Repository navigation
Days 249–264. Several things yoyo used to do quietly it now refuses or asks about, and one key stopped travelling. ANTHROPIC_API_KEY is sent only to --provider anthropic; /save, /export and /spawn -o refuse to replace an existing file without --force; a bash command the new parse tree can't read asks for confirmation even under "always" or an allow rule; a .yoyo.toml that isn't valid TOML stops the run with the file and line; and --provider custom (or an unknown provider with --base-url) with no --model refuses instead of sending claude-opus-5-5 to your own server. Those are why this is a minor bump. The rest of the span moves config reading onto the toml crate (five shapes of valid TOML that yoyo misread, several of them fences that never built), moves bash safety onto a real shell parser (tree-sitter-bash), and stops several credential leaks: to a cross-origin redirect, to an untrusted project's base_url, and onto your screen from config show.
Changed (behaviour you will notice)
/save,/exportand/spawn -orefuse to overwrite an existing file. Each says<path> already exists. Nothing was written.and names the--forceform;/spawn -orefuses before the sub-agent runs, so nothing is spent./savestill replaces a file it wrote itself this session.--save-session <path>and the CLI--output/-oflag still replace silently (#1211; the remainder is #1214).- Bash commands are checked on a parse tree, and what the tree can't read asks. yoyo now parses each command with tree-sitter-bash and runs its safety checks on every simple command in it, including the bodies of a literal
bash -c "…"/sh -c '…'(#1209). A parse error,eval, or a command name chosen by an expansion goes to confirmation, even after "always" or under an exact allow rule. A wildcard allow rule (git *,cargo *,*) auto-approves only a single plain command, sogit log; rm -rf xorcargo $(…)asks. Brace lists, backslash escapes, globs and$'…'words are never treated as literal text for auto-approval (#1205). The release binary grew by about 1.5 MB for the parser. Work on this continues in #1202. - A
.yoyo.tomlthat isn't valid TOML stops the run. Every loaded config file is parsed with thetomlcrate at startup; a syntax error prints the file, line and column, and nothing is sent (#1184, still open for the remaining line readers). --provider custom(or an unknown OpenAI-compatible provider) with no model named refuses instead of sendingclaude-opus-5-5to your own server; the REPL's/providerdoes the same (#1175, #1183).--provider githubis refused: GitHub Models was retired on 2026-07-30, and its endpoint answers every path with a plain-textOK. The setup wizard and README no longer offer it (#1166).-pexits non-zero when the reply was cut by a limit you didn't set, or refused, and--output-format jsongains anincompletekey (null,{reason: output_limit, limit_set_by}, or{reason: refusal}) withis_erroragreeing with the exit code; stream-json prints the same note on stderr (#1026).yoyo -p hi extrarefuses the stray word (exit 2, nothing sent) instead of silently dropping it, and the refusal names a flag that swallowed another flag as its value (#1150, #1151).- The temperature range follows the provider: up to 2.0 where the vendor documents 0–2, 1.0 otherwise, read by the run,
/config setandconfig getfrom one rule, so/config set temperatureaccepts only what the run would send (#1204, #1207). (For Claude 4.7 and later, see Fixed.) - GIF is no longer accepted by
/add,@pathor--image(yoagent 0.27 refuses it), and every "supported formats" message comes from one list (#1143).
