Repository navigation
fix: v0.10.1 — realistic time-box defaults; find the worker's worktree without cd - #4
Merged
Conversation
…e without cd - Defaults 40k/60k tokens, 10/15 min killed real RED/GREEN attempts (60–190k even on success) before they wrote a line; now 150k/250k tokens, 20/30 min. - The budget hook located the worker's worktree from the tool call's cwd only, so a worker using `git -C <wt>` / absolute paths from the main repo never had its .agentic/task.env override applied. It now finds the worktree from a binding, the paths the call names, the cwd, or the single unclaimed worktree for the role, and re-resolves the budget when it becomes known (clock kept; a hard verdict reached under the wrong budget is withdrawn). - `budget status` / `budget resolve` / budget messages show where the budget came from. - Regression test scripts/test/budget-worktree.test.py, run in CI. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01YRfiFoSvw3SH9hJZqPhhYK
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
adeelahmad
pushed a commit
that referenced
this pull request
Oct 18, 2026
…rs to the gate's suppression check (run 22) e2e-live.sh (#38) passed the Claude prompt as an argument after --allowedTools, which takes every following word, so `claude -p` got no prompt and exited in 2s ("Input must be provided"). The prompt goes on stdin again; --dry-run says so (test: e2e-live-hosts). Live run 22 (wordfreq, 361f8d7): 13/13, 16m18s, $1.14, 2 wakes, 0 plugin reads, 3 blocks. The final-gate persona failed the sprint on RED's #![allow(dead_code)] in tests/common/mod.rs by its own open-ended grep, costing a GREEN re-dispatch. The persona now names exactly assert_no_suppression's patterns and defers to `selfcheck final-gate` (test: suppression-contract.test.py; eval final-gate #4). Ledger, ROADMAP M0.8, TRACKING, CHANGELOG. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015QxyVf3NgwsgRyYpBgDDTC
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.
Problem (reported from a real sprint)
Fifteen RED/GREEN worker attempts died with zero commits, for two reasons.
cwd. Workers that usegit -C <wt>or absolute paths, instead ofcd-ing into the worktree, always reported the main repo as their cwd. So the hook never read the worktree's.agentic/task.env(BUDGET_*) and applied the tiny default for the whole attempt.The workaround at the time was a repo-level
docs/agents/defaults.md, which resolves against the main tree.Fix
New defaults: 150k / 250k tokens, 20 / 30 min.
Worktree discovery (
budget→find_worktree) checks these in order:-C,cd,--work-tree, absolute paths,file_path, patch headers.task.envnames this role.The winning worktree is remembered for the rest of the attempt.
Late discovery: if the worktree is found after the first call, the budget is re-resolved without resetting the attempt's clock. A hard verdict reached under the wrong budget is withdrawn, and its marker is removed.
Diagnostics:
budget statusshows each attempt's budget source (tasks.md line,task.env,defaults.mdor built-in) and how its worktree was found.budget resolve --worktree WTprints the source too.Test plan
scripts/test/budget-worktree.test.py, 7 checks. It coversgit -Cfrom the main repo, absolute paths, a first read of a main-tree file, late discovery (clock kept, hard verdict withdrawn), the status output, and project-widedefaults.md.make ci(Rust, shellcheck, Python compile check, evals, host tests: 7 + 15 + 18 checks) andclaude plugin validate --strict🤖 Generated with Claude Code
https://claude.ai/code/session_01YRfiFoSvw3SH9hJZqPhhYK
Generated by Claude Code