Add safe agent implant scaffold and CLI hook patch#10
Conversation
…n stats - `crumb compress`: two-stage compression pipeline inspired by Google's TurboQuant. Stage 1 (PolarQuant-like) does semantic dedup, Stage 2 (QJL-like) does signal-scored pruning. Configurable target ratio. - `crumb bench`: gamified compression scoring — grades crumbs A-F on density, compressibility, structure, and conciseness (out of 100). - `crumb inspect` now shows token cost, content density, keyword density. - `crumb dream` now reports compression ratio after consolidation. CLI now at 26 commands. https://claude.ai/code/session_018rgag4FMarjfsCwLpaRXSZ
- `crumb from-git`: auto-generate task crumb from git history — recent commits, changed files, branch context. Instant handoff for any dev switching contexts mid-branch. - CI bench action (.github/workflows/bench-pr.yml): runs `crumb bench` on all .crumb files in PRs, posts score table as a comment. Reusable workflow + copy-paste template for other repos. - VS Code extension (vscode-extension/): command palette commands (new task, validate, compress, bench, "crumb it"), .crumb syntax highlighting, and 5 snippet templates. Ready for marketplace publish. CLI now at 27 commands. https://claude.ai/code/session_018rgag4FMarjfsCwLpaRXSZ
* feat: ship MeTalk and PyPI-ready crumb-format 0.2.0 * fix: install crumb-format from checkout in CI workflows * chore: auto-generate .crumb handoff for PR #3 * fix: use local reusable bench workflow in CI template * chore: auto-generate .crumb handoff for PR #3 --------- Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
* feat: add crumb REST API for cross-AI access * chore: auto-generate .crumb handoff for PR #4 --------- Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
CRUMB Bench Results
Powered by crumb-format — TurboQuant-inspired context compression |
CRUMB Bench Results
Powered by crumb-format — TurboQuant-inspired context compression |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d1379fd4a6
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| def cmd_agent(args: argparse.Namespace) -> None: | ||
| action = args.agent_action |
There was a problem hiding this comment.
cmd_agent is implemented here but never reachable because this commit only ships the parser wiring as a text patch (patches/cli_crumb_agent.patch) instead of modifying cli/crumb.py itself. In the current tree, crumb --help still has no agent command, so all crumb agent template/run/eval flows described in the PR are non-functional until a separate manual patch step is done.
Useful? React with 👍 / 👎.
| "agent": asdict(experiment.agent), | ||
| "backend": asdict(experiment.backend), | ||
| "metadata": experiment.metadata, | ||
| "seed": experiment.seed, |
There was a problem hiding this comment.
stable_run_id hashes title/goal/context/constraints/agent/backend/metadata/seed but omits experiment.evaluation, so two experiments with different required/banned terms resolve to the same run directory and overwrite each other’s metrics.json/log artifacts. This makes run outputs non-isolated and can silently corrupt experiment comparisons when only evaluation criteria change.
Useful? React with 👍 / 👎.
2 participants
This PR uploads the safe CRUMB-native agent implant discussed in chat.
Included in this branch:
cli/agent.py— safe local agent scaffold withtemplate,run, andevalactionsexamples/agent_experiment.json— starter experiment specpatches/cli_crumb_agent.patch— exact patch to wirecrumb agent ...intocli/crumb.pyNotes:
cli/crumb.py; the connector path used here allowed file creation cleanly, so I included the live patch file in-repo for application/review.Suggested follow-up after merge/apply: