[AlleyBo55]

…rong ones

No new parity rows. The provider layer, the permission boundary and the agent loop get their first tests; what the tests found gets fixed; and the things that decide whether "works with any model" is true get built.

Fixed

• Streams that broke mid-reply were recorded as complete (both providers returned silently on a bad frame or dropped connection). Now surfaced as the error event the REPL renders. Anthropic mid-stream error frames too.
• Model fallback never fired for rate limits: the chain judged the RetriesExhausted wrapper (status 0), not the 429 inside it. OpenAI's context_length_exceeded lives in error.code and was never read.
• Streamed token counts were zero: OpenAI-compatible streams never asked for usage; Anthropic input tokens live in message_start and only message_delta was read.
• Trusted-tool prefixes matched substrings of the raw JSON: "BashTool:ls *" approved rm -rf ./tools, "FileWriteTool:src *" approved src/../../etc/passwd. Now word-boundary on the command, directory boundary on the cleaned path, and every command in a compound line must be covered; substitution and file-writing redirects are never auto-trusted.
• Malformed tool arguments executed with empty input; the model now gets the bad JSON quoted back and a request to resend.
• Codex provider path was unreachable and read the wrong keys from ~/.codex/auth.json.
• Context estimate ignored tool results, so auto-compaction fired far too late. A file ending in a newline no longer gets a phantom empty last line.

Added

• Zero-config local models: an Ollama-style id routes to a running Ollama or LM Studio with no key. With no --model and no keys, gocode picks a local coding model and says so. With nothing set up, it prints three ways to start instead of a credentials error. Bare gocode in a terminal starts chat; --help is grouped and the porting-harness commands are hidden.
• Loop guard: identical tool calls with identical results three times warn the model, five times stop the loop (stop_reason "stuck").
• --max-cost on chat and prompt: a dollar ceiling for unattended runs, enforced before tools run, transcript kept valid.
• Per-model cost ledger behind /cost with a real price table; cache reads priced at Anthropic's rates.
• Prompt caching on Anthropic: cache_control on the system block (covers tools) and on the newest message, set on request copies so breakpoints never accumulate. GOCODE_DISABLE_PROMPT_CACHE opts out.
• Tool output capped at 40 KB per call (head and tail kept, marker explains how to ask for more); FileReadTool pages at 2000 lines.
• gocode doctor --model X probes streaming, system prompt, tool calls and parallel tool calls; cached in .gocode/capabilities.json; chat warns at startup from the cache.
• Swarm made real: every spawned sub-agent joins under its own name, is offered send_agent_message and list_agents, reads its inbox each turn; background agents report to main. Previously the tool was never offered and no mailbox was ever read.
• evals/: 15 tasks with hidden tests and reference solutions, a runner that scores gocode per model, and a self-test proving each task is solvable and discriminating without spending tokens. Opt-in, not in CI.
• Tests for apiclient (recorded SSE fixtures for both wire formats), agent, permissions, doctor, swarm, orchestrator, toolimpl.

Changed

• /cost is multi-line and priced per model; --output-format json total_cost uses the same pricing.
• Bare gocode on a TTY starts chat instead of printing help.
• --model codex with no credentials errors at startup.
• BashTool:git * no longer covers git log | head; trust head * too.
• Anthropic requests send system as a block list.

Known gaps: proxy providers prefer the env key over --api-key; the price table is hand-maintained; the deepseek alias points at a retired model; nothing here has yet been run against a live model.