Repository navigation
Releases
v0.88.10-alpha
What changed
Added
- A host without a command-prefix key can still win the bookkeeping. opencode's
"shell"only names a binary, so the prefix is a two-line wrapper that exports
BASH_ENV, plus abashenvthat setspipefailand puts the tool directory on
PATH. The same three checks apply, and they were run through the host
(opencode runreturned exit 1 for a failing pipeline) rather than read off the
config schema. The "what a host must implement" table gains the row.
Changed
- The harness reference now states the part prompt text cannot buy: a session that
carried a "one call per question" rule made 85 consecutive direct calls and zero
scripted ones. Bookkeeping belongs to the harness.
v0.88.9-alpha
Changed
- The shell rule is an outcome, not a tool. The previous entry told the worker to
put "two or more commands" into acodemodescript. Measured over a 2 000-call
session: 99% of calls already carried more than one command (4.2 on average), and
codemodewas written 0 times in 85 calls after the rule was added. A rule that
prescribes an instrument is inert when the behaviour it wants already happens — so
the entry now states the invariant ("one call per question, filter inside the call,
one screen out") and namescodemodeonly for what shell does badly: tool
orchestration and JSON the shell would mangle. harness.mdputs the bookkeeping on the harness. ~30% of 8 265 shell segments
werecd(2 005),export PATH(487) andtimeout(810) — a stateless shell making
every call restate where it is and what it can run. The fix is a shell prefix and a
persistent cwd, plusset -o pipefail: a pipeline ending in| tail -2reports the
exit status oftail, which is how this project once tagged a release from a red
suite.
v0.88.8-alpha
Changed
harness.mdsays whatcodemode.mode: "only"actually bought, measured.
Four runs on pi 1.5.2, one host, one model: on a task that wanted enumeration,
"only"produced the right answer at 44.8k tokens against 35.9k and the same
four calls; on a task that wanted one edit, three calls against five. It does
not reduce round trips — the claim it is sold on — so the entry keeps"on"
as the default and reserves"only"for volume jobs, with a note to measure
before believing it.
v0.88.7-alpha
Fixed
harness.mdno longer recommends-bash, because it amputates instead of
enforcing. Acodemodescript can call only active direct tools plus
codemode/deferredtools, so removingbashfromdefaultToolsremoves it
from the session entirely: measured on pi 1.4.0 and 1.5.2,tools.bash does not exist,ALL_TOOLSlists nobash, and the agent is left with no shell. The
entry now enablescodemodeand removes nothing, namescodemode.mode: "only"
as the one hard enforcement that keeps capability, and says whatexposure: "codemode"would cost for a built-in (re-registering it, truncation and hooks
included).harness.mdsays what PTC buys: roughly 4.3M → 149k tokens on a measured
50-item triage, break-even on a single small task, with the two failure modes
worth naming (a script that silently covers half the items, and a wrong tool made
cheaper to call).
v0.88.6-alpha
Changed
- The human-facing question stops naming a retired ladder. The run-knobs
migration replaced the appetite, butask-patterns.mdstill told the worker to
ask about "Pattern 7 (Appetite)" andplan-critiqueto say the estimate was
"bigger than the appetite" — against a config that no longer exposes one. The
question now names the declaredqualityand exploration knobs and the surface
that would come off, and the legacyappetite:line stays readable for
projects that still carry it. Merged from the branch opened on 2026-10-27, after
rebasing onto main.
v0.88.5-alpha
Changed
harness.mdnames the surgical enforcement, not the blunt one.
{"defaultTools": ["+codemode", "-bash"]}changes the inherited selection, so
read/edit/writestay declared and the shell becomes reachable only from a
script;codemode.mode: "only"hides everything, including the exact-match edit
that a script does worse. Verified on pi 0.88: a fresh session lists
read, edit, write, codemode, nobash.- opencode 2.3.0 needs no flag — Code Mode is built in (
experimental.codemode_vm
is deprecated and ignored, "Code Mode always runs on the VM") and MCP servers are
exposed through it by default. A per-tool deny needs a plugin plus apolicies
entry.
Fixed
AGENTS.mdno longer states the navigation ladder as a reflex. ripwire and
cymbal are conditional, with the situation that pays for each and the measured
reason (up to 28% more cost, arXiv 2609.30725).
v0.88.4-alpha
Added
cli-tools/harness.md: the three harness capabilities a stage cannot ask
for, with the setting that enforces each. A prompt can request economy; only
the harness can make it true. Programmatic tool calling is the big one —
{"defaultTools": ["+codemode"], "codemode": {"mode": "only"}}hides the
built-in tools and makes the model reach them through scripts, where the
default"on"only appends a note to each description. The file also covers
the two-tool MCP surface (search_tools+execute), the workdir-script
fallback when a host has neither, tool-result pruning, why a gate must name a
workspace member rather than a role, and the two measurements to repeat on the
next card.
Changed
code-map.mdsays when a navigation rung costs more than it saves. A map
says where; the edit still needs the code, so a rung pays only when it
replaces a search. Named symbol →ffgrep+read --offset: two calls, the
second returns the code.ripwire --forfor an unfamiliar repo touching three
or more files. Structure-aware retrieval raised agent cost by up to 28% on
SWE-bench Verified (arXiv 2609.30725) — the
extra hop is the mechanism, so the ladder no longer chains ripwire into cymbal
by reflex.
Fixed
npm run buildfailed ontypes/stages.ts—StageSkipCondition's index
signature did not admitundefined, which its optionalany_ofrequires. CI
runstypecheck, notbuild, so the documented command was broken without a
red pipeline.
v0.88.3-alpha
Added
doctorchecks that a recipe declaration has a file.stages.yamlnames
a recipe per stage and a runner reads a 404 as "this stage has no recipe", so
a declaration with nothing behind it was silent.doctornow resolves every
execution.recipe:againstrecipes/<id>.yamland reportsrecipe-missing
(error) with stage, recipe and the path it looked in,recipes-dir-missing
when the directory itself is gone, andstages-missing/
stages-unreadablewhen the declaration file cannot be read. Arecipe:the
reader cannot bind — outside anexecution:mapping, or inside an
execution:written inline on one line — isstages-unparsed(warn) naming
the line, never a finding charged to the stage above it.STELOW_STAGESand
STELOW_RECIPESoverride the two paths (documented instelow doctor
schema output).- A check that could not run doesn't read as a clean bill. Absent stages
file isstages-unchecked(info), an unreadable one is an error, and
doctor: ok (no drift)prints only when there are no findings at all. Two
warn classes were also undocumented:orphan-pidandtransitions-missing.
v0.88.2-alpha
- Number the promises. The spec mints a stable id per promise (
S1,S2), and design, tech, build and audit cite the id instead of restating the promise. On the Multica smoke run the audit had to translate between its own findings and the spec's promises, because both sides described them in prose: a requirement with two names is a requirement nobody can trace.
v0.88.1-alpha
The English pass in 0.88.0 was checked with an accent scan, and an accent scan
cannot see Mesmo ator, Etapa, Nenhum or compras corporativas. Those
survived into the rendered contracts, which is how a host ends up enforcing a
vocabulary the methodology claims it retired.
Fixed
-
The Portuguese that carried no accent.
05-job-actors.mdkeptMesmo ator
in five places (including a criterion another host reads as data),
09-emotional-social-jobs.mdkept a whole table header and theNenhum
status,11-switch-interview.mdkeptNenhuma, and two workflow references
(plan-critique-context.md,step-3-feature.md) kept Portuguese instructions
and step names. All English now. -
docs/study-goal-vs-plannotator.mdwas a Portuguese study note in a
repository that says documents are English.
Added
scripts/check-language.sh, wired intoverify:generated. It looks for
Portuguese words and phrases, not for accents, skips the frozen archive, the
changelog and URLs, and fails on a hit. A future translation pass that misses
an unaccented word now fails in CI instead of shipping in the data.