root/fusion
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

FN-199: request detailed reasoning summaries

Browse Source 
Request detailed reasoning bodies for Responses-family model traces while preserving provider fallback behavior.\n\n- Shape supported Responses payloads to request detailed reasoning summaries.\n- Retry once with the prior payload when detailed summaries are rejected.\n- Preserve titled trace bodies across dashboard surfaces and document the setting.\n\nFiles changed:\n .changeset/fn-199-reasoning-summary.md             |   7 +
 docs/dashboard-guide.md                            |   2 +-
 docs/settings-reference.md                         |   4 +
 .../__tests__/ThinkingTrace.surfaces.test.tsx      |  19 +++
 .../components/__tests__/ThinkingTrace.test.tsx    |   4 +
 .../src/__tests__/pi-reasoning-summary.test.ts     | 150 +++++++++++++++++++++
 .../__tests__/reasoning-summary-payload.test.ts    | 102 ++++++++++++++
 .../src/execution/reasoning-summary-payload.ts     |  72 ++++++++++
 packages/engine/src/pi.ts                          |  48 +++++++
 9 files changed, 407 insertions(+), 1 deletion(-)

Fusion-Task-Id: FN-199

Fusion-Task-Lineage: a3179f05-e709-4bae-9f4a-5507810717d8

Co-authored-by: Fusion <noreply@runfusion.ai>
This commit is contained in:
Fusion Agent
2026-08-27 11:15:44 +00:00
parent c31c15a8cd
commit 808f4c603b
9 changed files with 407 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
docs/dashboard-guide.md
View File

@@ -110,7 +110,7 @@ Press `Escape` to close the current/topmost dashboard popup. Popped-out task win
<!-- FNXC:ChatFindDocs 2026-08-21-16:29: FN-110 gives the active visible Chat host contextual Find ownership without adding a configurable dashboard shortcut. -->
### Thinking traces
Thinking panes split titled reasoning traces with captured bodies into independently expandable sections, with **Collapse all** and **Expand all** controls. Headings without captured reasoning stay inline in the flowing trace instead of becoming empty collapsible rows, and the defensive empty-state label appears at most once per section. **Raw trace** shows the original unsectioned capture and switches back with **Sectioned trace**; it is available for every trace that contained headings, including a titles-only trace rendered as one flowing block. The workflow live-log console remains raw and unsectioned by design. The same behavior applies while Planning Mode, Mission Interview, and Milestone/Slice Interview stream a generation.
Thinking panes split titled reasoning traces with captured bodies into independently expandable sections, with **Collapse all** and **Expand all** controls. Responses-family models now request titled sections with their reasoning bodies. A titles-only trace can still arrive from a provider Fusion cannot configure; those headings stay inline in the flowing trace instead of becoming empty collapsible rows, and the defensive empty-state label appears at most once per section. **Raw trace** shows the original unsectioned capture and switches back with **Sectioned trace**; use it to diagnose a provider-side titles-only payload. The workflow live-log console remains raw and unsectioned by design. The same behavior applies while Planning Mode, Mission Interview, and Milestone/Slice Interview stream a generation.
### Chat Find

4
docs/settings-reference.md
View File

@@ -103,6 +103,10 @@ Fusion persists one canonical ordered vocabulary: `off`, `minimal`, `low`, `medi
Custom-provider models are presumed thinking-capable and expose all seven levels by default; no capability checkbox or other configuration is required. Fusion registers these levels with pi as transmissible, while pi owns API-specific `off` translation and up-then-down clamping. Consequently, a global or inherited thinking level that custom providers previously ignored is now sent to the gateway. A strict gateway can reject an unsupported effort; select **Off** for that gateway, which pi translates to its explicit non-reasoning form or omits when the API has no such form.
#### Reasoning-summary detail
When a real thinking effort is resolved, Fusion requests `reasoning.summary: "detailed"` for Responses-family models: `openai-responses`, `openai-codex-responses`, and `azure-openai-responses`. Pi forwards the effort through its simple stream path but not this summary preference; without Fusion's request shaping, those APIs use their shortest `"auto"` summary. Non-Responses APIs are unchanged. If a provider rejects detailed summaries, Fusion retries that run with the previous request shape. Detailed summaries can consume slightly more output tokens.
| `ntfyEnabled` | `boolean` | `false` | Enable ntfy push notifications. |
| `agentClarificationEnabled` | `boolean` | `false` | Legacy default for programmatic Planning Mode session notification eligibility. Dashboard Planning Mode always starts its infinite, user-validated interview with follow-up questions enabled; this setting no longer suppresses questions or creates a final summary. |
| `failureNotificationMode` | `"sticky-only" \| "terminal-only" \| "all"` | `"sticky-only"` | Failure notification behavior. `sticky-only` defers failed-task notifications by `failureNotificationDelayMs` and suppresses transient self-recoveries. `terminal-only` suppresses while auto-retry is still active and only dispatches when `paused === true` or `column === "in-review"` with `status === "failed"`. `all` restores legacy immediate failure notifications. |