v0.3.0

• Added a doctor feature (cli-router doctor, --repair) to catch agent-CLI drift over time. doctor reports each provider's discovery health (CLI present? live model list parses?). With --repair, when a provider's CLI is installed and responding but its model list no longer parses, Doctor uses a working agent as an LLM parser — it runs the sick provider's own list command itself, hands only the raw text to the agent, and expects back a JSON array of model ids (LLM-as-parser, never LLM-as-operator). Recovered lists are written to ~/.cli-router/model-cache.yaml (cli_router.modelcache.ModelCache), which layers between live discovery and the static DEFAULT_MODELS fallback, so the fix survives across runs without editing source. Model discovery and the TUI model pickers now consult that cache.
• Made Doctor resilient to broken agents via a modular backend-failover chain. Instead of trusting one "healthy" provider, Doctor builds an ordered chain of (provider, model) backends — installed providers alphabetically, each provider's models cache-first then DEFAULT_MODELS — and walks it until one actually answers, pinning the winner and reusing it for the remaining repairs. As long as one provider+model works, Doctor can heal the rest. A stale model (e.g. a retired gpt-5) simply fails and falls through to the next candidate. The LLM call (run_agent) is an isolated seam so future Doctor tasks can reuse the same selection logic. Repairs are cancellable (a cancelled hook, or Ctrl-C), stopping gracefully with partial results preserved.
• Lengthened Doctor's discovery timeout to 10s (vs. the interactive picker's 1.5s, now a probe_models(..., timeout=) parameter) so slow-cold-starting CLIs are diagnosed instead of timing out.
• Stopped probing providers that have no safe, machine-readable model-list command. claude models hangs and claude model list would start a billable agent turn; hermes model is an interactive login/selector. These are removed from discovery, so Claude and Hermes now resolve straight to their static/cached model lists without ever shelling out, and Doctor reports them honestly as static rather than crying drift. (Codex and Grok keep live discovery.)
• Hardened _parse_model_catalog to decode the JSON object at each { (via raw_decode), tolerating a banner before the catalog and a log line printed after it — the exact format-drift Doctor exists to weather.
• Hardened Grok/text model-list parsing against banner drift: _parse_model_output is now section-aware (only trusts a Default model: line and entries under Available models:, falling back to scanning all lines only when there is no header) and requires each token to look like a model id (no spaces, carries a digit or -). This replaces the old prefix-denylist noise filter, which leaked banner words like You when the login/status wording drifted — a latent glitch that became load-bearing once the model is passed as -m.
• Routed the selected model and reasoning effort to the underlying CLI: provider_tool_config now bakes both into the command (codex exec -c model_reasoning_effort=<effort> -m <model>, claude -p --model <model> --effort <effort>, grok --single -m <model> --reasoning-effort <effort>), and editing a provider model config in the TUI regenerates the command. Previously the model and effort fields were stored but never passed, so every model config ran the tool's default model at its default effort. Custom/generic tools keep their hand-written commands; Hermes (single auto-routing model) takes neither flag.
• Refreshed the Codex static fallback model list to the current catalog (gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna, gpt-5.5), replacing the retired gpt-5.1/gpt-5 slugs. Live codex debug models discovery already surfaces the GPT-5.6 Sol/Terra/Luna models; the Claude list already includes claude-fable-5.
• Added per-stage output between stages instead of a bare exit 0: the plain CLI and the TUI end-of-run summary now print each stage's extracted answer condensed to a half-page preview, and the TUI live view shows a one-line result teaser per finished stage.
• Streamed provider progress live: stream_tool now delivers stderr lines through a new on_stderr_line callback (both streams funnelled through one dispatch loop), and the TUI renders that progress as dimmed, ANSI-stripped secondary output — so long stages (e.g. Codex, which streams to stderr) are no longer a silent wait.
• Added condense_extracted, first_meaningful_line, and strip_ansi helpers in streamfmt.
• Added grok as a built-in provider: default model list, grok models discovery, a grok --single command template, a packaged grok.yaml preset, and ANSI/noise stripping in model-list parsing so discovery output stays clean.
• Added stage_library config validation: it must be a list of mappings, each with id, tool, and input_template, referencing a known tool.
• Made [previous stage output] a real variable: it now injects the immediately preceding stage's final extracted text ({previous_output}), not the plan file path. Previously it was aliased to {plan_path}, so downstream stages (QA, Summary) never saw prior stage output and could report stale or contradictory results.
• Added [all stage outputs] ({all_stage_outputs}): every completed stage's final output so far, labeled by stage id — useful for QA and Summary stages.
• Added [plan file] ({plan_path}) as a distinct variable for referencing the plan file path.
• Slimmed the run manifest (run.yaml): per-stage records no longer embed full stdout/stderr (already saved as sidecar .stdout/.stderr files). Each stage now records byte counts plus the artifact filenames, keeping the rendered command for diagnostics. A run that dumped ~800 KB of duplicated streams now writes a ~4 KB manifest. runs show reads the flattened exit_code and still renders older manifests.
• Updated packaged presets and first-run TUI seed prompts: planner and QA stages now instruct "do not edit any files; respond only in Markdown", coder references the plan file, QA reviews the previous stage's output, and Summary consumes all stage outputs.
• Fixed TUI stage reordering (U/D) so the new order is written to the workflow config instead of only the on-screen list; reordered stages now run in the chosen order and survive add/remove and plain cli-router run.
• Persisted TUI edits back to whichever config is in use — a project-local cli-router.yaml or an explicit --config path, not just the home config — so reorders, stage add/remove, and Model Config changes survive across runs.