fix(sensorium): gate stale pointers and add volunteer cards - #48
Conversation
Preserve the live Sensorium pointer/outbox repairs on a clean branch.\n\n- add foreground pointer backoff/relevance gates and saved-residue doorway handling\n- classify historical prepared context-pointers as settled/non-actionable in dashboard surfaces\n- add compact volunteer cards/status orientation helpers\n- update plugin registration/tests/docs for the split tool surface\n\nVerification:\n- pytest -q
Verification:\n- ruff check agent_sensorium tests\n- pytest -q
![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: 8600e64da9
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".
| is_saved_residue = ( | ||
| candidate.get("status") == "archived" | ||
| and decision in {"SAVE", "PROMOTE_CONSCIOUS"} | ||
| and settlement.get("intake_task_id") |
There was a problem hiding this comment.
When a candidate is settled via apply_kanban_settlement with SAVE or PROMOTE_CONSCIOUS, agent_sensorium/settlement.py marks it reviewed, not archived. Because this exact-subject branch only recognizes archived rows, sensorium(action='status', reference_id=<saved candidate>) returns kind: candidate and omits kanban_settlement for the normal saved-settlement path, even though volunteer cards/pointers surface the same row as saved residue. This breaks the new reference-id recovery path for ordinary Kanban-saved candidates; key this off the settlement metadata or include reviewed here.
Useful? React with 👍 / 👎.
| last_presented = next( | ||
| ( | ||
| receipt for receipt in reversed(prior_presented) | ||
| if receipt.get("pointer_type") in {"candidate", "saved_residue"} | ||
| ), |
There was a problem hiding this comment.
When two live sessions share the same Sensorium store, this min-turn-gap check looks at the last candidate/saved-residue pointer from any session, while current_turn_index is computed per session_id. A receipt from another session at turn 5 can therefore suppress a relevant pointer on turn 1 of a new session (1 - 5 < min_turn_gap) even though that session has not seen a card yet. Filter this lookup to session_receipts when session_id is present so one conversation cannot starve another.
Useful? React with 👍 / 👎.
| res = apply_media_gift_choice( | ||
| store=store, | ||
| decision=decision, | ||
| why_now=why_now, | ||
| artifact_id=artifact_id, |
There was a problem hiding this comment.
For decline, choose_silence, and block_delivery, apply_media_gift_choice does not require artifact_id to exist, and _update_artifact_delivery_state silently no-ops if it does not. A typo in /artifacts/{artifact_id}/triage therefore returns success and writes a decision receipt against a nonexistent artifact while the intended held artifact remains held. Check that the artifact exists before applying these choices, not only on approve_delivery.
Useful? React with 👍 / 👎.
| allowed_prefixes = [ | ||
| Path("/home/entity/.hermes").resolve(), | ||
| Path("/home/entity/.gemini/antigravity-cli/scratch").resolve(), | ||
| Path("/tmp").resolve(), | ||
| ] |
There was a problem hiding this comment.
This verifier only trusts /home/entity/.hermes, but the dashboard and store default to Path.home() / ".hermes". In deployments/tests running under another account (for example /root/.hermes, a macOS home, or a service user), legitimate held text/image artifacts under the actual Hermes state tree are reported as UNVERIFIED/security_warning instead of VERIFIED_COMPLIANT. Build the allowed prefix from Path.home()/DEFAULT_ROOT rather than hard-coding one username.
Useful? React with 👍 / 👎.
| try: | ||
| from agent_sensorium.inner_life import signal_inbox_lineage_violations | ||
|
|
||
| raw_violations = signal_inbox_lineage_violations( | ||
| {"version": 2, "blocks": blocks_obj}, |
There was a problem hiding this comment.
This import is never satisfiable in this commit because agent_sensorium.inner_life does not define signal_inbox_lineage_violations, and the broad except then makes every /topology response report signal_inbox_lineage.ok=false with lineage_checker_unavailable. As a result, configured sensor-to-signal_inbox gaps are never actually checked and operators only see a missing-checker diagnostic. Add/export the checker or call an existing implemented API here.
Useful? React with 👍 / 👎.
| with open(resolved_path, "r", encoding="utf-8", errors="ignore") as f: | ||
| line1 = f.readline() | ||
| line2 = f.readline() |
There was a problem hiding this comment.
For allowed text/image artifacts, readline() without a size limit can read an entire file when the file has no early newline. Since artifact refs can point at regular files under /tmp, one large one-line file is enough to make /snapshot allocate a huge string or stall while merely checking the two required header markers. Read a bounded prefix/header instead of unbounded lines.
Useful? React with 👍 / 👎.
|
Closing this stale broad branch rather than reviving it. Its accepted feature slices were integrated through the later #49-#58 train, while the branch itself is now conflicting and carries unresolved P1/P2 findings. The still-relevant current-main pointer continuity findings are repaired narrowly in #68, and the remaining generic dashboard/profile boundary findings are repaired narrowly in #69. This preserves the useful intent without merging the obsolete 4,700-line proposal. |
1 participant
Summary
Verification
pytest -qCleanup context
This is the first upstream-preservation slice from the dirty live checkout checkpoint
backup/sensorium-dirty-20260708-154949/stash@{0}.