v0.6.3

23 PRs on top of v0.6.2 — the harness release: Fusion and local models that finish the task, and tools that stop damaging the user's work.

Fusion that delivers

• Workers see the request. Every brief quotes the operator's original request; briefs up to 32K chars; a worker's reads are confined to its working directory and the fan-out's write scope (#426).
• Task status is ground truth, not the worker's reply. A task that declared files and wrote nothing is no_changes; a reply on the forced final step is max_steps; the delegate result opens with a status table for every task and the call reports all_ok / partial / all_failed (#426, #428, #431).
• A contract between the parts of a fan-out. fusion.delegate takes owners, provides, requires, inputs and checks; parts run in waves ordered by requires → provides; one refusal names every problem at once (#431, #439, #459).
• Check before you accept. New read-only verify.syntax and verify.run (command, service or headless page) run against a throwaway copy of the working directory; contract checks run through them (#432).
• No more wasted workers. Worker time limits come from measured throughput; a worker is handed back only after it stalls, never mid-generation; reasoning effort and output caps reach the provider; llm.runMode.fusion.cloudWorkers caps a cloud fan-out; /runmode status says what will actually run (#434, #439).
• A stalled review must act. A Fusion orchestrator that keeps reading without delegating gets one notice, then can only fusion.delegate or reply (llm.runMode.fusion.reviewStallSteps, #459).

Local models

• The KV cache survives long turns. Prompts only append between history cuts, sessions keep their llama-server slot across steps and resumes, context and fan-out width are sized from the GGUF header and a throughput probe, sliding-window models launch with --swa-full (#429, #436).
• Bounded thinking. localModels.reasoningBudgetTokens (default 1,500) caps the reasoning prelude before each tool call; localModels.thinking: off now applies to every prompt (#453).
• Per-request grammar and tool roles. A local orchestrator can no longer generate calls the gate would refuse; workers and orchestrators get their own tool sets (#430).
• Families without a hand-built profile render through llama-server's own chat template (localModels.useServerTemplate, #435).

Cloud providers

• Native multi-turn messages for OpenRouter, OpenAI-compatible, AIMLAPI and Gemini, with Anthropic cache_control breakpoints, cache-capable routes for Google models and cached tokens on the usage record (#433).
• Reasoning fields and tagged tool calls parsed for every OpenAI-compatible kind; o* / gpt-5 models no longer get temperature / max_tokens (#433).
• Provider errors told apart: exhausted credit stops the turn resumably, retry hints are honoured up to 180 s, a context-length rejection learns the window, repacks and retries once (#435).

Tools that protect your work

• Your files are safe. Replacing a pre-existing file saves the previous bytes and says so; os.fs.restore brings them back; a file the request names as an input is not replaced without overwrite: true (#438, #459).
• Long shell commands detach instead of hanging. After tools.shell.defaultTimeoutMs (10 min) the agent gets the output so far and a job id to wait, kill or keep (#452).
• Reads outside the working directory ask first through the approval ladder (fs_read_outside; agent.readScope: unrestricted to opt out) (#452).
• Calls carrying model control markers or unknown argument keys are refused with the expected keys; batched writes run whole under auto-approval; .js/.json writes are parse-checked; os.fs.grep { literal: true } (#427, #435, #438).
• A reply batched with work tools is a progress note, not the end of the turn (#453).
• Bigger defaults for today's contexts: 200 conversation pairs, 32K batch results, 8K render cap, 60 worker steps (#437).

Also

• Atomic Chat (local) provider preset; an External llama.cpp URL on :1337 opens the wizard on it (#424).
• Re-run first-time setup any time: /onboarding (alias /setup), the Setup menu row "Run first-time setup again…", or atomic-agent tui --onboarding. Providers, keys, models, sessions and memory are untouched, and backing out writes nothing (#425).
• Downvoted profile facts, pinned ones included, now leave ### profile as memory.voting.profileFilterThreshold promised (#423).
• Model downloads no longer hang at 9x%. A connection the CDN throttles is swapped for a fresh one, and connections that finish their slice take over the rest of a slow one (#462).
• A message typed while a turn is running is labelled YOU · steered into the running turn, so the turn's answer is no longer read as its reply; [try again] on a stopped turn re-sends the turn's own request, not the steer (#461).
• The forced reply after a wandering loop stop says what actually happened instead of "a no-progress loop after 12 blocked attempts" (#460, refs #458).

Config migrates automatically to v69; values you set yourself are kept.