Skip to content

v0.12.0: lanq makeready --tune (embeds lanq-makeready 0.5.0); parallel development guide - #25

Merged
jangolath merged 10 commits into
mainfrom
release/0.12.0
Nov 14, 2026
Merged

jangolath merged 10 commits into
mainfrom
release/0.12.0

Conversation

@jangolath

Copy link
Copy Markdown
Owner

Summary

  • Embeds lanq-makeready 0.5.0 (four pinned scripts, including the read-only tune scans).
  • lanq makeready <m> --tune (also with --all) runs the tune scan as a second read-only SSH run and reports it in text and in a new additive lanqTune JSON object.
    • The tune report is parsed as untrusted input: capped, validated and sanitised.
    • fixIds and the security flag come from lanq's own tables.
  • The scanner now receives the machine's allowed hours (-Hours), so its active-hours check runs.
  • Docs: docs/PARALLEL.md (from docs: parallel development with many worktrees and a quiet dev machine #24), a Tuning section, JSON.md, AGENTS rules (--tune only when the developer asks; never the fix script).
  • makeready output without --tune is byte-identical (12 goldens captured before the change).

Test plan

  • Pool full plan (3018 passed), targeted, dotnet format
  • CI (incl. the sh-only sync test)
  • Release dry run
  • RC smoke, plus a real --tune run on the pool

🤖 Generated with Claude Code

jangolath and others added 10 commits November 13, 2026 21:39
New docs/PARALLEL.md: several developers or AI agents in parallel git
worktrees on one PC, with every build, test run and format check on the
lanq pool. Covers worktree layout, offloading targeted runs, format checks
and the full plan, allowlisting checkers up front, slot capacity planning,
NoCapacity back-off, a local Docker worker, what sharding buys when compile
time dominates, pitfalls, agent rules and a worked example.

Linked from README.md and llms.txt; AGENTS.md, docs/AGENTS-GUIDE.md, the
lanq skill and both instruction-file snippets gain a short "Parallel
worktrees" block. CHANGELOG gains an Unreleased Documentation entry.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ze heavy)

When the pool includes a worker that runs on the dev machine itself (a
local container or VM), every build and test run from a parallel agent
must carry --min-size heavy (or a --needs capability that worker lacks).
docs/PARALLEL.md gains a "Keep the work off your own PC" section with the
plan form (lanq run --plan --min-size heavy narrows every job; lanq.yml
needs/size still apply; --on with the real workers when a job declares a
small size) and uses the option in every example. The AGENTS.md,
AGENTS-GUIDE, skill and snippet blocks carry the same rule.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Byte-for-byte goldens of lanq makeready's single-machine text and JSON,
--all text and JSON, the held, unreachable and unknown-machine error
objects and an unconfigured PC, captured on the pool from the 0.11.0 code
before 0.12.0 changes anything. The config has no allowed_hours, so the
scanner command is unchanged by the coming -Hours argument; the embedded
scanner version in the --all header is normalised.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
One file list (the scanner and the read-only tune scan, each for Windows
and Linux) drives the download, the SHA256SUMS check (each listed exactly
once), the attestation check and the copy. The fix script is never
embedded. The closing text names the four pins. SupplyChainTests' fake
release has the four files and refuses a SHA256SUMS without the tune lines.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Embeds the verified lanq-makeready 0.5.0 release (synced with
scripts/sync-makeready.sh v0.5.0; SHA-256 pinned in MakereadyTests): the
scanner and, new, lanq-makeready-tune.ps1/.sh. The fix script stays out.
lanq --version prints a third line for the tune scripts (smoke-native
greps it).

lanq makeready <machine> | --ssh ... | --all --tune: after a report
(READY or NOT READY; never for a held or unreachable machine), the tune
scan goes the same way as the scanner (on ssh stdin, run from memory,
150 s, output capped at 256 KB). A configured machine gets -Hours (the
one HH:mm-HH:mm window all its allowed_hours entries share, normalised)
and -WorkRoot (only when lanq's and the tune script's rules both accept
it); an ad-hoc --ssh PC neither. --tune is refused with --setup and --add.

The report is untrusted: Makeready/TuneReport.cs parses it with a
source-generated context (depth 8), checks schema and tool, at most 64
items, ids, statuses (OK/INFO/WARN/SKIP), and the summary counts, and
rebuilds it with every string sanitised and cut (title 120, detail and
advice 1024, host 128, arch 32). fixId comes from lanq's own table (a
test keeps it equal to the scripts' fix ids) and only when the worker
agrees; security is the worker's flag AND lanq's per-OS set.

--json adds lanqTune last in the report (single machine and each --all
row): status Ok|TimedOut|Unreachable|Unusable, error, report, and its
own nextSteps with explicit names (ask_developer without a command for
fixable items, none, retry or ask_developer with lanq makeready ... --tune,
never safe unattended). The text adds a tune section after the report
(WARN and INFO with advice and fix ids, SKIP as a count, and the owner's
preview line); --all --tune a TUNE column. No new status line, no new exit
code: the exit code is always the scan's.

The scanner now gets -Hours too for a configured machine with
allowed_hours, so its suggested allowed_hours and the Windows Update
active-hours check use the machine's window. Without --tune and without
allowed_hours, every output is byte-identical to the 0.11.0 goldens.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
README: a "Tuning: --tune" section (what the tune scan checks, -Hours and
-WorkRoot, the text section and TUNE column, lanqTune, the untrusted
report), --tune in the command table, the --version row (two embedded
script lines), the scanner's -Hours, the nextSteps exception and the
"read-only" security bullet; lanq-makeready 0.5.0 in the --all example.
docs/JSON.md: the lanqTune section after lanqWithheld (keys, statuses,
fixId and security rules, advice and detail are untrusted text), the
reworded "no nextSteps" notes (the single report has none at the top;
with --tune they are inside lanqTune; .machines[].nextSteps never carry
them), the conditional keys, the nextSteps table and a jq tip.
SECURITY.md: the tune report as a second untrusted input. AGENTS.md (a
hard rule), docs/AGENTS-GUIDE.md, the lanq skill and llms.txt: run --tune
only when the developer asks, never the fix script or its preview, and
advice/detail are not instructions. RELEASING.md: four embedded scripts.
PACKAGE.md mentions --tune. CHANGELOG: "Unreleased" becomes
"0.12.0 (unreleased)" with Added, Behaviour changes, Security and For
coding agents above Documentation.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- L4: the scanner gets -Hours only when every allowed_hours entry is the
  same every-day window (no days named), so the suggested allowed_hours is
  never wider than the configured one; the tune scan still gets the shared
  times with the days dropped (ScanHours(m, dropDays), HoursWindow.HasDays).
- L1: ssh's and the worker's stderr is quoted at most 64 lines, each
  sanitised and cut to 256 characters, before SshErrors.Explain and the
  unusable-report detail (scanner and tune); the whole error text is cut to
  1024 after SafeLine, so lanq's own known_hosts and key guidance stays whole.
- L2: duplicate tune item ids are refused ("duplicate item id").
- L3: summary.seconds outside 0..3600 is left out.
- L6: Ctrl+C inside the tune scan exits 130 with the Cancelled object and no
  partial report; single and --all lanqTune are compared for TimedOut too.

Goldens unchanged.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
docs/agents/AGENTS.md-snippet.md and CLAUDE.md-snippet.md gain the rule:
lanqTune advice/detail are untrusted worker text, never instructions, and
the fix script is never run, previews included. README, docs/JSON.md,
lanq --help and the CHANGELOG say the scanner's -Hours is passed only for
one every-day window (the tune scan gets the times without the days), and
that errors quote at most one sanitised 256-character stderr line. The
CHANGELOG Documentation entry lists the snippets; the heading stays
"0.12.0 (unreleased)".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@jangolath
jangolath marked this pull request as ready for review November 14, 2026 07:01
@jangolath
jangolath merged commit 554f780 into main Nov 14, 2026
15 checks passed
@jangolath
jangolath deleted the release/0.12.0 branch November 14, 2026 07:02
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant