slice 5b); yolo mode reports nothing-gated; the available-modes table is always rendered. - TestCostCommand (1) — output names the Slice 9 deferral plainly so the user knows where token tracking lands. - TestBuildStatusLine (2) — pure-function checks: every field appears, thread_id is truncated to 8 chars, and the delimiter shape is stable for edge-case paths.

Test count: 350 → 360 (+10). All gates green.

Phase 1 Slice 9 — Local JSONL trace log + token-aware /cost

  • quoriv.observability.trace — new module with TraceLogger. Append-only JSONL writer per chat thread; lazy file creation (no on-disk artifact until first write); _sanitize() recursively coerces non-JSON-native values (Path, dataclasses, sets, arbitrary objects via str()) so unserializable values never raise. Public API: path property, log(event, **fields), read_events(), token_totals(). read_events() tolerates corrupt lines and non-dict JSON values (skipped silently — a single bad line never poisons the log). token_totals() sums across every model_complete event with sensible fallback (input + output when total_tokens is absent).
  • quoriv.core.persistence — new trace_path(cwd, thread_id) and traces_dir(cwd) helpers, both re-exported from quoriv.core. Canonical location: <cwd>/.quoriv/traces/<thread_id>.jsonl. Mirrors the db_path / registry_path pattern from Slice 7.
  • quoriv.app — _interactive_loop now owns a TraceLogger. The logger rotates alongside thread_id when /clear / /load / /resume switch threads (old log file remains on disk so a future /load can return to it and see its history). _drive_turn brackets the turn with turn_start / turn_end events; _stream_events records model_complete (with input_tokens / output_tokens / total_tokens extracted from the final AIMessage.usage_metadata when LangChain provides it), tool_start (with args), and tool_end (with output preview, truncated at 500 chars).
  • /cost is no longer a stub — it reads tracer.token_totals() for the active thread and prints aligned counts plus the trace file path. Surface still notes that per-provider dollar-cost calculation is deferred (waiting on a rate table).
  • _handle_slash gained a keyword-only tracer: TraceLogger | None = None parameter, defaulted so existing test calls without a tracer still type-check. When None, /cost falls back to a "no logger attached" message.
  • 28 new tests:
  • tests/unit/observability/test_trace.py (24): TestSanitize (7) — primitives passthrough, dict recursion, Path coercion, dataclasses, sets/tuples, fallback to str(), non-string dict keys. TestTraceLoggerWrites (7) — .path property, lazy file creation, parent-dir auto-creation, JSONL append, ISO-8601 UTC timestamps, supplied fields preserved, unserializable values sanitized. TestTraceLoggerReads (4) — missing-file empty list, round trip, malformed-line resilience, non-dict JSON skipped. TestTokenTotals (5) — empty log zeros, multi-event sum, total_tokens fallback from input + output, ignores non-model_complete events, ignores non-int token fields. TestTracePathIntegration (2) — canonical filesystem location, round-trip through a fresh logger instance.
  • tests/unit/test_app_slash.py (4) — TestCostCommand rewritten: no-tracer reports "No trace logger"; empty log reports zero calls + trace file path; populated log shows token totals (input/output/total/calls); trace file path always surfaced.

Test count: 360 → 388 (+28). All gates green.

Phase 1 Slice 6b — Parsed pytest counts in run_tests

  • quoriv.tools.tests._parse_pytest_summary — new pure helper that extracts {passed, failed, errors, skipped, duration_seconds} from the terminal summary line emitted by pytest. Regex anchors on the "in <duration>s" suffix so unrelated === separator lines never match; the last match in the output wins so per-session header lines do not pollute the result. Returns the all-None shape when no summary line is found — the caller can tell "couldn't parse" (e.g., pytest crashed at collection) from "0 of everything".
  • run_tests return shape gains a summary block with the parsed fields when framework == "pytest". Other frameworks get the placeholder all-None summary until Slice 6c lands cargo / go / npm parsers — keeping the shape stable lets the LLM check summary["passed"] once instead of branching per framework.
  • Parser reads stdout + stderr concatenated so CI environments that redirect pytest output to stderr still get counts.
  • 13 new tests in tests/unit/tools/test_runner.py:
  • TestParsePytestSummary (9) — passing-only, failed-only, mixed (passed/failed/errors), passed+skipped, "no tests ran", plural "errors", no-summary-line returns all-None, empty input returns all-None, last-match-wins when multiple === lines.
  • TestRunTests (4 new) — pytest summary surfaces counts on the result dict; stderr summary parsed; pytest with no summary line returns null counts; non-pytest framework gets the all-None placeholder.

Test count: 388 → 401 (+13). All gates green.

Phase 1 Slice 9c — Per-provider dollar-cost estimates in /cost

  • quoriv.observability.cost — new module with the shipping rate table for /cost:
  • ProviderRate — frozen dataclass {input_per_1k: float, output_per_1k: float} (USD).
  • RATES: dict[str, ProviderRate] — 17 entries keyed by provider:model prefix. OpenAI (gpt-5, gpt-4o, gpt-4o-mini, gpt-4-turbo, gpt-4, gpt-3.5-turbo), Anthropic (claude-opus-4, claude-sonnet-4, claude-haiku-4, claude-3-5-sonnet, claude-3-5-haiku, claude-3-opus, claude-3-haiku), Gemini (gemini-1.5-pro, gemini-1.5-flash), and local sentinels (ollama:, vllm: → free).
  • lookup_rate(model_id) — longest-prefix match so "openai:gpt-4o-mini" resolves to its own entry rather than the broader "openai:gpt-4o"; versioned ids like "openai:gpt-4o-2024-08-06" fall back to the prefix.
  • estimate_cost(rate, input_tokens, output_tokens) — returns {input_cost_usd, output_cost_usd, total_cost_usd} from the per-1k rate.
  • quoriv.observability.__init__ re-exports RATES, ProviderRate, lookup_rate, estimate_cost.
  • /cost is no longer dollar-blind. _handle_cost gained a keyword-only model_id parameter, threaded through _handle_slash. When the model has a rate, output now includes an "Estimated cost" block with input/output/total dollar amounts to 4-decimal precision. When no rate is configured, a friendly "update quoriv.observability.cost.RATES" hint is printed alongside the token totals — the agent still gets actionable info without a stale or fabricated dollar figure.
  • 20 new tests:
  • tests/unit/observability/test_cost.py (17): TestProviderRate (2) — frozen + value equality. TestRatesTable (5) — non-empty, every entry is ProviderRate, no negative rates, every known provider (openai/anthropic/gemini/ollama) has at least one row, every key contains a colon. TestLookupRate (6) — exact match, longest-prefix wins, versioned suffix falls back to the prefix, ollama sentinel matches every model, unknown provider returns None, empty id returns None. TestEstimateCost (4) — zero tokens, basic math, sub-thousand tokens, free rate zeros out.
  • tests/unit/test_app_slash.py (3 new in TestCostCommand): known model shows dollar estimate with provider id; unknown model shows "No rate configured for ... — update RATES"; ollama renders $0.0000.

Test count: 401 → 421 (+20). All gates green.

Phase 1 Slice 6c — cargo / go / npm output parsers

  • quoriv.tools.tests extended with three pure helpers that mirror _parse_pytest_summary's contract — input = combined stdout+stderr, output = {passed, failed, errors, skipped, duration_seconds} with the all-None fallback when nothing matches:
  • _parse_cargo_summary — matches test result: ok. N passed; M failed; K ignored; ...; finished in Xs lines (one per test binary). Multi-crate workspaces produce multiple summary lines; counts and durations are summed across all of them. ignored maps to skipped. cargo doesn't distinguish errors from failures so errors is always 0.
  • _parse_go_summary — counts per-test status lines (--- PASS: / --- FAIL: / --- SKIP:) for the count fields, and sums per-package summary durations (ok pkg X.XXs / FAIL pkg X.XXs). When only one or the other appears, the absent fields stay None or 0 in a documented way.
  • _parse_npm_summary — parses jest / vitest-style summary blocks (Tests: N passed, M failed, K total + Time: X.XX s). vitest's todo / pending categories collapse into skipped to keep the cross-runner shape stable. Other npm runners (mocha, ava, …) fall through to the all-None shape — they emit summaries in different formats and aren't covered by this slice.
  • _FRAMEWORK_PARSERS dispatch dict replaces the if chosen == "pytest" branch in run_tests. Adding a new framework now means adding a marker file, command builder, and parser entry — all three concerns sit next to each other in the module.
  • The all-None fallback still applies when a framework's runner emits output that doesn't match the expected summary shape (e.g., cargo compile error before tests run), so the caller can still tell "couldn't parse" from real zero counts. The Slice 6b test that exercised the placeholder path was renamed and rewritten to assert this contract under the new parsers.
  • 16 new tests in tests/unit/tools/test_runner.py:
  • TestParseCargoSummary (4) — single-package success, single-package failure, multi-package counts + duration sum, no-summary returns all-None.
  • TestParseGoSummary (4) — --- PASS / FAIL / SKIP counts with package-duration sum, package-summary-only zero counts but non-None duration, multiple passes only (no package summary → duration stays None), no-recognisable-output returns all-None.
  • TestParseNpmSummary (5) — jest-style full summary, passing-only, todo + pending collapse to skipped, no-summary returns all-None, no Time: line keeps duration None.
  • TestRunTests (3 new) — each new framework dispatches to its own parser end-to-end through run_tests.

Test count: 421 → 437 (+16). All gates green.

