fix(webui): restore smooth streaming and preserve model phases

[serrrfirat]

Summary

• Keep one projection subscription alive for the complete Axum SSE connection, eliminating the event-loss window created by repeatedly tearing down one-event subscriptions.
• Emit typed application-level keep_alive frames every 15 seconds while a subscription is idle so the browser watchdog sees healthy liveness; Axum comments continue protecting proxies.
• Bound the generic ProductSurface first-event wait so non-SSE callers cannot hang indefinitely on a silent subscription.
• Use a one-slot, backpressured product handoff and revalidate stream access independently every second so authorization I/O does not gate token frames.
• Replace the 75 ms cumulative-text window with a 16 ms latest-snapshot window: ordinary provider cadence passes through while sub-frame microbursts are conflated.
• Preserve intermediate assistant utterances by assigning each model-call phase a distinct live-text identity and replacing only the latest phase at finalization.
• Render active assistant Markdown with streamdown@2.5.0; completed replies retain marked + DOMPurify.
• Delegate browser SSE framing, Last-Event-ID, abort, and retry/backoff to event-source-plus@0.1.15 over fetch/ReadableStream. React retains only application policy such as connection status, terminal error classification, visibility, and active-run stall recovery.
• Send the SSE bearer in the Authorization header rather than the request URL, and increment connection_generation for every package-managed request.
• Suppress empty intermediate assistant phases so they cannot split adjacent tool activity or reserve blank transcript rows.
• Keep activity disclosures user-controlled: incoming reasoning, failed, and declined tool events update the summary without opening the collapsed details.
• Keep timestamp/copy controls on user and final assistant messages without reserving their 28 px row on intermediate utterances.
• Preserve one React row identity across live, final, and timeline-refreshed forms so final handoff does not remount or flash.

Change Type

• Bug fix
• New feature
• Refactor
• Documentation
• CI/Infrastructure
• Security
• Dependencies

Linked Issue

None — reported through Railway preview testing.

Validation

• cargo fmt --all --check
• cargo clippy -p ironclaw_host_api -p ironclaw_product -p ironclaw_webui --all-targets --all-features -- -D warnings
• cargo build — not run separately; test and clippy builds covered affected crates and targets.
• Relevant tests pass:
  • cargo test -p ironclaw_host_api
  • cargo test -p ironclaw_product --test reborn_services_contract -- --skip trace_reads_are_available_as_product_views (248 passed)
  • cargo test -p ironclaw_webui --all-features
  • cargo test -p ironclaw_architecture
  • 906 frontend tests
  • 42/42 Reborn WebUI v2 browser smoke cases, including packaged fetch-stream retry, terminal response, approval gate, and delayed activity ordering
  • pnpm lint
  • pnpm typecheck
  • pnpm build (/chat 208.1 KB gzip, 1.9 KB budget headroom)
  • git diff --check
  • cargo clippy --all --benches --tests --examples --all-features
• cargo test --features integration — not applicable: no database, persistence backend, runtime lane, WASM, or MCP behavior changed.
• Manual testing:
  • On 931d6b8e5, synchronized Railway wire/DOM capture observed roughly 690 cumulative text snapshots in about one minute while the browser showed the phase only after terminal timeline refresh. This proved backend delivery remained lossless but left the client reception path unresolved.
  • On 9761db3ee, a 16 ms producer coalescer still produced zero visible continuation updates for 55 seconds. This ruled out the producer queue/coalescer as the root cause.
  • On b498051b0, after replacing the custom native EventSource lifecycle, a fresh Railway run produced 210 visible DOM changes in 55 seconds.
  • The previously failing existing-thread follow-up produced 57 visible DOM changes in 55 seconds without navigation. Intermediate phases remained ordered and visible through terminal completion.
  • On c4c7b98d5, the canonical PR-6876 Railway preview rendered five activity groups and five intermediate utterances with a consistent 20 px row gap, zero empty intermediate rows, and zero intermediate timestamp/copy rows. The final reply upgraded to data-final-reply=true and retained its timestamp/copy controls.
  • On a0a97d8c8, a deterministic .invalid HTTP request produced Activity - 1 tool, 1 failed while the toggle remained aria-expanded=false and zero detail rows were mounted. Manual expansion exposed the failure details, and manual collapse restored zero mounted detail rows.
  • The live model initially timed out for 60 seconds waiting for response headers and later emitted the final answer in long bursts; scoped logs attribute that quiet period to the external provider/retry path before browser-visible content existed.
• If a coding agent was used and supports it, review-pr or pr-shepherd --fix was run before requesting review.

