build(chat): consume partial-markdown 0.5.0 — streaming open-line rendering

[blove]

What

Bumps @cacheplane/partial-markdown ^0.4.1 → ^0.5.0 to pick up B.2 — streaming open-line rendering. The chat already calls materialize(parser.root), so no code change is needed: the open (unterminated) line now renders mid-stream (parsed, status: 'streaming') instead of buffering invisibly until a newline. Smoother streaming for long newline-free paragraphs; completed constructs render as they close.

Verification

• Chat unit suite green — 898 passed (incl. streaming-markdown integration + NG0956 guards).
• nx build chat green; lockfile synced (no platform bindings dropped).
• The markdown-surfaces e2e is unaffected (it finish()es → committed tree, which B.2 doesn't change) — CI confirms.

Part B of the resilient-markdown work is now fully shipped: B.1 escapes, B.3 fuzz suite, the O(n²) perf fix, and B.2 streaming open-line — all in partial-markdown 0.5.0.

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
threadplane	Ready	Preview, Comment	Jun 25, 2026 5:11am

[github-actions]

Automated approval: this PR received an intelligent (AI) code review. See the review comments on this PR.

[github-actions]

Claude finished @blove's task in 1m 7s —— View job

---

PR Review

• Read CLAUDE.md and gather context
• Review PR diff
• Analyze changes
• Post findings

---

This PR looks good. Clean, minimal dependency bump — no issues found.

Summary

The bump from ^0.4.1 → ^0.5.0 is applied consistently across all three locations (libs/chat/package.json, root package.json, package-lock.json) with the correct 0.5.0 integrity hash.

No code changes are needed because streaming-markdown.component.ts:123 already calls materialize(this.parser.root), which is exactly the entry point that 0.5.0's open-line rendering hooks into. When the parser has an unterminated line in the buffer, materialize() now surfaces it with status: 'streaming' rather than buffering it, so long newline-free paragraphs will appear character-by-character instead of all at once.

Interaction with the sentinel-\n logic

The existing finish path (lines 108–109 in streaming-markdown.component.ts) pushes \n before calling finish() when isStreaming flips to false. With 0.5.0 this is still correct: the sentinel closes the open line and finish() commits the tree. Completed content renders as before; only mid-stream rendering changes.

Test coverage

The chunked integration test (streaming-markdown.integration.spec.ts:89-102) pushes content one character at a time with per-chunk change detection, which implicitly exercises the open-line path for every sample — each push before the first \n hits an unterminated line. No targeted open-line streaming test is added here, but the existing coverage is reasonable for a pass-through dependency bump.

No bugs, no security issues, no API regressions, no type-safety concerns.

[github-actions]

Claude encountered an error after 1m 12s —— View job

---

PR Review

• Read CLAUDE.md and gather context
• Review PR diff
• Analyze changes
• Post findings

[github-actions]

Claude finished @blove's task in 1m 28s —— View job

---

PR Review

• Read CLAUDE.md and gather context
• Review PR diff
• Analyze changes
• Post findings

---

This PR looks good. The dependency bump is clean and the bundled e2e fixes are legitimate.

Dependency bump

The ^0.4.1 → ^0.5.0 bump is applied consistently across all three required locations:

File	Change
libs/chat/package.json	^0.4.1 → ^0.5.0
package.json	^0.4.1 → ^0.5.0
package-lock.json	resolved 0.4.1 → 0.5.0, new integrity hash

The lockfile also reflects the @threadplane/* packages moving from 0.0.52 → 0.0.53 (from the prior release commit already on main), which is expected and correct.

Integration verified

streaming-markdown.component.ts:123 already calls materialize(this.parser.root) — the exact entry point that 0.5.0 hooks into to surface unterminated lines with status: 'streaming'. No code changes needed here.

The sentinel-\n finish path (lines 108–109) remains correct: it closes any open line before finish() commits the tree. 0.5.0 doesn't change committed-tree semantics, so existing behaviour is preserved.

Bundled e2e fixes

Two commits are bundled that the PR body doesn't call out, but both are legitimate:

• project.json — dependsOn: ["inject-env"]: Adds a task ordering constraint so inject-env always runs before serve. Without this, the dev server could start before env vars are injected, breaking e2e runs. Correct fix.
• global-setup.ts — 120s → 240s timeout: Cold CI runners re-optimize Vite deps after a lockfile change (which this PR introduces), which can exceed the old 120 s window. The in-code comment explains the reason clearly.