Phase 1 Slice 4b — Tree-sitter multi-language symbol intelligence

  • Migrated the ast extra from the abandoned tree-sitter-languages (no Python 3.13 wheels) to the maintained tree-sitter-language-pack>=0.6.0 and bumped tree-sitter>=0.24.0. The new pack ships bundled C wheels for Python 3.10-3.13 on Linux / macOS / Windows and covers ~80 languages.
  • quoriv.repo.ast — new module: LANGUAGE_BY_EXTENSION (40+ entries), detect_language(path) by suffix, get_parser(language) lazy-loaded from the pack, is_available() so callers without the extra installed degrade gracefully. The lazy imports keep the rest of Quoriv working when someone installs without [ast].
  • quoriv.repo.symbols — new module with Symbol frozen dataclass and two public functions: extract_definitions(source, language, *, target=None) and find_references(source, language, target). Per-language DEFINITION_KINDS maps tree-sitter node kinds to Quoriv symbol kinds for python / javascript / typescript / tsx / go / rust / java / kotlin / c / cpp / csharp / ruby / php / lua / elixir / swift. CONTAINER_KINDS tracks scopes (class / struct / trait / impl / module / namespace / protocol / interface / enum) so nested method definitions record their parent. Uses direct tree walks (not QueryCursor) because the language pack ships its own Node class that's binary-incompatible with the public tree_sitter Query API.
  • quoriv.tools.ast_tools expanded:
  • find_symbol is now multi-language. Python (.py / .pyi) keeps the stdlib ast path (no extra needed); every other extension routes through tree-sitter via quoriv.repo.symbols.
  • New @tool callables: go_to_definition(name, path=".") — strict alias of find_symbol named for the agent's "jump-to-def" intent; find_references(name, path=".") — every identifier-like node whose text equals name (definition + callers + type uses + field accesses).
  • _iter_source_files walks the path, skipping common build / vendor dirs (.venv, venv, __pycache__, .git, build, dist, node_modules, target).
  • QUORIV_TOOLS now exposes 11 tools: find_symbol, go_to_definition, find_references, the 7 git tools, and run_tests.
  • 63 new tests across three files:
  • tests/unit/repo/test_ast.py (26) — extension → language for 21 file types, table sanity, case insensitivity, parser smoke tests for python/go, unknown language → LookupError.
  • tests/unit/repo/test_symbols.py (14) — Symbol frozenness, table coverage, Python def/class/method extraction with parent, Python reference search hits both def and call sites, Go type/method/function extraction, Go reference search across declarations and uses, TypeScript interface/type/class/method/function, Rust struct/trait/impl/function, graceful empty-list for unsupported languages.
  • tests/unit/tools/test_ast_tools.py (23 new) — TestFindSymbolMultiLanguage (5) covers go/ts/rust + node_modules / target skip + a mixed-language file tree. TestGoToDefinition (3) verifies alias semantics + registration. TestFindReferences (6) covers Go callsites, TS field access, empty/missing cases, registration.

Test count: 437 → 500 (+63). All gates green.

Phase 1 Slice 9d — Config-driven cost rates

  • quoriv.config.schema — two new Pydantic v2 models: CostRate (USD per 1,000 tokens with Field(..., ge=0.0) on both input_per_1k and output_per_1k) and CostConfig (rates: dict[str, CostRate], defaults to empty). Both carry extra="forbid" so a typo in ~/.quoriv/config.toml fails loudly at validation time. QuorivConfig gains a cost: CostConfig = Field(default_factory=CostConfig) section.
  • quoriv.observability.cost.effective_rates(config) -> dict[str, ProviderRate] — merges the user's cost.rates over a fresh copy of the built-in RATES. The built-in table is never mutated; the returned dict is fresh each call. effective_rates(None) returns a copy of RATES so callers without a config object still get the standard table.
  • quoriv.observability.cost.lookup_rate gained an optional second rates argument. Passing None falls back to the built-in RATES (legacy behaviour preserved). Passing the result of effective_rates(config) means longest-prefix lookup operates over the merged table — a user's fine-grained anthropic:claude-opus-4-7 entry naturally wins over the broader built-in anthropic:claude-opus-4 prefix.
  • quoriv.observability.__init__ re-exports effective_rates.
  • quoriv.app.run_chat precomputes cost_rates = effective_rates(config) once per session and threads it through _interactive_loop → _handle_slash → _handle_cost. The new keyword-only cost_rates parameter carries a None default on every layer so older test entry points keep working.
  • The "no rate configured" hint in /cost now points the user at [cost.rates."{provider}:{model}"] in ~/.quoriv/config.toml rather than at the in-source RATES dict, matching the new override path.
  • config.example.toml — documents the [cost.rates."provider:model"] block with example overrides, the non-negative-float constraint, and the longest-prefix lookup rule.
  • 20 new tests:
  • tests/unit/config/test_schema.py TestCostConfig (8) — empty defaults, QuorivConfig.cost.rates == {}, rate accepts 0.0, rate rejects negative input / output, missing fields rejected, extra fields rejected on both CostRate and CostConfig, full round-trip through QuorivConfig.model_validate.
  • tests/unit/observability/test_cost.py TestLookupRateCustomTable (3) — uses supplied table not built-in, longest-prefix within custom table, explicit None falls back to built-in.
  • tests/unit/observability/test_cost.py TestEffectiveRates (6) — None returns a built-ins copy that is safe to mutate, empty config matches built-ins, user override replaces built-in by key (other entries survive), user can add a new provider, calling effective_rates does not mutate RATES, a more specific user key wins over a broader built-in prefix via merged-table longest-prefix.
  • tests/unit/test_app_slash.py TestCostCommand (3 new) — user rate override shadows the built-in (1k input @ $0.05 + 1k output @ $0.20 → totals appear in /cost); user rate can add an unknown model so a previously rateless id now renders an estimate; legacy "No rate configured" hint absent when an override exists.

Test count: 500 → 520 (+20). All gates green.

Phase 1 Slice 8b — Live /mode switch

  • /mode <name> now rebuilds the compiled DeepAgent in place against the same AsyncSqliteSaver checkpointer. The running thread's conversational state survives the switch — only the interrupt_on= dict changes (via interrupt_on_for_mode(new_mode)). No restart, no new thread id.
  • quoriv.app._SlashResult gained a third slot, new_mode: PermissionMode | None. The interactive loop branches on it after a slash dispatch: when set, it calls build_agent(config, model_override=..., cwd=..., mode=new_mode, checkpointer=saver) and reassigns the local agent. The _toolbar closure reads the latest permission_mode at call time, so the status line reflects the new mode on the next prompt redraw without explicit refresh.
  • _handle_mode is now mode-aware: with no argument it preserves the Slice 8 display (current mode + gated tools + menu); with an argument it normalises to lowercase, validates against ALLOWED_MODES, short-circuits on same-mode requests with a friendly "Already in <mode>" note, and surfaces unknown values with the valid set listed inline. The closing line of the display variant now reads Switch live with /mode <name> instead of the stale Live-switch lands in a later slice.
  • quoriv.app.run_chat and _interactive_loop thread config, model_override, and the open AsyncSqliteSaver through as keyword-only args. All three default to None so legacy single-mode test entry points keep working without modification.
  • SLASH_COMMANDS["/mode"] description updated from "Show the current permission mode and what each mode gates" to "Show permission mode (no arg) or live-switch (/mode <name>)" so /help advertises the new form.
  • 6 new tests in tests/unit/test_app_slash.py::TestModeCommand:
  • test_no_arg_does_not_switch — display-only path returns _SlashResult(new_mode=None).
  • test_valid_arg_returns_new_mode — /mode yolo from ask returns _SlashResult(new_mode="yolo") silently (the loop, not the handler, prints the confirmation).
  • test_valid_arg_with_uppercase_normalized — /mode YOLO normalises to "yolo".
  • test_same_mode_does_not_switch — /mode ask while in ask prints "Already in" and returns no switch.
  • test_invalid_arg_reports_error — /mode banana prints the unknown-mode error with the offending input and the full valid set; returns no switch.
  • test_all_modes_can_be_targets — round-trips every valid mode as a target from a different starting mode to catch any PermissionMode literal-narrowing regression in the dispatch path.

Test count: 520 → 526 (+6). All gates green.

Phase 1 Slice 9b — End-to-end stubbed-LLM turn test

  • tests/integration/test_e2e_stubbed_chat.py — first integration test: drives a full user turn through quoriv.app._drive_turn against a real build_agent-compiled DeepAgent whose model is a _StubChatModel (a langchain_core.language_models.fake_chat_models.GenericFakeChatModel subclass with bind_tools short-circuited to self, since the base class raises NotImplementedError and DeepAgents binds tools at compile time). quoriv.core.agent.get_model is monkeypatched to return the stub, so the test exercises the same code path the CLI uses — including the LangGraph event stream, the StreamRenderer Live, the TraceLogger writes, the MemorySaver checkpointer, and the PathProtectionMiddleware. Mode is yolo to skip HITL interrupts so the agent finishes in one model turn.
  • 4 tests in TestEndToEndTurn:
  • test_drive_turn_writes_turn_start_and_end — verifies the trace bracket: first record is turn_start with the original thread_id/user_input/mode, last record is turn_end with the matching thread_id. That bracket is the contract /cost and any future observability tooling depend on.
  • test_drive_turn_records_model_complete — verifies at least one model_complete record lands per turn (one stub AIMessage → one on_chat_model_end → one trace entry).
  • test_drive_turn_renders_stub_response — sanity that the LLM payload is actually rendered to the console buffer through StreamRenderer, not just traced.
  • test_status_line_built_from_session_context — _build_status_line(model_id, mode, cwd, thread_id) is unaffected by a completed turn and still returns a well-formed string with the expected fields, mode marker, truncated thread id, and three separators.
  • Catches future regressions where _drive_turn / _stream_events / TraceLogger drift out of sync — a renamed LangGraph event key, a missing tracer call, or a status-line format change all surface here instead of in production.

