v0.20.1

A patch release from a whole-codebase bug hunt. Candidate issues were
confirmed by reading or running the code, independently challenged, and fixed
with a regression test each.

Fixed

• Tool-call and tool-result lines no longer parse tool arguments or output as
  Rich markup. Text such as ls [/tmp] used to crash the turn and leave an
  unanswered tool call in history, and code previews silently dropped
  bracketed text (d[key] showed as d). Recalled-context and info lines
  get the same treatment.
• Tool failures reported as Error reading/writing/listing ... are now
  classified as failures instead of "worked".
• search_files keeps ripgrep matches when ripgrep also reports errors (for
  example one unreadable folder) and marks the result as partial. Long
  stderr no longer stops the search early. The Python fallback now honors
  path globs such as **/*.py, dir/*.py and !*.md, case-insensitively on
  Windows.
• run_shell keeps leading whitespace on the first output line, so formats
  such as git status --short stay intact.
• A schemeless OLLAMA_HOST (for example 127.0.0.1:11434) no longer fails
  the readiness check and blocks every prompt, and the supplemental gateway
  receives a normalized host.
• Agent-block models are routed to Ollama Cloud or local Ollama by the block's
  own model instead of the active model.
• --oneshot --json always ends with a done event, reporting partial when
  saving the configuration fails.
• /agent fixes: a model error marks the block as failed; tasks that begin
  with show, switch, resume or fork are only treated as thread
  commands when followed by a thread reference; apostrophes and quotes in the
  task no longer drop --pipeline or --roles; Ctrl+C on a team run stops
  waiting for running specialists.
• Saving a memory never overwrites a memory.json that exists but cannot be
  read safely; an empty file still counts as no memories. One invalid legacy
  fact no longer breaks /remember or turns off recall, and the load error
  clears after a successful repair.
• Lessons written as paragraphs (the template format) are indexed, not only
  ## sections.
• Automatic memory capture rejects complaints and reported speech such as
  "You always forget to run the tests".
• When the context summarizer fails, the lossy fallback summary is reported
  instead of being used silently.
• Code-RAG re-embeds chunks after an embedding-width or embedder change
  instead of failing on every turn, keeps older indexes usable after upgrade,
  and drops deleted files from the repository map.
• A source file with a future modification time no longer forces a full
  harness re-index on every load.
• Ctrl+C during a streamed answer keeps the partial answer in the
  conversation and says so.
• /agent resume and /agent fork no longer treat English words made of hex
  letters (add, beef, decade) as thread references; a reference followed
  by a task must contain a digit or match a full thread id.
• A schemeless OLLAMA_HOST now also works for embedding identity and the
  gateway upstream, so code, lesson and memory retrieval stay semantic.
• File contents that happen to begin like a tool error are no longer recorded
  as failed tool calls, in the chat loop, /agent, one-shot and program runs.
• Partial search results are recorded as not fully successful.
• Malformed search globs such as *[^] no longer crash the Python fallback.
• /agent team cancellation stops write_file before it writes and
  terminates a running run_shell command.
• Ctrl+C during a parallel tool batch returns within a bounded wait and keeps
  every tool call paired with a result.
• /diff, /changes, agent-block panels, /help, the memory table and the
  runtime overview show bracketed text literally instead of treating it as
  Rich markup.
• Existing lesson indexes are rebuilt with the new chunker, and lesson text
  containing comment markers is kept intact.
• Memory capture: needed/embedded standing rules are accepted, and
  first-person complaints (I always forget ...) are rejected.
• Files created after the last code-RAG scan are found on the next scan.
• The settled thinking panel shows both the opening and the ending of long
  reasoning, and the live panel keeps indentation.

Changed

• Ctrl+C at the prompt clears a half-typed line; on an empty line it asks for
  a second Ctrl+C (or Ctrl+D) before exiting.
• The footer reflects /safe, /auto, /theme and /cd changes
  immediately.
• The model-wait spinner shows the elapsed time, and switches to "loading" for
  slow local model loads.
• The live thinking panel follows the newest reasoning instead of freezing
  after the first 1,200 characters.
• Errors show their class and, for connection, timeout and missing-model
  failures, a short hint about what to check.

Security

• The public-release scan also rejects local keychain item names used for
  companion credentials, in source and in built artifacts.

Release Limits

• Native signing, M8 external-browser metrics and other blocked hardening
  work remain blocked and are not claimed.