Repository navigation
Releases
Releases · Rixmerz/locks
v0.111.2
Fixed
locks statsno longer counts a worker's tool call timed longer than the whole attempt it ran in (a
machine that slept read as an 8-hoursed); it says how many it left out (Store.tool_time_impossible).locks footprintcompares releases only on briefs of tasks outside a workflow, whose briefs carry what
earlier steps said, and names a release's growth only when both sides holdMIN_TASKS(10) briefs.
v0.111.1
Changed
- The orchestrating session no longer asks the person to type a
lockscommand. A resume that grants
(--allow,--allow-approval,--keep) waits for the person's yes in the conversation, then the session
runs it and Claude Code's permission prompt is their approval; where Claude Code refuses it outright, the
session points to the stop's button on the board instead of a line to type (commands/orchestration.md,
locks guide endings).
v0.111.0
Changed
- The compaction advice is judged against the window the session really runs in and kept ahead of the
harness's own compaction: the model a session's SessionStart names ([1m], which the transcript's model id
never says) and any compaction the transcript keeps (preTokens) set the window,
CLAUDE_CODE_AUTO_COMPACT_WINDOWnarrows it, andCLAUDE_AUTOCOMPACT_PCT_OVERRIDEor the session's own
automatic compactions say where the harness compacts by itself; the worth, pressing and urgent shares stay
below that (Context.shares), and with the harness's compaction off nothing is pressing. While work the
session sent still runs, fromRUNNING_SHAREof the window, the session is told it is the moment to decide
whether to compact before that work's ending or after it, before the harness decides for it.
v0.110.0
Added
- Each tool call keeps how long it took, as Claude Code's
PostToolUsereports it (duration_ms), andlocks statsshows where the workers' tool calls spent their time: by signature, the calls, their total and the longest. The store's schema moves to version 2 (one added column). /locks:insightruns in a fork of the session (context: fork, in the foreground): the traces, the analysis and doctor it reads stay out of the context every later turn re-reads, and only its report comes back.
v0.109.0
Added
- TRACER as a way of working, not only documents:
locks method tracer onputs a project on the method, andlocks methodsays the phase it is in (read from whatdocs/tracer/holds), its slices and the bundled workflow of that phase. Five new workflows do the phases before the slices:tracer-contract,tracer-spec,tracer-empathy,tracer-designandtracer-slices; their writing steps are held todocs/tracer/andopenspec/by their surfaces, and the person's decisions (the contract's sign-off, the expectations that enter, the design base) are question steps. - A build request in a project on the method hears the phase it would skip, or the next slice: advice, never a block.
- A slice's Definition of Ready is checked in code before any worker (
ready: tracerontracer-slice's build step): one whose requirements, empathy round, capability, critical flows or floors are missing stops asunready, naming what is missing. locks methodcounts the expectations the slices' verifiers found that no CORRECT line, floor or catalogue entry covered.
v0.108.1
Fixed
- A worker that locks' own guard refused to start locks work (a run, a workflow, the queue) stops the task as a
decisionat once, its work kept, instead of climbing: more effort met the same refusal twice before a third cell said so.
v0.108.0
Added
scripts/demo.py DIR [--shots]: demo data for the README's screenshots, written by locks' own
commands with the scenarios' fakeclaudein two generic repositories (no money spent), times and
costs spread to a real worker's size;--shotsretakesdocs/images/with Playwright. The
screenshots are retaken with it.
Fixed
- The board's Spend screen said
$-0.00for sessions when every dollar was a worker's (a float's
rounding); it says$0.00.
v0.107.0
Added
- What each workflow cost here, on the board. The Workflows screen gives every definition its record
in the project (api.workflow_record): runs started, runs done and the median cost of a done run, read
from the steps of up toRECORD_RUNSruns. Observed, never a budget.
Fixed
- A continued session's attempt no longer counts the session it continued twice. The CLI's
total_cost_usdfor a--resumeor a--fork-sessionis the whole session's (measured: a fork and a
resume of a $0.0048 session each said $0.0059), and locks kept it as the attempt's own: every attempt
given more turns or more effort, a trunk's fork and a resume after a death counted what came before
again. In a realfeature-parallelrun a part that forked the plan was kept at $1.00, $0.64 of it the
plan's. Now each worker session's whole cost is kept (orchestrate.session_cost, the store's meta),
and an attempt that continues one keeps what it added (launch.own_cost,LaunchSpec.resume_cost_usd);
the CLI's figure stays asWorkerResult.session_cost_usd. A session whose cost was never seen (one
from before this release) keeps the CLI's figure, as before.
v0.106.1
Fixed
- The bundled
feature-parallel's plan is read blind. Its plan step namesstart: opus/high, and
the critique counted that as a person's pin: it never drew an arm, so the one bundled workflow with
parts never ran it (found by a real run ofbench/speed/matrix.py). A start a workflow step's
definition names is no pin. - A tracked tool artifact is never committed as the work. A kept branch dropped its tools'
artifacts withgit rm --cached, which staged the deletion of the ones a repository tracks: afeature
run committed eight.pycits check removed, its review read them as the change's, and two fix rounds
could not undo them (stopped asreview). They are now reset to HEAD.
v0.106.0
Added
- A review's fix goes on in the session of the worker that built the change. In
feature,bugfix
andtracer-slicethe fix step forks its builder's session (trunk), told it is its own change and
given the rows the review confirmed, instead of a fresh worker orienting again in code it never saw.
Only on the builder's model; whether it pays is the trunk's existing fork/fresh trial. scripts/mutate.py: mutation testing of one module against the tests named for it, stdlib, the
module always put back. Its first run:findings.pyagainsttests/test_checking.py, 21 of 26
mutants caught, the five left all text-clipping lengths, no rule.docs/workflows.md: which bundled workflow fits which work, with one measured run of each.
Fixed
- The PreCompact hook has 30 seconds, not 5. It writes the compaction's digest from the whole
transcript; on a long session it was cancelled ("Hook cancelled") and the digest was not written.