Test count: 526 → 530 (+4). All gates green. With Slice 9b done, Phase 1 is complete.

Phase 2 Slice 1 — Memory wiring

  • quoriv.core.memory — new module: MemoryCandidate NamedTuple (label, path), memory_candidates(cwd) returning the ordered list of two candidates (global first as ~/.quoriv/memory.md, project second as <cwd>/PROJECT.md), and resolve_memory_files(cwd) filtering to ones that exist via Path.is_file() (so a directory named PROJECT.md is correctly rejected). The order matters: DeepAgents concatenates the list in load order under <agent_memory> in the system prompt, so global-then-project lets a project file refine a global note — same precedence rule the TOML loader uses.
  • quoriv.core.agent.build_agent now calls resolve_memory_files(root), converts to strings, and passes the result to create_deep_agent(memory=...). When the resolved list is empty, the argument is None — DeepAgents documents None (not []) as the contract for "don't attach the middleware", so we honor that.
  • quoriv.core.__init__ re-exports MemoryCandidate, memory_candidates, resolve_memory_files.
  • quoriv.app._handle_memory rewritten to source its candidate list from memory_candidates(cwd) rather than hardcoded paths. Each present file now gains a (loaded) tag, making it clear the agent's MemoryMiddleware has actually seen the file — not just that it exists on disk.
  • quoriv.app._render_welcome adds a Memory: PROJECT.md, memory.md line to the welcome panel when at least one file is loaded; silent when neither exists so first-time users don't see clutter.
  • 19 new tests:
  • tests/unit/core/test_memory.py (10): TestMemoryCandidates (5) — load-order, global path uses fake_home/.quoriv, project path tracks the supplied cwd, NamedTuple round-trip, different cwds yield different project paths. TestResolveMemoryFiles (5) — neither / project-only / global-only / both / directory-with-the-name rejected.
  • tests/unit/core/test_agent.py::TestBuildAgentMemoryWiring (4) — monkeypatched create_deep_agent captures kwargs and verifies memory=None when no files, memory=[...PROJECT.md] when only project file present, memory=[...memory.md] when only global, and global-then-project ordering when both exist.
  • tests/unit/test_app_slash.py::TestMemoryCommand (2 new) — (loaded) tag appears next to present files; absent files do not show the tag. The two pre-existing /memory tests gained the fake_home fixture so a developer's real ~/.quoriv/memory.md doesn't leak into the assertion.
  • tests/unit/test_app_slash.py::TestWelcomePanel (3) — no memory line when neither file present; PROJECT.md surfaces in the panel; global memory.md surfaces.

Test count: 530 → 549 (+19). All gates green.

Phase 2 Slice 2 — quoriv init

  • New quoriv init [PATH] command in quoriv.cli scaffolds a starter PROJECT.md at the target directory (defaults to cwd). Refuses to overwrite an existing file by default — exits non-zero with a "pass --force to overwrite" hint so a CI script can't silently nuke a hand-edited PROJECT.md. --force / -f overrides.
  • quoriv.core.memory.PROJECT_MEMORY_TEMPLATE — the starter content. One-screen template: top-level note explaining what Quoriv does with the file, then sections for Project overview, Architecture, Conventions, Useful commands, Things to avoid. Designed to fit on one screen so users can scan the shape before editing.
  • Removed an outdated # noqa: TC003 on from pathlib import Path in cli.py — the import is now used at runtime by the new init command's Path arguments and Path.cwd() fallback, not just in type annotations.
  • 6 new tests in tests/unit/test_cli.py::TestInit:
  • test_creates_project_md_in_target_dir — happy path with explicit PATH; "Created" message; file lands at <dir>/PROJECT.md.
  • test_refuses_to_overwrite_by_default — pre-existing PROJECT.md survives untouched; exit code non-zero; output mentions --force.
  • test_force_overwrites_existing — --force replaces the file; "Overwrote" message; old content gone; new content carries the template header.
  • test_force_short_flag — -f short form works the same as --force.
  • test_template_covers_expected_sections — guards the UX of the starter: all six headings must be present (regression catcher for an accidental template gut).
  • test_no_path_writes_to_cwd — without an argument, writes to Path.cwd() (monkeypatched to tmp_path because Typer's CliRunner doesn't chdir).
  • TestTopLevel::test_help_lists_commands updated to include init in the expected command list.

Test count: 549 → 555 (+6). All gates green.

Phase 2 Slice 3 — "Always allow" session allowlist

  • quoriv.permissions.allowlist — new module: SessionAllowlist, an in-memory set of tool names the user has promoted from a one-time HITL approval to a session-persistent one. __contains__, allow, clear, __len__, tools() returning an immutable frozenset snapshot. Keyed by tool name only (matches the granularity interrupt_on= itself uses).
  • quoriv.ui.prompts.DecisionType gained "approve_always". The interactive prompt now reads approve / reject / always [a/r/A]. parse_choice accepts A, aa, and the spelled-out always (case-insensitive) for the new decision; bare lowercase a still means "approve once" so a user can't accidentally promote a tool by typing the same key as before.
  • quoriv.app._collect_decisions consults the allowlist before calling prompt_approval: matching tools auto-resolve to approve with a [dim]auto-approved <tool> (allowlisted this session)[/dim] note. When the user picks approve_always, the tool name is added to the allowlist and a [green]Will auto-approve …[/green] confirmation is rendered. auto_deny (read-only mode) always wins over the allowlist — a remembered approval doesn't unlock read-only.
  • quoriv.app._decision_payload maps approve_always → {"type": "approve"} on the wire. DeepAgents only speaks approve / reject / edit / respond; the allowlist promotion is a Quoriv UX layer.
  • quoriv.app._interactive_loop creates one SessionAllowlist per run_chat invocation and threads it through _drive_turn → _collect_decisions. Survives /clear (the user promoted these tools deliberately; rotating the thread shouldn't silently un-promote them).
  • 21 new tests:
  • tests/unit/permissions/test_allowlist.py::TestSessionAllowlist (7) — empty default, allow adds, idempotency, __contains__ tolerates non-strings, tools() returns an immutable snapshot, clear, independent multi-tool tracking.
  • tests/unit/ui/test_prompts.py::TestParseChoice::test_approve_always_aliases (1 parametrized × 6 inputs) — covers A, aa, always, Always, ALWAYS, and whitespace-padded forms. The test_approve_aliases parametrize lost the bare "A" entry, since capital A is now reserved for approve_always; the comment in-place documents the split.
  • tests/unit/test_app_decisions.py (9) — TestDecisionPayload (3): approve passthrough, reject keeps message, approve_always → approve. TestCollectDecisionsAllowlist (6): allowlisted tool skips prompt; non-allowlisted still prompts; approve_always promotes; second call uses the promoted entry (single prompt across two calls); auto_deny wins over allowlist; None allowlist preserves legacy "always prompt" behavior.

Test count: 555 → 576 (+21). All gates green.

Phase 2 Slice 4 — Per-task model routing (built-in subagents)

  • quoriv.core.subagents — new module: three built-in SubAgent specs (researcher, debugger, reviewer) with fixed system prompts and a build_subagents(config) helper that resolves each role's configured model token ("default" / "fast" / "strong" / a literal "provider:name") into a model instance via quoriv.models.factory.get_model. Going through the Quoriv factory rather than letting DeepAgents call init_chat_model directly keeps the keychain-aware key lookup consistent across the main agent and every subagent.
  • quoriv.config.schema — new SubAgentRoleConfig (model: str = "default") and SubAgentsConfig (researcher / debugger / reviewer, with defaults "fast" / "strong" / "strong"). extra="forbid" on both — invented role names and stray fields fail fast at validation. Added to QuorivConfig as the subagents section.
  • quoriv.core.agent.build_agent now calls build_subagents(config) and passes the result to create_deep_agent(subagents=...). Returned shapes are typed as DeepAgents' SubAgent TypedDict via cast (imported under TYPE_CHECKING so the runtime import surface stays small).
  • config.example.toml documents the new [subagents.*] blocks with the token taxonomy and per-role defaults.
  • 16 new tests (and 1 integration-test fix):
  • tests/unit/core/test_subagents.py::TestResolveModelToken (5) — default / fast / strong resolve through [model]; literal provider:name passes through; overridden [model] section flows to all tokens.
  • tests/unit/core/test_subagents.py::TestBuildSubagents (6) — three roles in fixed order; every role carries name/description/system_prompt/model; researcher uses model.fast by default, debugger/reviewer use model.strong (verified by monkeypatching quoriv.core.subagents.get_model and capturing the requested ids); user can redirect a role to a literal model; user can redirect a role to "default"; role descriptions mention their job (the only signal the main agent has when routing).
  • tests/unit/core/test_subagents.py::TestSubAgentsConfigSchema (4) — default routing, partial-override preserves other roles, unknown role rejected, extra-field-within-role rejected.
  • tests/unit/core/test_subagents.py::TestBuildAgentSubagentsWiring (1) — build_agent passes a list of three subagents named [researcher, debugger, reviewer] into create_deep_agent (uses the same create_deep_agent capture pattern as the memory-wiring test).
  • tests/integration/test_e2e_stubbed_chat.py updated to monkeypatch both quoriv.core.agent.get_model and quoriv.core.subagents.get_model, since each subagent now resolves its own model. Existing 4 integration tests still pass unchanged.

Test count: 576 → 592 (+16). All gates green.

