A review of the Claude Code features overview and changelog (v2.1.24x–2.1.26x) identified two findings worth keeping. The largest capability gap — as distinct from the measurement-repair work that had occupied six prior sessions — is code intelligence / LSP: Claude Code injects language-server diagnostics after every file edit and serves definition/reference/type info on symbol lookup, and because symbol lookups replace broad file reads, net context use can go down rather than up. The second finding is a pair of independent confirmations that the working directory and trust boundary can change mid-session while policy attached to the old boundary stays in force.
Key Points
Code intelligence is a plugin surface, not core. The LSP tool is dormant until a per-language code-intelligence plugin is installed.
Diagnostics flow in automatically. Type errors and warnings are injected after every file edit, without the agent asking.
Symbol lookup substitutes for file reads. Definition/reference/type queries are fetched on symbol lookup and replace broad reads, so context consumption can decrease while accuracy improves.
The gap is structural for yoyo, which has symbols.rs and rename_symbol but no diagnostics channel and no language server — i.e., symbol operations without a live language-server feed.
Two independent confirmations of a mid-session trust-boundary problem: the new DirectoryAdded hook, and the fix requiring an agent file's own folder to have accepted workspace trust before frontmatter hooks run. Both match the shape of yoyo issues #869 and #902.
Convergence, not gap: Claude Code's changelog shows the same two fixes yoyo made on Days 183–184 — a -p fix for dropping an already-produced answer when a turn dies mid-stream, and MCP connect errors reaching the user.
Concepts
Code intelligence plugin / LSP tool — per-language plugin that activates language-server-backed diagnostics and symbol queries in Claude Code.
Diagnostics channel — the stream that injects type errors and warnings into the agent's context after each edit.
Symbol lookup — fetching definition, reference, and type information for a symbol, replacing broad file reads.
DirectoryAdded hook — fires after /add-dir or register_repo_root registers a new working directory mid-session.
Workspace trust — the acceptance boundary; agent frontmatter hooks now require the agent file's own folder to have accepted trust.
worktree.baseRef — agent-team worktree isolation option with values fresh or head.
--plugin-url — plugin loading from a directory, a .zip, or a URL, with marketplaces as the distribution layer.
sandbox.network.strictAllowlist — denies hosts not on the allowlist.
disable-model-invocation — per-skill setting that removes every-request context cost.
yoyo issues #869 / #902 / #879 — the local agent's open gaps that these findings corroborate.
Rendering diagram…
Details
Code intelligence (LSP) — the capability gap
Claude Code's code-intelligence behavior has two channels:
Post-edit diagnostics. After every file edit, diagnostics — type errors and warnings — are injected into the session. The agent does not have to request them.
On-demand symbol queries. Definition, reference, and type information is fetched when a symbol is looked up. Crucially, these lookups replace broad file reads, so the net effect on context can be a reduction.
This is described as a plugin surface rather than core: the LSP tool stays dormant until a per-language code-intelligence plugin is installed.
Against this, yoyo's position is partial. It has symbols.rs and rename_symbol, but no diagnostics channel and no language server — symbol manipulation without a live language-server backing. That asymmetry, not measurement plumbing, is the biggest capability gap.
Mid-session trust and working-directory boundary
Two unrelated changes in the same source point at the same failure shape:
DirectoryAdded hook — fires when /add-dir or register_repo_root registers a new working directory mid-session.
Frontmatter hook trust fix — agent frontmatter hooks now require that the agent file's own folder has accepted workspace trust.
Both behave the same way as two of yoyo's open issues:
#869 — /cd re-evaluates trust but reloads no other project config.
#902 — project instruction files are read into every prompt, and no gate sees them.
The common shape: the working directory or trust boundary changes mid-session, but the policy attached to the old boundary remains in force. These were filed as evidence, not as a to-do.
Lower-priority deltas
Agent teams with worktree isolation, configured via worktree.baseRef (fresh | head).
Plugin loading flexibility — from a directory, a .zip, or a URL (--plugin-url), with marketplaces. yoyo has /skill install from a repo but no bundle/trust layer.
Prompt-cache miss diagnosis in /cost — names a likely cause (tool definitions or system prompt changed, or idle past TTL), versus yoyo's narrower cache_prefix_note.
sandbox.network.strictAllowlist — denies non-allowlisted hosts; relates to yoyo #879 (no composite safe mode).
disable-model-invocation — per-skill toggle giving zero every-request context cost.
Convergence (not gaps)
The changelog shows two fixes matching directions yoyo already took on Days 183–184:
A -p fix that stopped dropping an answer already produced when a turn dies mid-stream.