`prompt-hook` injection cap (16k chars) exceeds Claude Code's 10k inline hook-output limit, so the context is persisted to a file and the model sees a 2 KB preview

[bompus]

Summary

codegraph prompt-hook caps its injected <codegraph_context> block at MAX = 16000 characters (src/bin/codegraph.ts, "Cap the injection so a large-repo explore can't flood the prompt"). Claude Code persists any hook stdout over 10,000 characters to a file and shows the model a 2 KB preview instead. So on a repo where explore fills the cap, the hook delivers a file path and the first 2 KB, not the context.

Claude Code's limit is not documented; I measured it (below). The fix is one constant.

What the model sees

From a Claude Code transcript with the installed hook, on every prompt of a 6-cell run (hook stdout 14.9–15.8 KB each time):

<persisted-output>
Output too large (15.8KB). Full output saved to: .../tool-results/hook-<id>-stdout.txt

Preview (first 2KB):
<codegraph_context note="Structural context from CodeGraph for this prompt — treat returned source as already read; call codegraph_explore for more.">
**Exploration: ...**

Found 47 symbols across 1 file.

**Blast radius — what depends on these (update/verify before editing)**
...
</persisted-output>

The model then either Reads the persisted file (one extra tool call per prompt) or answers from the preview.

Measured threshold

Claude Code 2.1.261, headless claude -p, a stub UserPromptSubmit hook that prints N characters:

hook stdout (chars)	delivered
8,031	inline
8,991	inline
9,631	inline
9,871	inline
10,031	persisted, 2 KB preview
12,031	persisted
16,031	persisted

So the inline limit is 10,000 characters, and a payload at the hook's own 16k cap is always persisted.

Effect

Same 6 discovery tasks, Opus, headless Claude Code, repo rules identical, paired cells:

setup	tool calls / task	codegraph_explore calls / task	correct
MCP server only (no hook)	3.83	1.33	6/6
codegraph install (hook + MCP)	4.33	0.83	6/6

With the hook the tool was called less, not more: the persisted preview is treated as "context already delivered" and the follow-up explore the note asks for often does not happen.

Related: #1654 notes that a keyword hit "injects the full explore payload unconditionally". This is what that payload becomes on the Claude Code side once it exceeds 10 KB.

Proposed fix

MAX = 9000 (leaves headroom under 10,000 for the wrapper and the projectPath nudge lines), keeping the existing "…(truncated; call codegraph_explore for the rest)" notice. The blast-radius list already leads the payload, so the surviving 9 KB is the part the model uses. Optionally read the cap from an env var for other hosts. Happy to send the one-line PR if that direction is right.

Environment

• codegraph v1.6.0 (npm global install, Windows 11, bundled Node 24.16.0)
• Claude Code 2.1.261
• Repo: ~450 TS/JS files