Phase 2 Slice 5 — Python plugin API (entry-point loader)

  • quoriv.plugins.loader — new module. Third-party packages register tools by declaring an entry point under the quoriv.plugins group ([project.entry-points."quoriv.plugins"] in their pyproject.toml). The named callable takes no arguments and returns an iterable of LangChain BaseTool instances. discover_plugin_tools(disabled=...) is called from build_agent at session start and merges the returned tools after QUORIV_TOOLS into create_deep_agent(tools=...). list_plugins() exposes a PluginRecord per entry point (name, target, captured tool names, load error) for introspection.
  • Defensive on purpose: an entry point that fails to import, a factory that raises, a non-iterable return, or a mixed list with non-tool items each logs a loguru warning and is dropped rather than failing the session. Users should be able to start a chat even if one plugin's tree is currently broken.
  • quoriv.config.schema.PluginsConfig (disabled: list[str], extra="forbid") added to QuorivConfig. Lets a user opt a specific plugin out without uninstalling its package — match by entry-point name.
  • quoriv.core.agent.build_agent calls discover_plugin_tools(disabled=config.plugins.disabled) and merges the result after QUORIV_TOOLS before handing the list to create_deep_agent(tools=...).
  • config.example.toml documents the new [plugins] block with a worked example showing the pyproject.toml snippet a plugin author writes.
  • 19 new tests:
  • tests/unit/plugins/test_loader.py::TestDiscoverPluginTools (4) — empty registry → empty list; one plugin → one tool; tools from multiple plugins merge; generator-returning factory normalises to a list.
  • tests/unit/plugins/test_loader.py::TestDisabledFiltering (3) — disabled list filters by entry-point name; disabled accepts any iterable (set works); empty disabled loads everything.
  • tests/unit/plugins/test_loader.py::TestDefensiveLoading (4) — import error → working plugin still loads; factory raising → working plugin still loads; non-iterable return → silently dropped; mixed valid+invalid items → valid kept.
  • tests/unit/plugins/test_loader.py::TestListPlugins (3) — one record per entry point; load failure captured with error string; factory failure captured with error string.
  • tests/unit/config/test_schema.py::TestPluginsConfig (5) — empty default; QuorivConfig.plugins.disabled == [] on empty input; explicit disabled list round-trips; extra="forbid"; full round-trip through QuorivConfig.model_validate.

Test count: 592 → 611 (+19). All gates green.

Phase 2 Slice 6 — MCP client (Model Context Protocol)

  • quoriv.config.schema.MCPServerConfig (+MCPConfig) — new sections. Each server picks a transport (stdio or sse, default stdio) and the per-transport fields: command/args/env for stdio, url/headers for sse. A model_validator enforces both the "required field per transport" rule (stdio needs command; sse needs url) and the "no cross-transport fields" rule — a stdio server with a url is clearly a typo and fails loudly. extra="forbid" throughout. Added to QuorivConfig as the mcp section, with mcp.servers: dict[str, MCPServerConfig] keyed by user-supplied server name.
  • quoriv.plugins.mcp.client — new module. load_mcp_tools(servers) is an async loader: translates each MCPServerConfig into the connection dict shape langchain_mcp_adapters.client.MultiServerMCPClient expects, instantiates the client, and calls await client.get_tools(). The lazy import for langchain_mcp_adapters is wrapped in try/except ImportError so users without the [mcp] install extra get a logged warning and an empty list rather than a hard crash. Every other failure path (client construction, get_tools()) is also caught + logged + returns [] — defensive throughout, same pattern as the entry-point plugin loader.
  • quoriv.plugins.mcp.__init__ re-exports load_mcp_tools.
  • quoriv.core.agent.build_agent gained a keyword-only extra_tools: list[Any] | None = None parameter. The tools list passed to create_deep_agent is now [*QUORIV_TOOLS, *plugin_tools, *extra_tools] — MCP-discovered tools land last so any future tool-shadow rule (a user override) stays explicit by position.
  • quoriv.app.run_chat calls await load_mcp_tools(config.mcp.servers) before constructing the agent and threads the result through _interactive_loop → build_agent and the live /mode-rebuild call path. The async load happens in run_chat (the only async point in the chat startup) precisely because build_agent stays sync.
  • pyproject.toml [mcp] install extra adds langchain-mcp-adapters>=0.2.0 alongside the existing mcp>=1.0.0.
  • config.example.toml documents the new [mcp.servers.NAME] blocks with both stdio and sse worked examples (the reference mcp-server-fetch stdio server and a hypothetical GitHub SSE endpoint with a bearer token).
  • 27 new tests:
  • tests/unit/plugins/test_mcp_schema.py::TestMCPServerConfigStdio (5) — minimal stdio, args + env, default transport, missing command rejected, cross-transport url rejected.
  • tests/unit/plugins/test_mcp_schema.py::TestMCPServerConfigSse (4) — minimal sse, headers, missing url rejected, cross-transport command rejected.
  • tests/unit/plugins/test_mcp_schema.py::TestMCPServerConfigStrict (2) — extra="forbid" and unknown transport rejection.
  • tests/unit/plugins/test_mcp_schema.py::TestMCPConfig (4) — empty default; QuorivConfig.mcp.servers == {} round-trip; mixed stdio + sse servers map round-trip; extra="forbid" at the MCPConfig level.
  • tests/unit/plugins/test_mcp_client.py::TestConnectionDict (5) — stdio minimal, env included when set, env omitted (not None) when not set, sse minimal, headers included when set.
  • tests/unit/plugins/test_mcp_client.py::TestLoadMcpToolsEdgeCases (2) — empty servers map → empty list (fast-path, no import); missing langchain-mcp-adapters (simulated by monkeypatching builtins.__import__) → logged warning + empty list.
  • tests/unit/plugins/test_mcp_client.py::TestLoadMcpToolsHappyPath (2) — monkeypatched MultiServerMCPClient captures the translated connection dict (verifies fetch stdio + github sse round-trip correctly); discovered tools flow back through.
  • tests/unit/plugins/test_mcp_client.py::TestLoadMcpToolsFailures (2) — MultiServerMCPClient.__init__ raising → empty list; await client.get_tools() raising → empty list.
  • tests/unit/core/test_agent.py::TestBuildAgentMemoryWiring::test_extra_tools_appended_to_tool_list (1) — extra_tools lands at the end of the tools list passed to create_deep_agent (after QUORIV_TOOLS and plugin tools). This is the same create_deep_agent capture pattern used for the memory tests.

Test count: 611 → 638 (+27). All gates green. Phase 2 is complete.

Phase 3 Slice 1 — Anthropic provider

  • quoriv.models.anthropic — new provider module mirroring quoriv.models.openai. Builds a langchain_anthropic.ChatAnthropic instance with model=spec.name, resolves the API key through quoriv.config.keychain.get_api_key("anthropic") (env var ANTHROPIC_API_KEY first, then OS keychain), raises MissingAPIKeyError with the correct env-var name when both are absent. Requires the [anthropic] install extra.
  • quoriv.models.factory._PROVIDERS registers anthropic. list_providers() now returns ["anthropic", "openai"] (sorted). The four remaining Phase 3 providers (Gemini / Ollama / vLLM / OpenRouter) stay commented out — they ship in later slices.
  • CI install line in .github/workflows/test.yml and the mypy job in .github/workflows/lint.yml add anthropic to the extras: pip install -e ".[dev,ast,mcp,anthropic]". Same pattern as the Slice 6 fix that pulled in [mcp] — keeps the test suite and mypy able to import the new provider's deps.
  • 7 new tests:
  • tests/unit/models/test_anthropic.py::TestMissingKey (1) — provider + env-var name on the raised MissingAPIKeyError.
  • tests/unit/models/test_anthropic.py::TestBuildFromEnv (1) — ANTHROPIC_API_KEY env var → ChatAnthropic with the model name preserved.
  • tests/unit/models/test_anthropic.py::TestBuildFromKeyring (1) — keychain fallback works when env is unset.
  • tests/unit/models/test_anthropic.py::TestIdentifierShapes (1) — modern dash-versioned ids like anthropic:claude-opus-4-7 flow through ModelSpec.parse unchanged.
  • tests/unit/models/test_anthropic.py::TestKwargsForwarded (2) — temperature and max_tokens forwarded to the underlying ChatAnthropic (with max_tokens_to_sample fallback for older lib versions).
  • tests/unit/models/test_factory.py::TestListProviders::test_anthropic_registered_in_phase_3 (1) — sanity that the factory dispatch picks it up.

Test count: 638 → 645 (+7). All gates green.

Phase 3 Slice 2 — Ollama provider (local, no API key)

  • quoriv.models.ollama — new provider module. Builds langchain_ollama.ChatOllama with model=spec.name. Unlike the OpenAI / Anthropic providers, Ollama runs locally (or on a user-controlled host), so no API key is required — the provider is intentionally absent from PROVIDER_ENV_VARS. No network call at construction time, so build_agent can succeed even when the Ollama server isn't running yet (connection errors surface at first invocation, not at startup). Requires the [ollama] install extra.
  • quoriv.models.factory._PROVIDERS registers ollama. list_providers() now returns ["anthropic", "ollama", "openai"] (sorted).
  • The CI install line in .github/workflows/test.yml and the mypy job in .github/workflows/lint.yml add ollama to the extras: pip install -e ".[dev,ast,mcp,anthropic,ollama]". Same pattern as P3-1.
  • Identifier shape preserved through ModelSpec.parse: Ollama tags carry an embedded colon (qwen2.5-coder:32b), and the parser splits on the first colon only — already documented and tested in test_base.py, now re-asserted at the provider level for qwen2.5-coder:32b and llama3.1:70b-instruct-q4_0.
  • 7 new tests:
  • tests/unit/models/test_ollama.py::TestBuild (2) — returns ChatOllama with the model name preserved; no network call at construction (build succeeds for an unpulled model).
  • tests/unit/models/test_ollama.py::TestModelNameWithTag (2) — second colon (the tag) stays attached to spec.name; real-world llama3.1:70b-instruct-q4_0 shape round-trips.
  • tests/unit/models/test_ollama.py::TestKwargsForwarded (2) — base_url forwarded (for non-default hosts like Docker / remote boxes), temperature forwarded.
  • tests/unit/models/test_factory.py::TestListProviders::test_ollama_registered_in_phase_3 (1) — factory dispatch picks it up.

