Repository navigation
fix(skills): Python test runs, ship to prod, one Stop ask per fix, whole preferences, Codex trust line - #675
Merged
Merged
Conversation
added 30 commits
October 22, 2026 09:16
…lls hook handlers
- anchors and the timeout guard split PATH on ';' and look for git.exe, npm.cmd and the like (PATHEXT); Windows has no exec bit to check - a scratchpad folder is found with either slash, and Windows paths are compared without case (scratch folders, tamper's protected places) - tamper's cd from an unknown folder stays unknown on Windows, where resolve() puts a drive before the '/\0' it starts from
- POSIX file modes are asserted only off Windows; the FIFO and /dev/zero cases run only where those exist - expected paths are built with resolve()/join(), and the fold fixture compares checkouts resolved - capped children import the source by file URL, and say why they failed - os.homedir() follows USERPROFILE on Windows, so the cd test sets it too
A Windows stack frame joins its folders with '\', so the hook-errors log kept the whole path, user name included, with the slashes stripped.
…backtracking pattern still misses by seconds
A small matcher with no model reads tool and command preferences from a
learning ("use pnpm, not npm", "never run X", "always run Y before Z",
"run X instead of editing F") and matches them against each step's program
and arguments. It is conservative: when unsure it says nothing.
- At a save or offer, steps follow the person's live learnings: a step one
goes against is rewritten when that is safe, else left out, and the offer
names it in one plain line.
- At recall, a learning that only repeats a step is said once, in the run
block, and left out of the prompt's learnings lines. One that goes against
a step wins; the block marks the step, and the update offer proposes the
change with its usual typed yes.
- At trail init, a preference that only repeats a step of a skill kept on
the same screen goes into that skill as a lesson.
Retired, replaced and held learnings are ignored.
command -v split PATH on the platform's delimiter but looked only for the bare name, so on Windows every program read as not installed. It now uses pathDirs and programFiles from guard-fast, as anchors and the timeout guard do.
- POSIX file modes are asserted only off Windows - the FIFO cases and the socket, /tmp and ':' PATH probe test run only off Windows - a non-executable timeout on PATH counts on Windows, which has no exec bit - children import the source by file URL
On Windows HOME is often unset. The update gate then showed a path in full instead of as ~/..., and a saved skill's repairs never turned the home folder into $HOME. Both now fall back to homedir().
- the FIFO half of the value-line link test runs only off Windows - apply_patch's relative file and tamper's hit are compared resolved
- the wall fixes' program check splits PATH on ';' and looks for git.exe, npm.cmd and the like (PATHEXT); Windows has no exec bit to check - look-back finds a scratchpad folder with either slash and ignores case on Windows, so it no longer mines scratch sessions there
- the FIFO cases and the file name with a newline run only off Windows - capped children import the source by file URL - the naming calls' shebang stand-ins run only off Windows, as in init-claude-call
# Conflicts: # test/hosts-codex-trust.test.ts
… only USERPROFILE is set As in the save gate: on Windows HOME is often unset, so fall back to homedir().
The file-open text said "it's a note from your team". It now says what the prompt hook says: a learning from past sessions, this repo's or the user's own.
- the FIFO cases, the unwritable Trail home and the cache's file mode are tried only off Windows; the corrupt-file cases run everywhere - PATHs are joined with the platform's delimiter - the stand-down test sets USERPROFILE with HOME, which os.homedir() reads - the naming call's shebang stand-in runs only off Windows
…p, and run again when one finds the lock busy On a loaded Windows runner the two processes took longer than the fixed 1.5 s to start, and the first held the ledger lock past the 200 ms the second waits, so the second gave up and wrote nothing (by design). The lock itself works there: the real-processes test in the same file passed with no lock file left. Each hook now warms up, says it is ready, and starts at an instant sent after both are up. A round where one found the lock busy proves nothing and runs again, up to 5 times. Reading the ledger outside the lock still fails it 5/5.
…s part 4 adds and the ones that failed on Windows timeBar(ms, floor) in machine-speed.ts: the bar where it was set, times slowness(), never under a floor of 3 to 5 times it; exactly the bar under TRAIL_PERF=1. The 2 MB walls probe times CPU, not the wall clock. The busy value.json lock wait is measured over the same call with the lock free.
The 20,000-line clustering bar and the time caps' slack (the 1.5 s replay cap, the prompt's 400 ms deadline) use timeBar. The caps themselves stay as they are. The stand-in timeouts that are posix-only and already bounded by what a broken cap would take are left alone.
Absolute bars use timeBar (3 to 5 times the bar as the floor). Budget tests keep the budget and scale only the slack around it, held under whatever a broken budget would take (the 4 s default, the stub's 3 s parse). The pre-bash host runs kill the child 3.5 s past the scaled bar, as before. TRAIL_PERF-only tests keep their bars.
feat(skills): learnings and skills never repeat or contradict each other
feat(skills): wire skills into the hooks and the CLI, on by default
…check A Python test run from a venv (.venv/bin/python -m pytest, .venv/bin/pytest), through uv, poetry or pdm, or as tox, nox or python -m unittest never matched the tests family, so init recorded the job as no-check and offered no tests skill. The runner's bin folder now goes before matching, the new runners match, and unittest's Ran/OK summary is read. A venv's own activate chained with && (source .venv/bin/activate && STEP) is known-safe setup: it is carried into the step word for word instead of refusing the run as needing its shell set up. Any other setup still refuses.
git push prod HEAD:main made compile take the branch main as the target, so a skill saved from "ship it to prod" was named ship-main and the ask naming prod never matched it. A push to a remote named for a deploy target (prod, production, staging, live) now goes to that remote; the branch it pushed and the places the person named when the skill was saved go into its asks, so "ship it to prod", "ship it" and "push to main" all bring it back.