Security
-
Provider requests no longer follow a cross-origin redirect with your API key (
x-api-key,api-key,Authorization, or aheaderscredential), on anthropic, azure, google and OpenAI-compatible providers (yoagent 0.27.8, pinned by a recording-stub test; #1181). -
An untrusted project's
.yoyo.tomlbase_urlno longer receives your API key.base_urlis now a trust grant like[permissions], hooks and MCP servers (#1179). -
config show/config getno longer print credentials raw. A key yoyo never reads has its value shown as***(a pastedopenai = "sk-…"or a DSN with a password), and keys whose names contain a credential word stay masked;max_tokensis no longer masked (#1194, #1197, #1199). -
Bash safety checks see more spellings:
$(…)and backticks inside double quotes (#1189),GIT_DIR=/GIT_WORK_TREE=glued to$(or a backtick (#1191), checks glued into$(,(or a backtick (#1193), and user deny rules against quoted or escaped command words./readand/planrefuserm,rmdir,unlinkandshred(#1192).echo "kill -9 1 at $(date)"is judged the same asecho "kill -9 1"(#1196); a quoted mention inside a non-print command can still ask (#1210). -
ANTHROPIC_API_KEYis no longer sent to non-Anthropic providers. It used to be the fallback key for every provider whose own variable was unset, so with an Anthropic key in your shell,--provider ollama,--provider custom --base-url ...,openai,bedrock,xaiand the rest sent it to that host asAuthorization: Bearer sk-ant-...(in-p, the REPL andyoyo review). Now only--provider anthropicreads it. The genericAPI_KEYand a config-fileapi_keystill apply to every provider. If you relied on it: a provider that needs a key and has none now stops withNo API key foundnaming its own variable (e.g.OPENAI_API_KEY) instead of sending the wrong key and getting a 401;ollamaandcustomrun with no key,yoyo reviewincluded (#1170).
Fixed
- A comment after a section header no longer drops the section.
[directories] # keep outfollowed bydeny = ["/tmp"]is legal TOML, but yoyo only recognised a header line that ended in], so the whole section was skipped: the deny list was never applied, and the new unknown-key warning blamed yourdenyline. Every section in.yoyo.toml([permissions],[directories],[mcp_servers.*],[model_pricing.*]) now reads a header with a trailing comment (#1174). - A config key yoyo never reads now gets a warning instead of silence.
modle = "gpt-4o"in.yoyo.tomlused to be ignored without a word, so the run went to the default model while you believed you had picked gpt-4o. Every run now prints one stderr line per unread key or section, with the file, the line and a did-you-mean (for example "did you meanmodel?" or "did you mean[permissions]?"). It is a warning, not an error: a key from a newer yoyo or for your own tooling does not stop the run (#1171). --temperature/temperature = ...no longer breaks Claude 4.7 and later. Anthropic removedtemperaturefor those models (including the defaultclaude-opus-5-5, on Anthropic and on Bedrock), and every request that set it failed with HTTP 400 and no hint. yoyo now doesn't send it to them and prints one stderr note saying so; older models (e.g.claude-haiku-4-5) still get it. The install guide's example config no longer sets it (#1145).- The banner and
/git stageno longer misread the first changed file. When the first line ofgit statuswas an unstaged edit (M Cargo.toml), yoyo dropped its leading space: the banner said1 staged, 1 modifiedinstead of2 modified,/git stageofferedargo.toml(a path that doesn't exist), and a one-character filename vanished from the list./git status,/git branchand/diff's--statsummary also had their first line shifted left. yoyo now keeps git's output exactly as git prints it, except for the trailing newline (#1169). - An unparseable
.yoyo/memory.jsonis no longer erased. One trailing comma used to make the file read as empty, and the next/remembersaved over every earlier memory. Now/remember,/forgetand auto-remember refuse with the file, line and column (not saved, nothing was written: … is not valid JSON: trailing comma at line 1 column 67), and startup warns that project memories were not loaded. An absent file is still an empty store (#1216, still open). - More valid TOML is read the way TOML says. A multi-line
deny = [ … ]list in[permissions]or[directories]builds its fence; an inline table (directories = { deny = [...] }, orfs = { command = "npx" }under[mcp_servers]) is read;[ directories ]and dotted keys are headers and keys to every reader; a#inside a quoted value is not a comment; and a key under some[section]no longer overrides the top-levelmodel/provider/base_url(#1177, #1178, #1180, #1182, #1184). config get/config showreport the value the run actually uses (model = ""is unset,thinking = "max"runs as high, temperature after clamping), name a second home config file that is being ignored, and no longer list~/.yoyo.tomlas its own shadow when you run from$HOME.config getwith no key exits 2 (#1172, #1185, #1195, #1154).- After a model refusal, the next request no longer fails with a 400 about
cache_controlon a thinking block (#1206, still open for the remaining cases). - A checkpoint stop in the REPL (or piped with
--continue-on-silence) is not undone by auto-continue (#1161). - A peak-load rate limit is no longer mistaken for context overflow (yoyo now uses yoagent's classifier) (#1167). Google's and xAI's bad-key replies get the auth hint (#1163, #1165). The DNS hint matches reqwest's real error text (#1126).
-p/ piped API-error hints name--provider/--modelinstead of REPL slash commands, and the REPL billing hint names a command that really switches to a local model (#1128, #1139).- A CA environment that loads no certificates (
SSL_CERT_FILEmissing, empty or not PEM) is refused with a message instead of a provider panic, and a panicked worker is no longer reported as "reply cut off" (#1138, #1142). - Bedrock: runs with thinking on send an explicit
maxTokensabove the thinking budget,/sideand/architectuse the samemax_tokensrule as the main agent, the picker stops offeringanthropic.claude-opus-4-7(which ran on guessed limits), and the guessed-limits note stops claiming a cap that isn't sent (#1121, #1160, #1173, #1198, #1203). - Model lists: dead or nonexistent Claude ids (
claude-sonnet-4-7,claude-sonnet-4-20250514, the retiringclaude-sonnet-4-5) are no longer offered, Claude context and output limits come from yoagent's table, the Gemini list matches Google's models page with limits from Google's own pages, OpenRouter Anthropic ids yoagent knows run on real limits in both spellings, other OpenRouter models get yoagent's vendor limits instead of a 128K / 4096 guess, and every provider default resolves through a preset or states why not (#1117, #1118, #1119, #1144, #1157; non-Claude ids are still being checked in #1147). - Shell verbs:
/grepreads every word after the pattern as a path, never a git revision (#1201);yoyo mapnames the backend that actually produced the symbols and no longer says "no supported source files" when every symbol is private (#1135); shell verbs printyoyo <verb>usage instead of/verb(#1127, #1155);skill show|install|initwith no argument exit 2 (#1153);goal set,init,skill init|install|searchandrisk snapshotexit 1 when their write or fetch fails (#1137); an unknown dash-led word is refused for every verb and subcommand (goal set --bogusno longer saves the typo) (#1132, #1133, #1136); and a flag's value slot is no longer read as a flag (yoyo --system --model x -p hi) (#1114). - The "treating unknown provider X as OpenAI-compatible" warning prints once per run, not five times (#1176).
--helpstops claiming--max-tokensdefaults to 8192 (#1190).
Added
- Subdirectory instruction files. The first time the model reads a file under a directory that has its own
YOYO.md/CLAUDE.md/AGENTS.md, that file is loaded too (deny-listed directories excepted). - Long bash output keeps its end. The bash tool keeps the head and the tail of long output, with the number of omitted bytes in the marker, so a build's final error is no longer cut away (#1159).
- The partial-read guard reaches sub-agents. yoagent's per-agent ReadLog is wired into
read_file/write_filefor the main agent and for every sub-agent and explore-agent (#1094, open for the never-read overwrite question). /costprices through yoagent's price table first, and the yoyo rows it shadowed are gone (#1012).- The docs say plainly that
-pand piped yoyo approve every tool call (#1149).
Dependencies
- yoagent 0.26.8 → 0.27.9. New:
tree-sitter0.25 andtree-sitter-bash0.25 (bash parse tree),toml1 (config), andrustls-native-certs0.8 on Linux/BSD (the CA pre-flight, #1138).
Price table audit
price_drift_audit(DeepSeek rows): passed, comparing 0 rows (deepseek-flash,deepseek-v4-flash,deepseek-v4-flash-vision-expanddeepseek-v4-proare registered divergences;deepseek-r1/deepseek-v3are not listed by models.dev, so nothing audits them).- General sweep (run 2026-11-19):
price audit: SUMMARY compared 115, matched 54, drifted 48, cache_read_only 13, unpriced 69 of 184 catalogue rows (rel_tol 1%). The 48 drifted rows are UNRECONCILED and owned by #1012 (23 at v0.4.0); every row is listed there. By provider: OpenAI 17, Google 14, Z.ai 8, DeepSeek 4, Mistral 3, Anthropic 1, xAI 1. 12 of the 48 differ only in the cache-write rate. Vendor pricing pages were not read for this release, and models.dev is community-maintained, so/costfigures for those models may be wrong in either direction. The 13cache_read_onlyrows are the known unmodelled-cache gap.