Minor observation

The PR body says "no code change is needed" — strictly true for the @threadplane/chat library itself, but the e2e infrastructure files do change. Not a bug, just a description gap.

No correctness bugs, no security issues, no type-safety concerns, no public API regressions.

[github-actions]

Claude finished @blove's task in 1m 31s —— View job

---

PR Review

• Read CLAUDE.md and gather context
• Review PR diff
• Analyze changes
• Post findings

---

This PR looks good. One minor documentation gap; no code issues.

Dependency version: ^0.5.0 → resolves to 0.5.1

The range in package.json / libs/chat/package.json is ^0.5.0, but the lockfile pins to 0.5.1 (not 0.5.0). The final commit (64b5ff8 build(chat): resolve partial-markdown to 0.5.1 (adds optimistic open-line)) makes this intentional — the PR body describes 0.5.0's streaming open-line feature but 0.5.1 also adds "optimistic open-line" on top. No issue, but the PR description is slightly behind the lockfile. Worth a one-line note in the body for future readers.

Integration (streaming-markdown.component.ts:123)

materialize(this.parser.root) at line 123 is the correct hook for 0.5.x's open-line path. Unfinished lines now surface with status: 'streaming' mid-push rather than buffering until the next \n. The sentinel-\n + finish() logic at lines 108–109 and 115–116 remains correct and unchanged: committed-tree semantics are unaffected by 0.5.x.

Bundled e2e fixes

Both are legitimate and explained by their commit messages:

• dependsOn: ["inject-env"] in project.json — correct Nx task-dependency placement on the serve target. Ensures env injection completes before the dev server starts; fixes a race that broke e2e runs.
• global-setup.ts 120s → 240s — the in-code comment (cold CI runner re-optimizes vite deps on first serve...) fully justifies the increase given the lockfile change this PR introduces.

Lockfile integrity

SHA-512 integrity for 0.5.1 is 88 base64 characters (86 + == padding), which is correct for a 64-byte digest. No anomalies.

No correctness bugs, no security issues, no type-safety concerns, no public API regressions.

[github-actions]

Claude finished @blove's task in 2m 5s —— View job

---

PR Review

• Read CLAUDE.md and gather context
• Review PR diff
• Analyze changes
• Post findings

---

This PR looks good. Five well-scoped commits; no correctness bugs, no security issues, no public API regressions.

Dependency bump

^0.4.1 → ^0.5.0 applied consistently across libs/chat/package.json and the root package.json. The lockfile resolves to 0.5.1 (within the range) — intentional per commit 64b5ff8 which documents the patch as "adds optimistic open-line" on top of 0.5.0's streaming open-line. The integrity hash (sha512-eLaVKdz...) is valid for a 64-byte SHA-512 digest.

streaming-markdown.component.ts:123 already calls materialize(this.parser.root) — the exact entry point 0.5.x hooks into to surface unterminated lines with status: 'streaming'. The sentinel-\n + finish() path at lines 108–109/115–116 remains correct: committed-tree semantics are unchanged in 0.5.x.

Bundled e2e fixes

Both are correct and well-motivated:

project.json — dependsOn: ["inject-env"] on serve
The serve:development build target consumes generated-keys.local.ts (gitignored), which inject-env generates. Without this ordering, nx serve (called from global-setup.ts) would exit 1 before the port ever opened. Adding it to serve.dependsOn is the right Nx idiom — it runs on every cold start and is documented to be idempotent.

global-setup.ts — 120s → 240s
Justified: this PR introduces a lockfile change, which triggers Vite dep re-optimisation on the first cold serve in CI, reliably exceeding 120 s. The in-code comment calls this out explicitly. Acceptable window increase.

Test fix (subagent-card.spec.ts)

The original test asserted chat-tool-call-card with text "research", but since #692 a delegation tool call renders as <chat-subagent-card> (full durable transcript) instead of a bare tool-call card. The fix:

• Switches the locator to chat-subagent-card
• Updates the comment to accurately describe the rendering path (durable subagent group, not active-only filtering)
• Renames researchCall → researchCard for clarity

The hasText: /research/i filter on chat-subagent-card is reliable — the card binds subagent().name verbatim, and the aimock harness names it "research".

No issues found

No bugs, no security concerns, no type-safety issues, no subscription leaks, no public @threadplane/* API regressions.