Additional dependency evidence:

• streamdown@2.5.0 is Apache-2.0.
• event-source-plus@0.1.15 is MIT-licensed and passed the repository supply-chain policy during installation.
• pnpm audit --prod reports existing React Router advisories; the new SSE package is not on those advisory paths.

Local environment limitations:

• CARGO_PROFILE_TEST_DEBUG=0 CARGO_INCREMENTAL=0 cargo test --workspace --lib built the full workspace and ran the library suites, but finished non-green on two untouched Trace Commons enrollment tests in ironclaw_host_runtime (543 passed; 2 failed) because this environment returned NetworkDenied instead of the expected unenrolled guidance.
• bash scripts/pre-commit-safety.sh no longer flags files added by the review pass, but remains non-green on five pre-existing findings elsewhere in this PR branch (filesystem/runtime/composition changes outside this streaming fix).
• trace_reads_are_available_as_product_views fails unchanged in isolation because this workstation has a Trace Commons enrollment and the test reaches its external account endpoint instead of the expected unenrolled zero-state. All other 248 product service contracts pass.
• The production-shaped composition canary stalls during local runtime startup in macOS Keychain (SecItemCopyMatching) before reaching its streaming assertion. CI and Railway provide production-composition verification.

Test Strategy

User behavior:

Given an active chat using a streaming model, when the provider emits cumulative assistant text and tool milestones, the browser renders text continuously, preserves intermediate model phases, and transitions the latest phase in place to sanitized final Markdown. The user must not need to switch threads to reveal buffered output.

Risk areas:

• Model behavior
• Browser
• Side effect
• Persistence
• Security or permissions
• External provider
• Cross-component behavior

Tests added or updated:

• Unit or contract:
  • Product contracts prove a subscription-capable projection uses one continuous subscription and retains the legacy one-shot drain path for compatibility callers.
  • Product contracts prove access revocation is detected on an independent control cadence without authorization reads in the event hot path.
  • Host API contract proves the process-local continuation is omitted from serialization; unit coverage exercises its debug and receiver-identity contracts.\n - Product live-stream coverage proves terminal milestones clear per-run text-phase state.\n - Changed-line coverage sabotage tests exclude nested cfg(test) modules while still failing closed for absent production sources.
  • Axum handler contract sends two back-to-back events through one product stream invocation with no poll/resubscribe gap.
  • Frontend contracts cover empty intermediate-phase elision, attachment preservation, compact intermediate metadata, package-managed request generations, authorization headers, retryable and terminal responses, replay rebasing, visibility, active-run stall recovery, route disposal, provider-rate updates, multi-phase ordering, stable row identity, collapsed reasoning/failure updates, manual activity disclosure, streaming Markdown, final handoff, and hostile Markdown.
• Reborn integration: Existing production-shaped composition coverage verifies projection cursor advancement. Full local production composition is blocked by macOS Keychain startup; CI and Railway cover this path.
• Recorded fixture: Not applicable: no model choice, tool arguments, or outbound request shape changed.
• Browser E2E: the smoke harness now emulates the packaged fetch/ReadableStream SSE transport; all 42 WebUI v2 smoke cases pass. Railway raw-SSE/DOM cadence sampling reproduced the zero-update browser failure, then verified both fresh-thread and existing-thread streaming after the packaged client replacement.
• Backend or runtime: Not applicable: no database, persistence backend, runtime lane, WASM, or MCP behavior changed.
• Live canary: Railway commit b498051b0 streamed a fresh tool-using run through 210 visible DOM changes in 55 seconds. An existing-thread follow-up—the reported zero-update case—streamed through 57 visible changes in 55 seconds. Intermediate phases remained ordered and visible through terminal completion.