Test count: 645 → 652 (+7). All gates green.

Phase 3 Slice 3 — Google Gemini provider

  • quoriv.models.gemini — new provider module. Builds langchain_google_genai.ChatGoogleGenerativeAI with model=spec.name and google_api_key=... (the LangChain wrapper's chosen kwarg name; differs from api_key used by OpenAI / Anthropic). Resolves the API key through get_api_key("gemini") — quoriv.config.keychain already routes both gemini and google provider names to the GOOGLE_API_KEY env var, so the precedence rules match the other cloud providers. Raises MissingAPIKeyError("gemini", "GOOGLE_API_KEY") when both env and keychain are absent. Requires the [gemini] install extra.
  • quoriv.models.factory._PROVIDERS registers gemini. list_providers() now returns ["anthropic", "gemini", "ollama", "openai"] (sorted).
  • CI install line in .github/workflows/test.yml and the mypy job in .github/workflows/lint.yml add gemini to the extras: pip install -e ".[dev,ast,mcp,anthropic,ollama,gemini]".
  • 7 new tests:
  • tests/unit/models/test_gemini.py::TestMissingKey (1) — MissingAPIKeyError("gemini", "GOOGLE_API_KEY") shape.
  • tests/unit/models/test_gemini.py::TestBuildFromEnv (2) — env-var key resolves to a ChatGoogleGenerativeAI with the model preserved; gemini-1.5-pro round-trips alongside gemini-1.5-flash.
  • tests/unit/models/test_gemini.py::TestBuildFromKeyring (1) — keychain fallback when env is unset.
  • tests/unit/models/test_gemini.py::TestKwargsForwarded (2) — temperature forwarded; max_output_tokens forwarded (note: differs from OpenAI's max_tokens / Anthropic's max_tokens_to_sample — Gemini uses its own name).
  • tests/unit/models/test_factory.py::TestListProviders::test_gemini_registered_in_phase_3 (1) — factory dispatch picks it up.

Test count: 652 → 659 (+7). All gates green.

Phase 3 Slice 4 — vLLM provider (OpenAI-compatible)

  • quoriv.models.vllm — new provider module. vLLM serves an OpenAI-compatible HTTP API, so under the hood the provider builds a langchain_openai.ChatOpenAI instance pointed at the user's vLLM endpoint via base_url. No extra SDK needed (uses the OpenAI SDK already in core deps).
  • Defaults that match local-first deployment:
  • base_url: explicit kwarg > $VLLM_BASE_URL env var > http://localhost:8000/v1 (the vLLM server's default OpenAI-compatible endpoint).
  • api_key: explicit kwarg > $VLLM_API_KEY env (already in PROVIDER_ENV_VARS) > keychain > placeholder "EMPTY". Unlike OpenAI / Anthropic / Gemini, vLLM never raises MissingAPIKeyError — most local vLLM deployments don't enforce auth and ChatOpenAI accepts any string as api_key, so falling through to a placeholder keeps the provider building cleanly. The keychain map already lists vllm.
  • quoriv.models.factory._PROVIDERS registers vllm. list_providers() now returns ["anthropic", "gemini", "ollama", "openai", "vllm"] (sorted).
  • No CI extras change needed — vLLM uses the OpenAI SDK already in the core dependency list.
  • 10 new tests:
  • tests/unit/models/test_vllm.py::TestDefaults (3) — returns ChatOpenAI with model preserved; default base_url applied when nothing is configured; placeholder "EMPTY" api_key when neither env nor keychain has one (read via SecretStr.get_secret_value() since LangChain stores it secret-style).
  • tests/unit/models/test_vllm.py::TestEnvVarPrecedence (3) — $VLLM_BASE_URL overrides the built-in default; $VLLM_API_KEY is honored; keychain wins when env is absent.
  • tests/unit/models/test_vllm.py::TestKwargOverrides (2) — explicit base_url= kwarg beats $VLLM_BASE_URL; explicit api_key= kwarg beats $VLLM_API_KEY.
  • tests/unit/models/test_vllm.py::TestKwargsForwarded (1) — additional kwargs like temperature flow through to ChatOpenAI.
  • tests/unit/models/test_factory.py::TestListProviders::test_vllm_registered_in_phase_3 (1) — factory dispatch picks it up.

Test count: 659 → 669 (+10). All gates green.

Phase 3 Slice 5 — OpenRouter provider

  • quoriv.models.openrouter — new provider module. OpenRouter routes hundreds of models from different vendors through a single OpenAI-compatible API, so the provider builds a langchain_openai.ChatOpenAI instance pointed at https://openrouter.ai/api/v1 (overridable via base_url= kwarg for OpenRouter proxies / staging endpoints). API key resolved via get_api_key("openrouter") (env OPENROUTER_API_KEY first, then keychain — both already in PROVIDER_ENV_VARS). Unlike vLLM, the key is required — OpenRouter is a paid service — so MissingAPIKeyError is raised when both sources are empty.
  • Identifier shape: openrouter:<vendor>/<model> (e.g. openrouter:anthropic/claude-3.5-sonnet, openrouter:meta-llama/llama-3.1-405b-instruct). The vendor/model slash is part of the name half; ModelSpec.parse preserves it because the split is on the first colon only.
  • quoriv.models.factory._PROVIDERS registers openrouter. list_providers() now returns ["anthropic", "gemini", "ollama", "openai", "openrouter", "vllm"] (sorted) — all 6 Phase 3 providers wired.
  • No CI extras change — OpenRouter uses the OpenAI SDK already in core dependencies.
  • 8 new tests:
  • tests/unit/models/test_openrouter.py::TestMissingKey (1) — MissingAPIKeyError("openrouter", "OPENROUTER_API_KEY") shape.
  • tests/unit/models/test_openrouter.py::TestBuildFromEnv (2) — env-var key resolves to ChatOpenAI with model preserved; default base_url points at OpenRouter's cloud endpoint.
  • tests/unit/models/test_openrouter.py::TestBuildFromKeyring (1) — keychain fallback when env is unset.
  • tests/unit/models/test_openrouter.py::TestIdentifierShapes (1) — vendor/model slash survives through ModelSpec.parse and reaches ChatOpenAI.model_name.
  • tests/unit/models/test_openrouter.py::TestKwargOverrides (2) — explicit base_url= overrides the default; temperature forwards through.
  • tests/unit/models/test_factory.py::TestListProviders::test_openrouter_registered_in_phase_3 (1) — factory dispatch picks it up.

Test count: 669 → 677 (+8). All gates green. All 6 Phase 3 providers (Anthropic, Ollama, Gemini, vLLM, OpenRouter, plus the pre-existing OpenAI) are now wired.

Phase 3 Slice 6 — web_fetch tool

  • quoriv.tools.web — new module shipping web_fetch(url, max_chars): a small @tool that wraps httpx.Client.get so the agent can pull text from a URL during a turn. Returns a structured dict (status_code, content_type, text, truncated, url) on success or {"error": ..., "url": ...} on network failure — mirrors the dict[str, Any] + "error" key contract every other Quoriv tool uses.
  • Bounded output: max_chars defaults to 10_000 (~2.5-3k tokens) — enough for most pages without blowing the model's context window. Larger bodies get truncated with an explicit "… (truncated, +N chars)" marker so the agent knows more content exists. Pages at the exact cap are not marked truncated.
  • Defensive: httpx.HTTPError (covers connect failures, DNS, timeouts, redirect-cap exceeded) is caught and turned into the error dict — a bad URL never propagates an exception up through the tool call. follow_redirects=True and timeout=30s are baked in.
  • Sync, not async — matches every other Quoriv tool. The agent's ToolExecutor runs it in a thread pool if needed.
  • Registered as the 12th entry in QUORIV_TOOLS. quoriv.tools.__init__ re-exports web_fetch.
  • web_search is intentionally deferred to a later slice — picking the search backend (Tavily, Brave, SerpAPI, …) deserves its own scope; shipping the fetch tool first keeps this slice self-contained and free of new API-key dependencies.
  • 11 new tests in tests/unit/tools/test_web.py:
  • TestWebFetchHappyPath (4) — expected dict keys, short body untouched, status_code + content_type round-trip, URL forwarded to the client.
  • TestTruncation (3) — body past cap truncated with +N chars marker, body at exact cap not truncated, default cap applies when not specified.
  • TestNetworkFailure (2) — httpx.ConnectError and httpx.ReadTimeout both surface as error dicts with the URL preserved and no text/status_code fields.
  • TestToolSurface (2) — web_fetch lives in QUORIV_TOOLS so the agent sees it; description mentions "fetch" + "url" so the model can route to it.

Test count: 677 → 688 (+11). All gates green.

