You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: docs/architecture/deep-review.md
+24-15Lines changed: 24 additions & 15 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -1,14 +1,23 @@
1
-
# DeepReview Architecture
1
+
# DeepReview / Strict Review Architecture
2
2
3
3
## Scope
4
4
5
-
DeepReview is a child-session workflow that runs a configurable Code Review Team against a review target. The current implementation has three layers:
5
+
DeepReview is the compatibility runtime for `Review: Strict`, the highest-strength mode of the unified Review experience. It remains implemented as a child-session workflow that runs a configurable read-only reviewer set against a review target, but it should not be presented as a second ordinary product entry next to Review.
6
+
7
+
Product-facing guardrails live in [review-verify-implementation-guardrails.md](../sdlc-harness/review-verify-implementation-guardrails.md):
8
+
9
+
-`Review` is the primary user-facing entry.
10
+
-`/review` is the intended long-term command entry; `/DeepReview` is only a transitional typed compatibility command for historical strict-review launches.
11
+
-`ReviewTeam` is an internal strict-review reviewer-set configuration, not a separate product concept users must learn.
12
+
- PR Review consumes review/verify results and readiness projections; it must not own another reviewer executor.
13
+
14
+
The current implementation has three layers:
6
15
7
16
- Frontend launch and UI orchestration in `src/web-ui`.
8
17
- Platform adapter commands in `src/apps/desktop/src/api/agentic_api.rs`.
9
18
- Platform-agnostic runtime policy, task admission, queue state, retry metadata, and report enrichment in `src/crates/assembly/core/src/agentic`.
10
19
11
-
The backend does not choose the review target or build the launch manifest. The frontend builds the effective `ReviewTeamRunManifest`, persists it on the DeepReview child session, and sends it with the first user message.
20
+
The backend does not choose the review target or build the launch manifest. The frontend builds the effective `ReviewTeamRunManifest`, persists it on the DeepReview child session, and sends it with the first user message. The manifest, session kind, agent type, storage keys, and queue event names stay compatible with historical DeepReview sessions.
12
21
13
22
## Runtime Roles
14
23
@@ -29,14 +38,14 @@ The reviewer agents use instruction-only context and read/search/git/diff tools.
29
38
30
39
## Launch Flow
31
40
32
-
DeepReview can be launched from session-file review controls or a `/DeepReview`slash command.
41
+
Strict Review can currently be launched from session-file review controls or the canonical typed `/review strict` command. Historical `/DeepReview`and `/deepreview` inputs remain compatibility aliases that route into the same strict Review path, but new product entry points should route through unified Review and `/review`, then select strict review only when task scope, difficulty, risk, explicit quality intent, and budget justify L3.
33
42
34
43
Frontend launch code lives in `src/web-ui/src/flow_chat/deep-review/launch`:
35
44
36
-
-`commandParser.ts` identifies `/DeepReview` commands and optional file or git targets.
45
+
-`commandParser.ts` identifies canonical `/review strict` commands, transitional `/DeepReview` compatibility aliases, and optional file or git targets.
37
46
-`targetResolver.ts` resolves slash-command targets from git status, changed files, and diffs when a workspace is available.
38
47
-`launchPrompt.ts` formats the user-facing launch prompt.
39
-
-`DeepReviewService.ts` builds the review-team manifest, creates a child session, opens it in the auxiliary pane, sends the launch prompt, and inserts the parent-session summary marker.
48
+
-`DeepReviewService.ts` builds the review-team manifest, creates a child session, sends the launch prompt, and inserts the parent-session summary marker. Launch does not automatically open the auxiliary pane; the summary-card detail action is the normal user-facing way to inspect the background review run.
40
49
-`src/web-ui/src/flow_chat/services/DeepReviewService.ts` is a compatibility re-export.
41
50
42
51
`launchDeepReviewSession` creates a child session with:
@@ -49,11 +58,11 @@ Frontend launch code lives in `src/web-ui/src/flow_chat/deep-review/launch`:
49
58
- context compression enabled
50
59
-`deepReviewRunManifest` stored on the child session metadata
51
60
52
-
If launch fails after the child session is created, the frontend closes the auxiliary pane, deletes the backend session when possible, discards local session state, and reports cleanup issues with the launch error.
61
+
If launch fails after the child session is created, the frontend runs idempotent UI/session cleanup, deletes the backend session when possible, discards local session state, and reports cleanup issues with the launch error.
53
62
54
-
## Review Team Configuration
63
+
## Strict Reviewer Configuration
55
64
56
-
The default review team contract is mirrored in Rust and TypeScript.
65
+
The default strict reviewer configuration contract is mirrored in Rust and TypeScript.
57
66
58
67
Rust source:
59
68
@@ -72,8 +81,8 @@ The desktop command `get_default_review_team_definition` returns the backend def
72
81
The persisted config path is `ai.review_teams.default`. The frontend config shape includes:
73
82
74
83
- extra subagent ids
75
-
-team strategy level
76
-
- per-member strategy overrides
84
+
-review strategy level
85
+
- per-reviewer strategy overrides
77
86
- reviewer and judge timeouts
78
87
- reviewer file-split threshold
79
88
- max same-role instances
@@ -83,7 +92,7 @@ The persisted config path is `ai.review_teams.default`. The frontend config shap
83
92
- provider capacity queue enablement
84
93
- bounded auto-retry enablement and elapsed guard
85
94
86
-
Extra team members must be enabled subagents with read-only review tooling. Core team members, `DeepReview`, and `ReviewFixer` are disallowed as extra members.
95
+
Extra reviewers must be enabled subagents with read-only review tooling. Core reviewers, `DeepReview`, and `ReviewFixer` are disallowed as extra reviewers.
87
96
88
97
## Manifest Shape
89
98
@@ -290,6 +299,6 @@ The review action bar persists UI state separately through `ReviewActionBarPersi
290
299
When changing DeepReview behavior, update all affected contracts together:
291
300
292
301
- Backend constants, team definition, execution policy, manifest gate, task adapter, queue events, and report enrichment.
0 commit comments