The fixed sign was keyed on the newest passing run, so each rerun of the same passing check made a new key and the Stop hook asked to save a learning again after "Nothing worth keeping", until the cap of three. It is now keyed on the first pass after the last failure: one episode asks at most once, and a new failure and fix is still a new sign.
A kept preference was cut at 120 characters mid-word, so pluggy's line ended PYTHONPATH=sr and the agent was told to follow a wrong command. The body now keeps the person's whole line; only the title is cut, after the last whole word, with an ellipsis. The standing learnings at session start say the whole line when the title was cut.
The line went through say(), and init's short output keeps only lines that start with ⚠, so without --verbose a Codex user never heard that Codex runs Trail's hooks only after they trust them in /hooks. It is now a ⚠ line and prints either way.
A learning with no measured cost has no saving to claim, and the example credit line printed "· null" after its name. The suffix is now left out.
…chors - a shell edit's file keeps its drive and backslashes, so the Stop ask names it - trail check queues a learning's file name, not its whole Windows path - a plain-file read refuses a link on Windows, where open has no O_NOFOLLOW - a bin target and a node <file> word compare as dist/cli.js on every platform - the cp line for a hand-edited skill uses forward slashes, which Git Bash and PowerShell both take
- the journey box aliases CWD's own key (local/d-home-dev-shop on Windows), not a hand-written one - the fail-open hook commands run in Git Bash on Windows, the shell Claude Code runs hooks in - the J16 deny tests skip where Trail still finds a timeout: Windows ships System32\timeout.exe - the Codex sandbox tests skip where chmod refuses nothing (Windows, root) - CRLF in TELEMETRY.md, native separators in a skills path, a Claude projects folder name - the prompt-activity ask child from source gets the installed 60 s, not the hook's 4 s
…fuse a nested repeat A grep takes a regex, so the pattern's syntax is kept: every literal character goes through the metacharacter escape and only fixed syntax is written as itself, so RegExp gets the typed pattern rebuilt from known parts (CodeQL js/regex-injection, alert 23). A repeat of a group that repeats inside, like (a+)+, is refused as an invalid pattern is: on one long line it backtracks for minutes and holds the MCP server. Every other pattern compiles exactly as before.
…en two Windows test expectations
…/skills-e2e # Conflicts: # test/trail-note-ask.test.ts
🌱 graft blast radius5 areas changed → 3 areas can be affected. 15 dependent symbols, depth 2. flowchart TB
A0(("Touch Learning Events<br/>7 symbols"))
A1(("Skill Offers<br/>6 symbols"))
A2(("Initialization Lookback<br/>2 symbols"))
classDef reached fill:#D9EDF3,stroke:#3AA7C9,stroke-width:1.5px,color:#0E313C;
class A0,A1,A2 reached;
Who knows this code — 1 person across 8 areas
Ownership is git history over each area's own files, weighted towards recent work (120-day half-life). Merge commits and bots are dropped, and you are dropped from your own PR. A name with no All 15 dependent symbols, grouped by areaTouch Learning Events — 7 symbols in 2 files
Skill Offers — 6 symbols in 4 files
Initialization Lookback — 2 symbols in 1 file
Test signal per changed area — 2 ✓ · 3 ✗Reached = a node under a test path has a resolved edge into the changed symbol. It undercounts anything called indirectly — through a CLI, a spawned process or a dynamic import — so read a low ratio as “look here”, never as a coverage gate.
7 test suites also reference this code7 symbols, kept out of the diagram and the table so they cannot crowd out the areas a reviewer has to look at.
Open the interactive graph → — click an area to see its dependent symbols at file:line. |
Merged
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes from a real end-to-end run of skills with the other features: real Claude Code sessions on pytest-dev/pluggy, with
trail init, learnings, skills, guardrails and the map commands together.What was wrong, and the fix
.venv/bin/pytest,uv run pytest,tox,nox,poetry runandpython -m unittestdidn't match. A leading<dir>/bin/is now stripped before matching, those runners are known, andsource <venv>/bin/activate &&(a plain path in the repo) carries into the steps. Other setup is still refused.git push prod maintook the branch as the target, so the skill wasship-main. A push to a remote named prod, production, staging or live now targets that remote, and the asks the skill was saved from count when matching./hooks. The line only printed in verbose mode. It's now a ⚠ line, so it always prints.Each fix has a test that failed first. Not in here: the offer's "worked out N times" counts every day the job ran, not only passing runs. That count also decides when the offer comes, so it needs its own change.
🤖 Generated with Claude Code