What the tests prove:

• The Axum SSE connection retains one underlying projection subscription rather than reopening one-event subscriptions.
• Back-to-back intermediate/model/tool events cannot fall into a no-subscriber gap.
• The product handoff is bounded to one waiting response and applies backpressure instead of adding a render-burst queue.
• Authorization is checked at open and independently every second; a slow access-store read does not delay text delivery.
• The continuation handle is process-local and does not alter the serialized product response.
• Each model call has a distinct text identity; finalization and timeline refresh preserve prior intermediate phases while replacing only the latest phase.
• Ordinary 25 ms provider cadence is preserved, while a microburst retains only the latest cumulative snapshot and flushes it before tool activity.
• The packaged SSE client owns parsing, retry/backoff, Last-Event-ID, and abort; each reconnect receives a fresh connection generation.
• The bearer is absent from the SSE URL and is refreshed through an authorization-header function on reconnect.
• Active, final, and refreshed forms of one assistant reply keep a stable React row identity.
• Incoming reasoning and tool-failure updates cannot auto-expand a collapsed activity run; users can still open it explicitly.
• Model-authored script, inline-handler, and unsafe-link payloads do not execute in covered streaming renderer cases.
• Drain-only product surfaces retain fallback polling behavior.

Security Impact

The stream remains caller-scoped. The product layer validates visibility before opening the projection subscription and revalidates it every second on an independent control task; revocation closes the continuation with a redacted product error. The continuation field is serde(skip) and covered by a wire-contract test.

The fetch-based SSE client sends the bearer through the Authorization header instead of a URL query parameter. Active model-authored Markdown passes through Streamdown's streaming renderer; completed replies retain marked + DOMPurify.

Reborn Trust-Boundary Checklist

• Public policy/evidence/trust-bearing types: N/A; no authority-bearing vocabulary changed.
• Untrusted content enters prompts only through an envelope/escaping primitive: N/A; prompt construction is unchanged.
• Hashes declare purpose: N/A; no hashing changed.
• New/changed status, exit, policy, runtime, or error variants: N/A; existing product error mapping is retained.
• Security/durability serde(default) fields fail closed or have migration tests: the process-local continuation is serde(skip) and a contract test proves it never crosses the wire.
• Queues/maps/buffers/counters have bounds and overflow-safe arithmetic: the product handoff has one slot; the event-stream subscription remains bounded.
• Driver/operator-visible errors have stable class semantics: existing projection and product errors are preserved.
• Sandbox/native/host names accurately describe trust boundary: N/A; no runtime boundary changed.

Database Impact

None. No schema, migration, PostgreSQL, or libSQL behavior changed.

Blast Radius

Touches the host/product event-stream response, product projection bridge, Axum SSE forwarding, packaged browser SSE lifecycle, live assistant-text projection, and active Markdown presentation. A regression could delay delivery, close a stream early, or regress Markdown display. Memory remains bounded, Streamdown remains lazy-loaded, and the production bundle remains within budget.

Rollback Plan

Revert the streaming commits on this branch to restore one-shot product polling, the previous native EventSource lifecycle, the prior producer behavior, and the prior active Markdown renderer. No data or schema rollback is required.

Review Follow-Through\n\n- CI closeout on exact head 7c8024a65: all 59 reported checks pass (9 conditional jobs skipped); Railway and CodeRabbit are green.\n- Changed production-line coverage passes at 90.96% (322/354) against the exact PR head. The first post-merge coverage lane hit a Docker Hub connection reset while pulling postgres:16-alpine; GitHub’s failed-job-only retry passed without a code change.\nRailway deployed exact head 3030f4958 with frontend asset assets/app-B3p3T4ll.js after one transient 19-second deployment failure was retried with an empty CI-only commit.

• Fresh-thread NEAR AI canary: 152 distinct visible DOM changes in 55 seconds. Intermediate phases appeared at about 2.3 s, 8 s, and 14.7 s; the final answer then grew continuously to 31 rendered paragraphs without a thread switch or reconnect indicator.
• Existing-thread follow-up: 115 distinct visible DOM changes. A roughly 14.8-second provider/tool quiet period was followed by continuous final-answer growth and terminal completion around 32 seconds, separating external response latency from frontend buffering.
• All five tool-activity groups across the successful canaries remained aria-expanded=false and no reconnect state appeared.
• Deterministic safe failure canary: the .invalid HTTP request produced Activity - 1 tool, 1 failed while the run was active; the failed tool and subsequent assistant/reasoning update left the group collapsed with zero mounted detail rows.
• The failure group remained collapsed through terminal final-answer handoff. Manual expansion exposed the HTTP failure and manual collapse removed the detail rows again.
• The bearer remained UI-entered only; the E2E contract separately proves header-only SSE auth and rejects token query parameters.

