Skip to content

Fix findings from a Claude Code and Codex dogfood run - #1

Merged
mo4islona merged 6 commits into
mainfrom
fix/dogfood-findings
Oct 29, 2026
Merged

mo4islona merged 6 commits into
mainfrom
fix/dogfood-findings

Conversation

@mo4islona

@mo4islona mo4islona commented Oct 29, 2026 •

Copy link
Copy Markdown
Owner

A dogfood run installed a wiki for a fake company in isolated homes, ran a simulated incident in Claude Code 2.1.312 and Codex 0.165, and followed it to a filed candidate. The loop worked end to end in Claude Code. These fixes cover what broke or misled on the way.

  • Candidates from a sandbox without network were lost in practice. Codex's default sandbox blocks the network, so candidate file kept the text in candidates.md, and nothing ever sent it. The new wikiloop candidate send command files the queue and keeps what still fails. While the queue holds entries, the session rule tells the user how to send them. The queue file is locked, so a parallel send cannot drop a new entry.
  • candidate search crashed with a traceback when the tracker was out of reach. It now says that the open candidates are unknown, and it exits 1.
  • Trigger evals ran with the maintainer's own setup. User settings brought other plugins, hooks and MCP servers into each run, and an installed copy of the same wiki competed with the checkout. The runs now use --setting-sources project,local --strict-mcp-config --no-session-persistence, so they also save no transcripts.
  • The installer ignored CLAUDE_CONFIG_DIR. With it set, the auto-update step failed after a successful marketplace add, and has_mcp read the wrong file.
  • wikiloop init now points to ./install.sh, as the README does.

Not in this PR: Claude Code labels the Stop hook's block as "Stop hook error", the stop text repeats the long plugin path, and Codex shows each hook as "Unnamed hook".

Test plan

  • python3 -m unittest discover -s tests -t .: 133 tests, OK (1 skipped). New: the queue and send against a fake gh, search offline, a config from an older core, and CLAUDE_CONFIG_DIR in the installer.
  • Example: build --check --strict, lint, check all pass.
  • End to end: candidate send filed the entry that the Codex run had queued to a real test repo, with the label and without the queue trailer. The queue was empty afterwards.
  • One eval case, run against the dogfood wiki: score 1.00. The init event listed only the wiki and the built-in plugins, with no MCP servers, and no transcript was saved.
  • CLAUDE_CONFIG_DIR paths confirmed in a scratch home: .claude.json and plugins/known_marketplaces.json both go to that directory.

🤖 Generated with Claude Code

mo4islona and others added 6 commits October 29, 2026 15:21
… reach

The local fallback had no way back: a Codex session in its default sandbox has no
network, so each candidate it filed stayed in candidates.md, and nobody was told how
to send it. A search in that sandbox crashed, so the agent could not see an open
duplicate either.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
User settings bring the maintainer's other plugins, hooks and MCP servers into each
run, and an installed copy of the same wiki competes with the checkout. Each run also
left a transcript in ~/.claude/projects.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With it set, Claude Code keeps the marketplace list and .claude.json in that
directory, so the auto-update step failed right after a successful marketplace add,
and the MCP check looked in the wrong file.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Codex 0.165 lists the plugin's hooks as new or changed on the start after the first
trust; the second trust holds.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@mo4islona
mo4islona merged commit cceb02e into main Oct 29, 2026
3 checks passed
@mo4islona
mo4islona deleted the fix/dogfood-findings branch October 29, 2026 12:49
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