Skip to content

Correct PRODUCT.md on tool-less turns #955

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
TheGreatAxios merged 2 commits into from
Sep 13, 2026
+4 −4
Merged

Correct PRODUCT.md on tool-less turns #955

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
66afc97
Correct Deterministic progress value prop on tool-less turns
TheGreatAxios Sep 12, 2026
e32a736
Remove auto-pause fiction from Stall failure mode
TheGreatAxios Sep 13, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
8 changes: 4 additions & 4 deletions docs/PRODUCT.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -34,7 +34,7 @@ The evidence is in how the product fails today: the personas already produce exc

## Key Value Propositions

1. **Deterministic progress** — Every turn must produce a tool call. No idle thinking; the director aborts a stalled run rather than spinning.
1. **Deterministic progress** — Primary chat is multi-turn and text-only turns are legal; sub-agent workers complete on a tool-less turn plus the report envelope. Loop protection is a wrap-up nudge after a long tool-only streak, not a hard abort for missing tool calls.
2. **Task tracking** — The agent can maintain a `manage_tasks` checklist for multi-step work; non-interactive `submit_output` is blocked while checklist items remain open. (A "task" here is a work item, not a child agent — spawning uses the separate `spawn_agent` fleet-agent surface.)
3. **Stall detection** — The director detects idle cycles and intervenes.
4. **Safe by default** — Consequential actions (writes, edits, shell) pass a permission gate; secret files and catastrophic commands are denied outright, regardless of intent.
Expand Down Expand Up @@ -123,11 +123,11 @@ Config-driven `postTurn` and `postRun` hooks (TypeScript or shell) run automatic

### Stall (tool-only turns with no narration)

**What the user sees:** The agent runs several turns in a row that are all tool calls with no explanation of what it's doing. After a one-shot nudge to explain itself, if the pattern continues the session **auto-pauses**: it stops issuing new inferences and replies with "Auto-paused: the model ran N steps in a row without explaining its progress. Send a message to resume." The session is not aborted — sending any message resumes it.
**What the user sees:** The agent runs several turns in a row that are all tool calls with no explanation of what it's doing. After a configured number of such turns the director injects a one-shot wrap-up nudge asking for an explanation — a check-in, not a stop. There is no automated stop or pause: the session keeps running until the operator intervenes.

The exact turn thresholds are model-family-dependent (tighter for models with observed runaway tool-only behavior); see "Main-session loop protection" in `docs/ARCHITECTURE.md`.
The exact turn threshold is model-family-dependent (tighter for models with observed runaway tool-only behavior); see "Main-session loop protection" in `docs/ARCHITECTURE.md`.

**Recovery:** Send a message to resume. To inspect state first, see `~/.corbits/projects/<project-key>/<session-id>/run.json` (or a legacy in-repo `.agent-state/` tree if not yet migrated).
**Recovery:** Nothing to recover — the session never pauses. To redirect it, send a message (steer) or interrupt. To inspect state, see `~/.corbits/projects/<project-key>/<session-id>/run.json` (or a legacy in-repo `.agent-state/` tree if not yet migrated).

### Permission denied (exec)

Expand Down
Loading