---

Review track: C (security-sensitive product/WebUI streaming transport and active Markdown rendering)

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[railway-app]

🚅 Deployed to the ironclaw-pr-6876 environment in ironclaw-ci-preview

Service	Status	Web	Updated (UTC)
ironclaw	✅ Success (View Logs)	Web	Jul 30, 2026 at 1:12 pm

[coderabbitai]

📝 Walkthrough

Summary by CodeRabbit

• New Features

  • Added subscription-backed streaming for Web chat so updates stay live with typed keep-alives and better reconnection behavior.
  • Updated streaming Markdown to use Streamdown for smoother, safer model-generated rendering.

• Bug Fixes

  • Improved authorization handling during active streams via periodic revalidation; streams terminate cleanly on revoked access.
  • Fixed assistant in-progress phase behavior to avoid overwriting earlier text and improved final-reply replacement/ordering during refreshes.
  • Kept intermediate activity panels and phases collapsed by default.

Walkthrough

The PR replaces polling-oriented WebUI event delivery with process-local subscriptions where supported, adds SSE keep-alives and recovery, introduces phase-qualified live assistant text, updates chat rendering and history reconciliation, and adjusts integration, browser, provider, and CI tests.

Changes

WebUI streaming and live projections

Layer / File(s)	Summary
Subscription-backed event delivery crates/ironclaw_host_api/..., crates/ironclaw_product/..., crates/ironclaw_webui/..., tests/integration/...	Stream responses can carry non-serialized subscriptions; product surfaces bridge projection subscriptions with authorization revalidation, while SSE handlers emit events and typed keep-alives and retain drain fallback.
Phase-qualified live text crates/ironclaw_product/src/projection/..., crates/ironclaw_reborn_composition/...	Live text IDs include model-call phases, phase state is cleared at terminal milestones, coalescing uses a 16ms window, and provider-cadence ordering tests are updated.
SSE client recovery and authentication crates/ironclaw_webui/frontend/src/lib/..., crates/ironclaw_webui/frontend/src/pages/chat/hooks/..., tests/e2e/...	The frontend uses EventSourcePlus, sends bearer credentials in headers, handles typed frames and retryable errors, and reconnects stalled active streams.
Phase-aware chat state and rendering crates/ironclaw_webui/frontend/src/pages/chat/..., crates/ironclaw_webui/frontend/package.json, crates/ironclaw_webui/frontend/src/styles/...	History merging, final-reply replacement, grouping, activity expansion, intermediate-message controls, and Streamdown-based Markdown rendering now distinguish intermediate assistant phases.
Compatibility, provider, and CI validation crates/ironclaw_llm/..., crates/ironclaw_reborn_*/tests/..., scripts/ci/...	Provider finish-reason tests, response fixtures, browser stream harnesses, integration replay, and changed-coverage filtering are updated for the revised contracts.

Estimated code review effort: 5 (Critical) | ~120 minutes

Possibly related issues

• Epic: Gateway state convergence — eliminate UI/backend state drift #2792 — The subscription-backed SSE continuity and periodic reauthorization changes align with this issue’s subscription/state-convergence objectives.
• Epic: Hermetic capability and journey testing platform #6524 — The subscription contract tests and stream test doubles contribute to the broader WebUI journey-testing and coverage goals.

Possibly related PRs

• nearai/ironclaw#5743 — Both update WebUI event replay and refresh behavior around stream_events and after_cursor.
• nearai/ironclaw#5821 — Both modify projection-based live assistant-text streaming.
• nearai/ironclaw#6337 — Both modify WebUI SSE recovery and streamed error handling.

Suggested reviewers: ilblackdragon

🚥 Pre-merge checks | ✅ 4

✅ Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	Conventional-commits style and accurately summarizes the WebUI streaming and model-phase changes.
Description check	✅ Passed	All required sections are present and substantially filled, including validation, test strategy, security, rollback, and review track.

✨ Finishing Touches 💡 1

⚔️ Resolve merge conflicts 💡

• Resolve merge conflict in branch codex/fix-stalled-webchat-stream

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[ironloopai]

🔎 Review · PR #6876

🟢 Completed · Review submitted

Submitted review →

Reviewed the complete trusted base-to-head comparison. The product projection subscription change and WebUI active-stream watchdog are coherent with existing cursor replay, authorization, SSE/WebSocket lifetime, and EventSource recovery contracts. No actionable defects found.