Phase 3 Slice 7 — web_search tool (Tavily backend)

  • quoriv.tools.web.web_search(query, max_results, include_domains, exclude_domains) — new @tool that hits Tavily's LLM-friendly search API and returns ranked result snippets. Each result row is normalised to {title, url, content, score} — Tavily's heavier raw_content field is intentionally dropped so the list stays compact in the agent's context.
  • Defensive everywhere: the tavily-python SDK lives in a new [search] install extra. A user without the extra still gets a working session; the tool just returns {"error": "Tavily SDK not installed (...)..", "query": ...}. Same shape if the TAVILY_API_KEY is missing or if the upstream API raises — the agent should be able to recover instead of dying mid-turn.
  • quoriv.config.keychain.PROVIDER_ENV_VARS gains a tavily → TAVILY_API_KEY entry. Key resolution flows through the existing get_api_key("tavily") helper (env var first, then keychain) — same precedence as the model providers.
  • pyproject.toml adds a new [search] install extra: search = ["tavily-python>=0.7.0"]. CI install lines in test.yml and the mypy job in lint.yml add search to the extras: pip install -e ".[dev,ast,mcp,anthropic,ollama,gemini,search]". The tavily.* module pattern is added to mypy's ignore_missing_imports overrides since the SDK ships without py.typed.
  • quoriv.tools.__init__ re-exports web_search and registers it as the 13th entry in QUORIV_TOOLS.
  • Annotation note: the tool's include_domains / exclude_domains use list[str] | None rather than Sequence[str] | None. LangChain's @tool decorator evaluates annotations at runtime via get_type_hints(), so Sequence (importable only from collections.abc) would either need to live at module top-level and draw a ruff TC003 warning, or sit under TYPE_CHECKING and break tool registration. list is a builtin, which sidesteps both problems.
  • 7 new tests in tests/unit/tools/test_web.py:
  • TestWebSearchMissingKey (1) — empty env + empty keychain produces a structured error dict (not an exception), and the error string names TAVILY_API_KEY.
  • TestWebSearchHappyPath (4) — env-var key resolves and returns normalised results; keychain fallback when env is unset (fake_keyring captures the key the fake client was constructed with); max_results forwards; include_domains + exclude_domains forward through to the SDK call.
  • TestWebSearchFailure (2) — Tavily client raising at search time becomes {"error": "Search failed: ..."}; non-dict responses from the SDK fall through to an empty results list (defensive guard against weirdly shaped upstream returns).
  • Existing TestToolSurface::test_registered_in_quoriv_tools extended to also assert web_search is in QUORIV_TOOLS.

Test count: 688 → 695 (+7). All gates green.

Phase 3 Slice 8 — Themes (light / dark / auto)

  • quoriv.ui.themes — new module exposing three palettes for the config.ui.theme setting that's been on the schema since Phase 0 Day 2 but until now never propagated past the doctor table. make_console(theme) returns a configured rich.console.Console; resolve_theme(name) collapses auto to a concrete dark / light based on the terminal's $COLORFGBG background hint (indexes 7 and 15 = light); RICH_THEMES["dark"] is explicitly None so the factory has no special-case branch for "use Rich defaults".
  • Palette content kept small on purpose: the only Rich styles overridden by light are dim (Rich's default low-contrast grey on dark becomes near-invisible on white — bumped to grey39) and panel.border (cyan washes out on light, switched to blue). Inline color tags like [green]ok[/green] and [red]err[/red] keep their Rich-default rendering because their contrast holds across both backgrounds.
  • quoriv.app.run_chat and quoriv.cli._console now call make_console(config.ui.theme) instead of bare Console(). The CLI helper wraps the lazy load_config call in try/except so a malformed config still gets a working console (falls back to auto).
  • The plain from rich.console import Console imports in app.py and cli.py move into TYPE_CHECKING blocks now that the runtime entry point is make_console — Console is only used as a type annotation in both files, so the runtime import is dead weight (ruff TC002).
  • 19 new tests in tests/unit/ui/test_themes.py:
  • TestPaletteMap (2) — dark is explicitly None, light is a real rich.theme.Theme instance.
  • TestLightBackgroundDetection (5) — no $COLORFGBG → no light detection; 0;7 and 0;15 → light; 15;0 / 15;1 / 15;8 → not light; malformed values fall through; a single field without the fg;bg shape isn't enough signal.
  • TestResolveTheme (4) — dark and light round-trip; auto resolves to light with the light-bg env var, dark without it.
  • TestMakeConsole (5) — dark returns a plain Console; light carries the grey dim override (verified by reading the style off the resolved Console); auto without signal falls through to dark; force_terminal=False kwarg forwards; an unknown theme name silently falls back to defaults rather than crashing.

Test count: 695 → 714 (+19). All gates green.

Phase 3 Slice 9 — Fallback chains

  • quoriv.config.schema.ModelConfig.fallbacks: list[str] — new ordered list of provider:model identifiers to try when the primary model raises a transient error (rate limit, 5xx, network failure). Empty default disables fallbacks.
  • quoriv.models.factory.with_fallbacks(primary, fallback_ids) — builds each fallback via get_model and wraps the primary with LangChain's Runnable.with_fallbacks(...). Defensive: a fallback id that fails to build (missing key, unknown provider, malformed id) is logged and skipped rather than aborting agent startup. If every fallback fails, the primary is returned unwrapped (no chain at all) so the user still gets a working session.
  • quoriv.core.agent.build_agent wraps the primary model via with_fallbacks(primary_model, config.model.fallbacks) before passing it to create_deep_agent. The return type can be either BaseChatModel (no fallbacks) or RunnableWithFallbacks (chain assembled); a cast at the call site keeps mypy happy without weakening the public typing of with_fallbacks.
  • 6 new tests in tests/unit/models/test_fallbacks.py:
  • TestEmptyFallbacks (2) — empty iterable returns primary unchanged with no get_model calls; an all-failing fallback list also returns primary unchanged (every fallback raises at build time).
  • TestChainAssembly (2) — single fallback wraps primary via .with_fallbacks(...) and get_model is called with the expected id; multiple fallbacks preserve user-supplied order all the way through.
  • TestPartialFailure (1) — middle fallback raises at build time, first and third still land in the chain (skip + continue, not abort).
  • TestIntegrationSmoke (1) — with_fallbacks against a real RunnableLambda primary returns a real RunnableWithFallbacks, catching future API drift in LangChain.

Test count: 714 → 720 (+6). All gates green.

Phase 3 Slice 10 — Hook registry (foundational)

  • quoriv.hooks — new module shipping HookRegistry, a tiny in-memory event bus for pre_tool / post_tool / on_message events. Callbacks are stored per-event in registration order, fire sequentially, and a callback that raises has its exception logged via loguru and dropped — one broken hook never breaks a turn. The registry is intentionally per-session (constructed once per run_chat), not a module-level singleton, so tests get clean isolation.
  • API surface: register(event, callback) (rejects unknown event names early so a typo doesn't silently disable instrumentation), fire(event, **kwargs) (forwards kwargs to every registered callback), handlers(event) (immutable snapshot tuple for /hooks-style introspection), clear() (test isolation).
  • Scope note: this slice ships the registry only. Wiring _stream_events to actually fire pre_tool / post_tool on on_tool_start / on_tool_end LangGraph events lands in a follow-up — the registry is a complete unit on its own (full test coverage, foundational shape locked in) and the integration is a separable concern that touches the chat-loop plumbing.
  • 10 new tests in tests/unit/test_hooks.py:
  • TestRegister (4) — empty registry has no handlers; register adds; multiple handlers preserve registration order when fired; unknown event names rejected with a ValueError naming the valid set.
  • TestFire (4) — kwargs forward verbatim; firing with no handlers is a silent no-op; firing an unknown event name is silent (programming error in the emitter, not user input); a callback that raises is logged and the remaining handlers still run.
  • TestClear (1) — clear() drops handlers across all three events.
  • TestHandlersSnapshot (1) — the returned tuple is a true snapshot; later register calls don't retroactively appear in held references.

Test count: 720 → 730 (+10). All gates green.

Phase 3 Slice 11 — Hooks integration (consumer side)

  • quoriv.app._stream_events now fires three events against the per-session HookRegistry shipped in Slice 10:
  • pre_tool on on_tool_start — kwargs: tool_name, args.
  • post_tool on on_tool_end — kwargs: tool_name, output.
  • on_message on on_chat_model_end — kwarg: message (the final AIMessage from LangChain).
  • run_chat constructs one HookRegistry() per session and threads it through _interactive_loop → _drive_turn → _stream_events as a new keyword-only hooks parameter. All three layers default hooks to None so existing test entry points (including the Phase 1 Slice 9b stubbed-LLM integration test) keep working unchanged.
  • The registry's defensive fire() swallows callback exceptions and logs them — combined with this integration, one broken hook still can't break a turn. The integration adds no failure modes that the registry doesn't already handle.
  • _stream_events picked up a # noqa: PLR0912 because adding three optional hook-fire branches pushed the flat event-kind dispatch from 12 to 15 branches. The function is intentionally a flat switch with one branch per LangGraph event kind, so splitting it would just hide the dispatch.
  • No new tests this slice — the registry has full unit coverage from Slice 10 (10 tests) and the wiring is three mechanical if hooks is not None: hooks.fire(...) lines. The existing 730-test suite catches regressions on every other code path that touches _stream_events.

Test count: 730 → 730 (no new tests; wiring-only slice). All gates green.

