Validate safe current-action continuation - #45
Conversation
![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: aee2c798a8
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".
| snapshot_content, | ||
| state, | ||
| pending_patches_count=len(pending_patches), | ||
| framing_goal_defined=bool(framing_goal), |
There was a problem hiding this comment.
Immediately after flg init, FRAMING.md#Goals contains (to be defined), but _first_section_line does not recognize that value as a placeholder, so bool(framing_goal) is true. Resume and Manifest consequently classify the valid initialization action (Run 'flg frame'...) as needs_reconciliation, while Handoff's separate placeholder-aware parser classifies it as current; this blocks the documented first-run continuation flow and violates the new canonical-action contract.
Useful? React with 👍 / 👎.
| @@ -0,0 +1,23 @@ | |||
| """Unit checks for the current-state freshness classifier.""" | |||
|
|
|||
| from scripts.check_current_state import product_changes_after_state_update | |||
There was a problem hiding this comment.
The CI workflow invokes the installed pytest console script, whose import path does not include the repository root; tests/conftest.py adds only src/, and the editable package also installs only packages under src. Consequently this import (and the analogous one in test_smoke_command_resolution.py) raises ModuleNotFoundError: No module named 'scripts' during collection, so every CI matrix job fails before running tests. Add the repository root to the test path or expose these helpers through an installed package.
Useful? React with 👍 / 👎.
| current_action = resolve_current_action( | ||
| snapshot_content, | ||
| state, | ||
| pending_patches_count=len(pending_patches), |
There was a problem hiding this comment.
After flg review accepts all decisions but before the patch is merged, Context's _pending_patch_summaries excludes that patch using decision_review_status == "accepted", whereas Handoff still counts it because its header remains pending_review. Passing this different count makes a Snapshot action such as Review pending patches resolve to needs_reconciliation in Resume/Manifest but current in Handoff, reintroducing the exact cross-view action disagreement this change is intended to eliminate.
Useful? React with 👍 / 👎.
| print("current-state: refreshed in working tree") | ||
| return 0 | ||
|
|
||
| last_update = _git("log", "-1", "--format=%H", "--", str(STATE_DOC)) |
There was a problem hiding this comment.
In the inspected .github/workflows/ci.yml, actions/checkout@v4 is used without a fetch-depth override, so the job has a one-commit shallow history. In that checkout, git log -- docs/product/current-state.md treats the shallow boundary commit as the file's latest introduction even when the file was not changed there, making last_update equal HEAD and the following diff empty; therefore CI reports the document current for exactly the later product-only commits this check is meant to reject. Fetch sufficient history before relying on this path log.
Useful? React with 👍 / 👎.
1 participant
What changed
Why
The continuation study showed that having project state is not enough when different views expose different next actions. A stale FLG view could be worse than no state because it encouraged a fresh Agent to continue obsolete work. This iteration makes current-action selection deterministic and tests the repair against the same held projects.
Result
Evidence boundary
The three-project result validates this targeted current-action repair on owner-operated projects using one model family. The AML #8 result belongs to the independent FlowGrid AML Retriever under the leaderboard Add/Search contract. Neither result establishes FlowGrid Core adoption, retention, product-market fit, or universal superiority.