_{Automatic · PR opened · attempt 1 of 3 · completed in 2m 15s}

Run details

• Repository: nearai/ironclaw
• Base: main at 2cc9408
• Head: codex/fix-stalled-webchat-stream at a9c545a
• Created: Jul 29, 2026, 8:59 PM UTC
• Updated: Jul 29, 2026, 9:01 PM UTC
• Run: c5e34161-0bda-4bab-80be-37ec22d133f8
• Latest attempt: 1 · Completed · 615fd576-3f8a-4cf9-9f01-4b0cc75a07a9

[ironloopai]

🔍 Review complete · PR #6876

✅ No actionable findings

Reviewed the complete trusted base-to-head comparison. The product projection subscription change and WebUI active-stream watchdog are coherent with existing cursor replay, authorization, SSE/WebSocket lifetime, and EventSource recovery contracts. No actionable defects found.

Validation and technical details

• Inspected all 5 changed files and surrounding product projection, SSE/WebSocket handler, EventSource lifecycle, and chat processing-state code.
• Verified trusted comparison refs/ironloop/base..refs/ironloop/head and ran git diff --check successfully.
• Focused frontend SSE suite passed: 16/16 tests.
• Frontend lint, source-convention checks, and TypeScript typecheck passed.
• Rust tests could not be executed because cargo is unavailable in the review environment; Rust behavior was validated through source and contract-test inspection.
• Base: main
• Head: codex/fix-stalled-webchat-stream at a9c545a
• Run: c5e34161-0bda-4bab-80be-37ec22d133f8

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@crates/ironclaw_webui/frontend/src/pages/chat/hooks/useSSE.ts`:
- Around line 35-40: The active-stream watchdog must not fire during healthy
runs with no runtime payloads. Update the SSE projection/keep-alive flow around
push_runtime_cursor_advance to emit application-level KeepAlive frames
periodically whenever any runtime is active, aligned below
ACTIVE_STREAM_STALL_DEADLINE_MS; alternatively increase the JavaScript deadline
in useSSE.ts beyond the guaranteed app-frame interval while preserving half-open
detection.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 4fe2f503-746a-4bcb-b667-c5e8a125aa63

📥 Commits

Reviewing files that changed from the base of the PR and between 2cc9408 and a9c545a.

📒 Files selected for processing (5)

• crates/ironclaw_product/src/reborn_services.rs
• crates/ironclaw_product/tests/reborn_services_contract.rs
• crates/ironclaw_webui/frontend/src/pages/chat/hooks/useChat.ts
• crates/ironclaw_webui/frontend/src/pages/chat/hooks/useSSE.ts
• crates/ironclaw_webui/frontend/src/pages/chat/lib/useSSE.test.ts

[coderabbitai]

Actionable comments posted: 2

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

crates/ironclaw_product/src/projection/tests/live_progress_stream.rs (1)

335-384: 🎯 Functional Correctness | 🟠 Major | ⚡ Quick win

text_bodies.len() == 64 asserts zero drops from a bounded channel and double-counts snapshot items.

Two coupling problems in one assertion:

1. text_bodies is appended per Text item per envelope (Line 345-351). Because ProjectionUpdate.state.items is a snapshot, any envelope that re-carries the run's Text item — including the one carrying the capability activity — contributes another body. The count therefore measures snapshot fan-out, not "64 preserved updates".
2. All 65 milestones are published before the first subscription.next(), so passing depends on PRODUCT_PROJECTION_SUBSCRIPTION_BUFFER = 128 absorbing the whole burst. Change the capacity or the overflow policy and this test fails for reasons unrelated to the provider-rate contract.

The contract worth locking is smoothness plus ordering, not an exact integer: the terminal body is already asserted at Line 375-377; add monotonic ordering and a floor instead. 0..70 at Line 335 has the same hidden dependency — it only works while the envelope count stays under it.

♻️ De-couple the assertion from buffer capacity and snapshot fan-out

-    assert_eq!(
-        text_bodies.len(),
-        64,
-        "the bounded WebUI stream should preserve provider-rate text updates: {text_bodies:#?}"
-    );
+    let mut distinct = text_bodies.clone();
+    distinct.dedup();
+    assert!(
+        distinct.len() > 1,
+        "provider-rate text must stream incrementally rather than arrive as one final blob: {text_bodies:#?}"
+    );
+    let mut ordered = distinct.clone();
+    ordered.sort_by_key(|body| {
+        body.rsplit(' ')
+            .next()
+            .and_then(|index| index.parse::<usize>().ok())
+            .unwrap_or_default()
+    });
+    assert_eq!(
+        distinct, ordered,
+        "live text updates must stay in provider order: {text_bodies:#?}"
+    );

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@crates/ironclaw_product/src/projection/tests/live_progress_stream.rs` around
lines 335 - 384, Update the live projection stream test around the collection
loop and final assertions to stop treating text_bodies.len() as a count of
preserved provider updates. Track unique or monotonically advancing text bodies,
assert their ordering, and retain a lower-bound smoothness check rather than
exact equality. Make the receive loop continue until the expected terminal
text/tool ordering is observed without relying on a fixed envelope count tied to
subscription buffer capacity.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@crates/ironclaw_product/src/projection.rs`:
- Around line 458-460: The subscriber capacity in SubscriberCapabilities must
use the manager-owned maximum or a named constant exported by
ironclaw_event_streams, replacing the product-local
PRODUCT_PROJECTION_SUBSCRIPTION_BUFFER, and verify overflow retains the newest
cumulative text item. In crates/ironclaw_reborn_composition/CLAUDE.md lines
399-404, remove the unsupported “maximum bounded subscriber buffer” and “16-item
default” claims or cite the owning constant so the documentation remains
checkable.

In `@crates/ironclaw_reborn_composition/CLAUDE.md`:
- Around line 442-445: Update the documentation bullet for
provider_rate_live_text_stays_smooth_and_precedes_tool_activity to describe that
the latest cumulative text is published before the capability activity, rather
than claiming every text update is ordered before the next milestone.

---

Outside diff comments:
In `@crates/ironclaw_product/src/projection/tests/live_progress_stream.rs`:
- Around line 335-384: Update the live projection stream test around the
collection loop and final assertions to stop treating text_bodies.len() as a
count of preserved provider updates. Track unique or monotonically advancing
text bodies, assert their ordering, and retain a lower-bound smoothness check
rather than exact equality. Make the receive loop continue until the expected
terminal text/tool ordering is observed without relying on a fixed envelope
count tied to subscription buffer capacity.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: f321025e-6151-4b0a-9160-f9e0f7e65a4a

📥 Commits

Reviewing files that changed from the base of the PR and between a9c545a and 6b9ce44.

📒 Files selected for processing (4)

• crates/ironclaw_product/src/projection.rs
• crates/ironclaw_product/src/projection/live_progress.rs
• crates/ironclaw_product/src/projection/tests/live_progress_stream.rs
• crates/ironclaw_reborn_composition/CLAUDE.md

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@crates/ironclaw_llm/src/nearai_chat.rs`:
- Around line 2019-2047: Extend the caller-level streaming tests around
complete_search_tool_streaming to cover parsed tool calls with no finish reason,
asserting FinishReason::ToolUse, and an explicit unrecognized finish reason,
asserting FinishReason::Unknown. Build each case through the existing HTTP/SSE
test server and caller wrapper rather than testing classifiers directly, while
preserving the existing terminal text Stop coverage.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: a4ab7340-7770-489b-8d89-6c93a51a6d8d

📥 Commits

Reviewing files that changed from the base of the PR and between 6b9ce44 and 5f3ab32.

📒 Files selected for processing (2)

• crates/ironclaw_llm/CLAUDE.md
• crates/ironclaw_llm/src/nearai_chat.rs

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@crates/ironclaw_product/src/projection/live_progress.rs`:
- Line 48: Replace the Mutex protecting text_phase_by_run in the live progress
state with an RwLock, using read access when retrieving the existing text_id and
write access when creating or removing phase entries. Preserve the current
phase-state behavior while updating all corresponding lock calls and
initialization.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 97e1d51a-2fca-488d-b809-40c7fc6da0ae

📥 Commits

Reviewing files that changed from the base of the PR and between 5f3ab32 and 2586a17.

📒 Files selected for processing (10)

• crates/ironclaw_product/src/projection/live_progress.rs
• crates/ironclaw_product/src/projection/tests/live_progress_stream.rs
• crates/ironclaw_webui/CLAUDE.md
• crates/ironclaw_webui/frontend/src/pages/chat/components/message-bubble.test.ts
• crates/ironclaw_webui/frontend/src/pages/chat/components/message-bubble.tsx
• crates/ironclaw_webui/frontend/src/pages/chat/hooks/useHistory.ts
• crates/ironclaw_webui/frontend/src/pages/chat/lib/stream-order-memory.ts
• crates/ironclaw_webui/frontend/src/pages/chat/lib/useChatEvents.test.ts
• crates/ironclaw_webui/frontend/src/pages/chat/lib/useChatEvents.ts
• crates/ironclaw_webui/frontend/src/pages/chat/lib/useHistory.test.ts

[coderabbitai]

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (3)

scripts/ci/ws12_workflow_contracts.py (2)

60-70: 🎯 Functional Correctness | 🟠 Major | 🏗️ Heavy lift

Detect folded if: ${{ false }} values.

The multiline branch only matches bare false; it misses valid YAML such as:

if: >-
  ${{ false }}

A lane can therefore be unconditionally disabled while this guard passes. Parse the workflow or extend the matcher, and add a regression case for folded expressions.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@scripts/ci/ws12_workflow_contracts.py` around lines 60 - 70, Update the
UNCONDITIONAL_SKIP matcher to recognize folded YAML if values containing the
expression ${{ false }}, including the shown >- form, or parse the workflow to
detect them reliably. Add a regression case covering a folded false expression
and preserve detection of the existing bare and block-scalar false forms.

---

74-87: 🎯 Functional Correctness | 🟠 Major | 🏗️ Heavy lift

Validate active workflow structure, not arbitrary source substrings.

marker not in text accepts markers left in comments or dead scalars, so a removed job can still satisfy WS12. It also makes valid folded commands format-sensitive. Tie each marker to parsed jobs, steps, commands, or triggers, with fixtures for comments and inactive steps.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@scripts/ci/ws12_workflow_contracts.py` around lines 74 - 87, Update
validate_workflow_texts to parse each workflow’s YAML structure and validate
REQUIRED_MARKERS against active jobs, steps, commands, and triggers rather than
using raw substring checks. Exclude comments, dead or inactive steps, and other
non-executable scalars, while normalizing folded command representations so
equivalent valid commands are accepted. Preserve missing-workflow and
unconditional-skip reporting, and add fixtures covering commented markers and
inactive steps.

.github/workflows/reborn-e2e.yml (1)

265-272: 🩺 Stability & Availability | 🟡 Minor | ⚡ Quick win

Avoid a secondary failure when coverage generation is skipped.

if: always() still runs this upload after the validation step fails, but product_surface_coverage is then skipped and its output/path may not exist. With if-no-files-found: error, the artifact step obscures the original failure with another error. Gate this step on steps.product_surface_coverage.outcome != 'skipped'.

Proposed fix

-        if: always()
+        if: always() && steps.product_surface_coverage.outcome != 'skipped'

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In @.github/workflows/reborn-e2e.yml around lines 265 - 272, Update the “Upload
product-surface coverage matrix” step’s condition to require both `always()` and
`steps.product_surface_coverage.outcome != 'skipped'`, preventing artifact
upload when coverage generation was skipped while preserving uploads after other
outcomes.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Outside diff comments:
In @.github/workflows/reborn-e2e.yml:
- Around line 265-272: Update the “Upload product-surface coverage matrix”
step’s condition to require both `always()` and
`steps.product_surface_coverage.outcome != 'skipped'`, preventing artifact
upload when coverage generation was skipped while preserving uploads after other
outcomes.

In `@scripts/ci/ws12_workflow_contracts.py`:
- Around line 60-70: Update the UNCONDITIONAL_SKIP matcher to recognize folded
YAML if values containing the expression ${{ false }}, including the shown >-
form, or parse the workflow to detect them reliably. Add a regression case
covering a folded false expression and preserve detection of the existing bare
and block-scalar false forms.
- Around line 74-87: Update validate_workflow_texts to parse each workflow’s
YAML structure and validate REQUIRED_MARKERS against active jobs, steps,
commands, and triggers rather than using raw substring checks. Exclude comments,
dead or inactive steps, and other non-executable scalars, while normalizing
folded command representations so equivalent valid commands are accepted.
Preserve missing-workflow and unconditional-skip reporting, and add fixtures
covering commented markers and inactive steps.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 7e72f906-470a-49b7-8a49-fb62d64e63b5

📥 Commits

Reviewing files that changed from the base of the PR and between 88d5eba and 3dc16f8.

📒 Files selected for processing (7)

• .github/workflows/reborn-e2e.yml
• crates/ironclaw_product/tests/reborn_services_contract.rs
• scripts/ci/reborn-crate-test-buckets.sh
• scripts/ci/test-reborn-crate-test-buckets.sh
• scripts/ci/ws12_workflow_contracts.py
• tests/e2e/scenarios/test_reborn_webui_v2_smoke.py
• tests/integration/webui_v2_product_api.rs

[coderabbitai]

Actionable comments posted: 1

Caution

Some comments are outside the diff and can’t be posted inline due to platform limitations.

⚠️ Outside diff range comments (1)

crates/ironclaw_product/src/projection/tests/live_progress_stream.rs (1)

404-462: 🩺 Stability & Availability | 🟡 Minor | ⚡ Quick win

Avoid wall-clock timing for delivery sequencing.

This test relies on short delays and timed next() reads to prove provider-cadence updates. Scheduler jitter can make it intermittently time out before the producer publishes. Use synchronization/barrier signals for sequencing, retaining only a generous bounded timeout as a failure watchdog.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@crates/ironclaw_product/src/projection/tests/live_progress_stream.rs` around
lines 404 - 462, Update the live projection sequencing test around the milestone
publishing loop and subscription reads to replace the 25ms sleeps and per-read
one-second timeouts with explicit synchronization/barrier signals that confirm
each provider-cadence update is delivered before publishing the next milestone.
Retain only a generous overall bounded timeout as the failure watchdog, while
preserving the assertion that text items precede the capability activity.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@crates/ironclaw_webui/src/webui_v2/handlers.rs`:
- Around line 1306-1311: Update sse_keep_alive_event to derive its payload from
the typed WebChatV2Event::KeepAlive serialization instead of duplicating a raw
JSON literal, ensuring the event name and wire payload remain aligned. Add or
update serialization, ordering, and transport tests for the keep-alive heartbeat
frame as required by the event contract.

---

Outside diff comments:
In `@crates/ironclaw_product/src/projection/tests/live_progress_stream.rs`:
- Around line 404-462: Update the live projection sequencing test around the
milestone publishing loop and subscription reads to replace the 25ms sleeps and
per-read one-second timeouts with explicit synchronization/barrier signals that
confirm each provider-cadence update is delivered before publishing the next
milestone. Retain only a generous overall bounded timeout as the failure
watchdog, while preserving the assertion that text items precede the capability
activity.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 39254ed5-8070-46bb-9533-500e33a95d2f

📥 Commits

Reviewing files that changed from the base of the PR and between 3dc16f8 and ca2a6f7.

📒 Files selected for processing (5)

• crates/ironclaw_host_api/src/product_surface.rs
• crates/ironclaw_product/src/projection/tests/live_progress_stream.rs
• crates/ironclaw_webui/src/webui_v2/handlers.rs
• scripts/ci/reborn_changed_coverage.py
• scripts/ci/test_reborn_changed_coverage.py

[coderabbitai]

🗄️ Data Integrity & Integration | 🟡 Minor | ⚡ Quick win

Keep the keep-alive payload tied to the event contract.

The event name is enum-derived, but the payload is now duplicated as a raw JSON literal. A future WebChatV2Event::KeepAlive wire-shape change could make the event name and payload disagree. Reuse typed serialization or add an explicit transport contract test for this exact frame.

As per path instructions, ephemeral heartbeat variants require serialization, ordering, and transport tests.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@crates/ironclaw_webui/src/webui_v2/handlers.rs` around lines 1306 - 1311,
Update sse_keep_alive_event to derive its payload from the typed
WebChatV2Event::KeepAlive serialization instead of duplicating a raw JSON
literal, ensuring the event name and wire payload remain aligned. Add or update
serialization, ordering, and transport tests for the keep-alive heartbeat frame
as required by the event contract.

Source: Path instructions