Phase 3 Slice 12 — Replay mode (post-mortem viewer)

  • quoriv.replay — new module shipping replay_thread(console, trace_path). Reads a per-thread JSONL trace via TraceLogger.read_events and renders each event with a Rich-formatted prefix and short label: ▶ turn_start (with user_input + mode), · model_complete (with model + input_tokens + output_tokens), ↳ tool_start (with tool_name + truncated args), ◀ tool_end (with tool_name + truncated output_preview), ■ turn_end. Unknown event kinds fall through to a compact JSON dump so nothing is silently dropped.
  • Read-only by construction: no model is invoked, no tool is executed, no permissions are checked. The viewer purely re-renders the JSONL log. Safe to point at any past session, including ones that ran in yolo mode against a paid provider.
  • quoriv replay <name-or-id> CLI command added. Resolves <name> through the per-cwd SessionRegistry (the same one /save / /load use); falls back to treating the argument as a raw thread id when no name matches. --cwd override mirrors the chat command. Missing-trace path raises typer.Exit(code=1) with a friendly nudge so a CI script catches the failure.
  • 9 new tests in tests/unit/test_replay.py:
  • TestFormatRecord (6) — one assertion per event kind (turn_start carries user_input, model_complete shows token counts, tool_start shows args, tool_end shows preview, turn_end renders minimally, unknown kinds fall back to JSON).
  • TestReplayThread (3) — missing file prints "No events" and returns 0; full trace renders every event with its fingerprint and reports the count in the header; malformed lines mid-file are silently dropped (the TraceLogger.read_events contract from Slice 9).

Test count: 730 → 739 (+9). All gates green.

Phase 3 Slice 13 — Eval harness foundation

  • quoriv.eval — new module shipping the pure scoring layer for a regression eval suite:
  • EvalCase(name, prompt, expected_substrings) — one prompt + the substrings the agent's final response must contain.
  • EvalResult(case_name, passed, failed_substrings, output) — what scoring produces.
  • score_case(case, output) -> EvalResult — pure function. Each expected substring must appear in the output (case-sensitive, unanchored). A case with no expectations always passes (smoke-test mode).
  • summarize(results) -> {"total", "passed", "failed"} and passed_fraction(results) -> float for CI thresholds.
  • SAMPLE_CASES — three minimal sanity cases (echo a distinctive token, basic arithmetic, file:line citation format) using assertions distinctive enough that random LLM drift is very unlikely to satisfy them by accident.
  • Scope note: this slice ships the scoring + bundled cases only. The runner that drives each case through _drive_turn against a chosen model is a follow-up — the scoring layer is a complete, foundational unit on its own. Builds on the stubbed-LLM integration pattern shipped in Phase 1 Slice 9b.
  • 16 new tests in tests/unit/test_eval.py:
  • TestScoreCase (5) — all substrings present passes; one missing fails with that substring listed; case-sensitive matching ("Hello" vs "hello"); empty expectations always pass (smoke mode); unanchored substring match.
  • TestSummarize (3) — empty / all-passing / mixed result lists.
  • TestPassedFraction (4) — empty is 1.0 (vacuous), all-pass / all-fail / half-pass.
  • TestSampleCases (4) — bundle non-empty; every case carries name + prompt + expectations; case names are unique (lookups don't collide); the assertions are satisfiable (synthetic outputs hitting each case's substrings pass).

Test count: 739 → 755 (+16). All gates green. Phase 3 is feature-complete — runner / CLI for the eval harness can ship as a follow-up enhancement.

Phase 4 Slice 1 — Telemetry opt-in scaffold

  • quoriv.config.schema.TelemetryConfig(enabled: bool = False, endpoint: str | None = None) — new top-level config section. Off by default. extra="forbid" keeps typos from silently flipping the flag. Added to QuorivConfig as the telemetry section.
  • quoriv.observability.telemetry — new module shipping the gating surface for Phase 4 telemetry:
  • is_enabled(config) — returns True only when the user explicitly opted in. Accepts a full QuorivConfig, a bare TelemetryConfig, or None (always disabled).
  • report(event_name, config=None, **fields) — no-op stub that short-circuits early when telemetry is disabled. When enabled, it logs to loguru at debug level for now — the network sink ships in a follow-up when we pick a backend (Sentry / PostHog / OpenTelemetry / self-hosted).
  • No backend, no transmission. Users who flip the flag in config.toml won't see traffic; the contract is "your opt-in is captured and respected; nothing transmits yet". When the backend lands, the gating check is already in place at every future call site.
  • quoriv.observability.__init__ re-exports telemetry_enabled and telemetry_report so emitters can from quoriv.observability import telemetry_report without reaching into the submodule.
  • 13 new tests in tests/unit/observability/test_telemetry.py:
  • TestTelemetryConfigDefaults (5) — enabled=False by default, endpoint=None by default, QuorivConfig.telemetry exposes the section with safe defaults, explicit opt-in round-trips through model_validate, extra fields rejected.
  • TestIsEnabled (4) — None config returns False, default config returns False, explicit enabled=True returns True, accepts a bare TelemetryConfig (callers that drilled down).
  • TestReport (4) — disabled config emits no log; None config no-ops without raising; enabled config doesn't raise (loguru sink is implementation detail); arbitrary structured kwargs accepted (future backend validates at the sink).

Test count: 755 → 768 (+13). All gates green.

Phase 4 Slice 2 — Eval-harness runner

  • quoriv.eval.run_case(case, *, config=None, cwd=None, mode="yolo", model_override=None, agent=None) — drives one EvalCase through quoriv.app._drive_turn against a fresh agent (or a caller-supplied one for tests), then extracts the final assistant text from LangGraph state via agent.aget_state(...) and feeds it to score_case. Per-case thread_id = f"eval-{case.name}" keeps cases isolated in the checkpointer. Discards rendered output to an in-memory Console so eval runs quiet.
  • quoriv.eval.run_suite(cases, ...) — sequential runner over a tuple/list of cases. A per-case exception is captured as a failed EvalResult (output = <error: ...>, every expected substring marked missing) so one bad case can't poison the rest of the run. Case order preserved in the returned list.
  • quoriv.eval._final_ai_text(messages) — walks a LangGraph message list in reverse looking for the latest non-empty AIMessage. Handles both string content and the list-of-dict-chunks shape some providers emit. Returns "" when no suitable message is found (score then reports every expected substring as missing).
  • quoriv eval [--model] [--cwd] — new Typer command. Resolves the model from config (with --model override), runs SAMPLE_CASES via run_suite, prints a Rich table summarising pass / fail per case + missing substrings, and exits non-zero if any case failed.
  • 18 new tests in tests/unit/test_eval_runner.py:
  • TestFinalAIText (5) — string content, list-of-dict-chunks, skips empty AIMessages, no-AIMessage list, empty list.
  • TestRunCase (5) — passing case, failing case (missing substrings reported), per-case thread id isolation, requires config or agent, smoke case with empty expectations.
  • TestRunSuite (3) — one result per case in order, per-case exception isolation (stubbed _drive_turn raise), empty suite returns empty list.
  • TestEvalCLI (5) — zero exit on all-pass, non-zero exit on any-fail, case names rendered in table, --model forwarded as model_override, --cwd forwarded as resolved Path.
  • Uses the same GenericFakeChatModel + bind_tools override pattern as tests/integration/test_e2e_stubbed_chat.py — no real LLM calls in the test suite.

Test count: 768 → 786 (+18). All gates green.

Phase 4 Slice 3 — Release CI workflow

  • .github/workflows/release.yml — tag-triggered release pipeline:
  • build job — checks out the repo, installs build + twine, verifies the pushed tag matches project.version in pyproject.toml (a stale tag now fails loudly instead of publishing the wrong artifact), builds wheel + sdist via python -m build (hatchling backend), validates the dists with twine check, and uploads them as a dist artifact.
  • publish job — needs: build, gated to push events on refs/tags/v* so a workflow_dispatch build doesn't accidentally publish. Uses pypa/gh-action-pypi-publish@release/v1 with OIDC trusted publishing (permissions: id-token: write) — no PyPI API token is stored in repo secrets. Bound to the pypi deployment environment so manual approval can be required before each release.
  • github-release job — downloads the dist artifact and attaches the wheel + sdist to the GitHub release page via softprops/action-gh-release@v2, auto-generating release notes from the changelog and commit history.
  • Trigger: push on tags matching v*.*.* (semver only). workflow_dispatch is also wired so the build job can be smoke-tested without cutting a tag.
  • 18 new tests in tests/unit/test_release_workflow.py:
  • TestWorkflowFile (2) — file exists, parses as YAML with the right name.
  • TestTriggers (2) — triggers on v*.*.* tags, supports workflow_dispatch. Handles PyYAML's quirk of mapping the bare on: key to Python True.
  • TestBuildJob (5) — build job exists, uses python -m build, validates with twine check, verifies tag matches pyproject.toml version, uploads the dist artifact.
  • TestPublishJob (6) — publish job exists, has id-token: write for OIDC, uses pypa/gh-action-pypi-publish, needs: build, gated to tag pushes (won't publish from workflow_dispatch), targets the pypi environment.
  • TestGithubReleaseJob (3) — exists, uses softprops/action-gh-release, has contents: write.
  • One-time setup before first release — configure trusted publisher on PyPI at https://pypi.org/manage/account/publishing/ binding the (repository, workflow name, environment) triple. After that, every git tag v… && git push --tags ships a release.

Test count: 786 → 804 (+18). All gates green.

Phase 4 Slice 4 — MkDocs documentation site

  • mkdocs.yml — mkdocs-material config with light/dark palette toggle, navigation features (instant, tracking, top-button, sections), search, code-copy buttons, and "Edit this page" → GitHub. Plugins: search, include-markdown, mkdocstrings[python] (pointed at src/). Markdown extensions cover admonitions, code highlighting, tables, tabbed blocks, task lists, footnotes, and pymdownx emoji.
  • docs/index.md, docs/changelog.md, docs/contributing.md, docs/security.md, docs/project-plan.md — thin shims that pull README.md / CHANGELOG.md / CONTRIBUTING.md / SECURITY.md / PROJECT_PLAN.md from the repo root via the include-markdown directive. Avoids duplicating content; root files stay canonical for GitHub viewers.
  • docs/architecture.md — short intro + includes the existing docs/DEEPAGENTS_REFERENCE.md so the architecture reference is available in the rendered site.
  • not_in_nav: DEEPAGENTS_REFERENCE.md — silences the orphaned-file warning for the reference doc, which is consumed via include-markdown rather than linked directly.
  • .github/workflows/docs.yml — GitHub Pages deploy workflow:
  • Triggers on push to main (path-filtered to docs / source / root markdown) + workflow_dispatch.
  • build job — installs docs extras + the include-markdown plugin, runs mkdocs build, uploads the rendered site/ directory as a Pages artifact via actions/upload-pages-artifact@v3.
  • deploy job — gated to refs/heads/main, uses actions/deploy-pages@v4 with OIDC (pages: write + id-token: write). Bound to the github-pages environment so the URL is surfaced on the workflow page.
  • Non-strict build for now — included root files carry pre-existing relative links (../PROJECT_PLAN.md, ../LICENSE) that MkDocs flags as warnings. Site renders correctly; tightening to --strict happens once those links are rewritten to be site-aware.
  • pyproject.toml [project.optional-dependencies.docs] — added mkdocs-include-markdown-plugin>=6.2.0 so pip install -e ".[docs]" produces a working mkdocs build.
  • 24 new tests in tests/unit/test_docs_setup.py:
  • TestMkdocsConfig (6) — file exists, site_name, repo_url / edit_uri, material theme, include-markdown plugin enabled, search plugin enabled. A custom SafeLoader subclass stubs out !!python/name: tags so the test runs without mkdocs-material being installed.
  • TestNavStructure (2 + 6 parametrized) — nav present, core pages mapped, every nav target resolves to an actual file under docs/.
  • TestDocsWorkflow (9) — workflow file present, triggers on main push + workflow_dispatch, has pages: write + id-token: write + contents: read, runs mkdocs build, uploads pages artifact via actions/upload-pages-artifact, deploys via actions/deploy-pages, deploy gated to main, targets the github-pages environment.
  • TestPyprojectDocsGroup (1) — mkdocs-include-markdown-plugin and mkdocs-material are listed in the docs extra.
  • One-time setup before first deploy — repo Settings → Pages → Source: "GitHub Actions". After that, every push to main updates https://burhanhussain1.github.io/quoriv/.

Test count: 804 → 828 (+24). All gates green.

Phase 4 Slice 5 — PyInstaller binaries

  • pyinstaller.spec — spec file used by pyinstaller pyinstaller.spec (CI invokes this on each matrix runner):
  • Entry point: src/quoriv/__main__.py with pathex=["src"] so the quoriv package resolves under PyInstaller's bootloader.
  • Explicit hiddenimports for every provider module under quoriv.models.* — they're loaded via importlib.import_module in quoriv.models.factory and the static analyser doesn't see them. Also lists the LangChain provider packages each one imports from (langchain_openai, langchain_anthropic, langchain_google_genai, langchain_ollama) plus langchain.agents.factory and langchain_anthropic.middleware.prompt_caching (DeepAgents pulls these in dynamically).
  • collect_submodules covers quoriv + langchain + langchain_core + langgraph + deepagents so chains / tracers / hub modules that evade analysis still ship.
  • excludes: pytest / mypy / ruff / mkdocs — dev-only stuff that shouldn't bloat the binary.
  • upx=False deliberately — UPX-packed Windows binaries get flagged by Defender SmartScreen.
  • .github/workflows/binaries.yml — cross-platform release pipeline:
  • Triggers on v*.*.* tag pushes (release path) and workflow_dispatch (manual smoke build).
  • build job — 3-row matrix (ubuntu-latest, macos-latest, windows-latest) with artifact + binary columns so the smoke / upload steps stay declarative. Installs binary extra + runtime provider extras, runs pyinstaller pyinstaller.spec --clean --noconfirm, renames the produced binary to a per-OS name (quoriv-linux-x86_64, quoriv-macos-arm64, quoriv-windows-x86_64.exe), smoke-tests it by invoking version, and uploads as a workflow artifact.
  • attach-to-release job — needs: build, gated to push events on refs/tags/v*. Downloads all three matrix artifacts and attaches them to the GitHub release via softprops/action-gh-release@v2. Manual dispatches stop at the workflow-artifact stage so a smoke build never accidentally publishes.
  • pyproject.toml [project.optional-dependencies.binary] — new extra installing pyinstaller>=6.10.0. pip install -e ".[binary,…]" is now enough to build a binary locally.
  • 23 new tests in tests/unit/test_binaries_workflow.py:
  • TestSpecFile (6) — exists, entrypoint, collect_submodules("quoriv"), every provider listed in hiddenimports, console=True, upx=False.
  • TestWorkflowTriggers (2) — v*.*.* tag push + workflow_dispatch.
  • TestBuildMatrix (3 + 3 parametrized) — every OS present, each matrix row names artifact + binary, Windows row carries .exe suffix.
  • TestBuildSteps (4) — installs binary extra, invokes pyinstaller pyinstaller.spec, smoke-tests the binary with version, uploads per-OS artifact.
  • TestAttachToRelease (5) — job exists, needs: build, gated to tag pushes, contents: write, uses softprops/action-gh-release.
  • TestPyprojectBinaryExtra (1) — pyinstaller appears in the binary extra.
  • First release run will likely surface platform-specific missing hidden imports; expect a short iteration loop after tagging v0.1.0-rc1 to refine the spec before v1.0.0.

Test count: 828 → 851 (+23). All gates green.

Phase 4 Slice 6 — Telemetry backend

  • quoriv.observability.telemetry.report now actually transmits when the user has opted in and configured an endpoint. The Slice 1 stub (debug-log only) is replaced by a sync HTTP POST via httpx:
  • Envelope shape: {"event": <name>, "fields": {<kwargs>}, "client": {"name": "quoriv", "version": __version__, "platform": sys.platform, "python": "<maj.min>"}, "timestamp": <ISO-8601 UTC>}.
  • When api_key is set on the config, sent as Authorization: Bearer <api_key>. Lets self-hosted sinks distinguish clients.
  • Timeout is 2 seconds (exported as _DEFAULT_TIMEOUT) — a misbehaving sink can stall a chat turn by at most that amount, no matter what.
  • Every transport error is swallowed. httpx.ConnectError, DNS failures, 5xx responses, broken TLS — all caught, logged at debug, never re-raised. Telemetry must never break the agent.
  • quoriv.observability.telemetry._build_envelope — new pure helper exposing the envelope shape so tests (and any future backend swap) can assert on it without touching the network.
  • quoriv.observability.telemetry._resolve_telemetry — internal helper that pulls the bare TelemetryConfig out of either a QuorivConfig or a leaf TelemetryConfig. Keeps report symmetrical with is_enabled on accepted inputs.
  • quoriv.config.schema.TelemetryConfig.api_key: str | None = None — new field for the optional bearer token. extra="forbid" is unchanged so a typo (enabled → enable) still fails validation loudly.
  • Endpoint docstring rewritten — it's no longer "ignored until a backend ships"; it now drives the HTTP transport.
  • No new dependencies. httpx is already a runtime dep for tool-side use; we reuse it here.
  • 14 new tests in tests/unit/observability/test_telemetry.py:
  • TestApiKeyField (2) — default None, round-trips through model_validate.
  • TestBuildEnvelope (4) — event name + fields preserved, client metadata present, timestamp is ISO-8601 UTC (+00:00 suffix), envelope is JSON-serialisable.
  • TestReportTransport (8) — uses a _PostRecorder callable that monkeypatches httpx.post to capture call args without hitting the network:
    • No endpoint configured → no POST (even when enabled).
    • Disabled config → no POST (even when endpoint is set).
    • Enabled + endpoint → POST with correct URL, JSON envelope, Content-Type header.
    • api_key set → Authorization: Bearer <key> header added.
    • Timeout is bounded to _DEFAULT_TIMEOUT (≤ 5s safety check).
    • httpx.ConnectError swallowed (no re-raise).
    • 5xx response swallowed (no re-raise).
    • Bare TelemetryConfig accepted directly (no QuorivConfig wrapper required).

Test count: 851 → 865 (+14). All gates green.

Changed

Architecture revision (post-DeepAgents audit)

  • Adopted DeepAgents-reuse model after a deep read of the installed deepagents 0.6.1 SDK
  • Added docs/DEEPAGENTS_REFERENCE.md — internal working reference for every DeepAgents feature Quoriv builds on
  • Rewrote PROJECT_PLAN.md: updated folder tree, architecture diagram, Phase 1 / Phase 2 scope, and the permission-mode → DeepAgents-config mapping
  • Narrowed scope of src/quoriv/core/__init__.py (DeepAgents IS the runtime; we just wrap create_deep_agent)
  • Narrowed scope of src/quoriv/tools/__init__.py (only Quoriv-specific tools — AST, git, tests, web — DeepAgents owns files/shell/grep/todo)
  • Narrowed scope of src/quoriv/permissions/__init__.py (mode translation only; FilesystemMiddleware enforces)
  • Narrowed scope of src/quoriv/repo/__init__.py (tree-sitter symbol layer powering AST tools)

Removed

  • src/quoriv/memory/ subpackage — DeepAgents' MemoryMiddleware loads PROJECT.md / ~/.quoriv/memory.md directly via the memory=[...] parameter. No custom loader needed.

Coming next (Phase 4 — kickoff)

  • Phase 3 is feature-complete. Phase 4 (release polish — MkDocs site, PyPI publish, PyInstaller binaries, CI matrix release pipeline, security policy, telemetry opt-in, v1.0.0 tag) starts in the next milestone — see PROJECT_PLAN.md.
  • v1.0.0 tag + announcement — once the release pipeline, docs, and binaries are in place.