`run_in_terminal` tool description advertises 60KB output truncation but the live code path truncates at 16KB

[anthonykim1]

The run_in_terminal tool description tells the model it gets up to 60KB of output, but the live output-retrieval path caps at 16KB.

Tool description strings at runInTerminalTool.ts#L135 and #L219 both say:

Output is automatically truncated if longer than 60KB to prevent context overflow

There are two MAX_OUTPUT_LENGTH constants with different values:

• runInTerminalHelpers.ts#L51 → 60000, only used by sanitizeTerminalOutput, which has no non-test callers.
• outputHelpers.ts#L10 → 16000, used by getOutput, which is the only output path reached by runInTerminalTool, getTerminalOutputTool, and taskHelpers.

So every run_in_terminal invocation that produces more than ~16KB of output returns only the last 16KB, regardless of what the tool description claims.

User-visible implications

truncateOutputKeepingTail keeps the tail, so the head is what gets dropped. That's where errors usually live for npm install, pip install, cargo build, pytest -v, git log, etc. The agent reads a truncated tail, confidently concludes "nothing actionable in the output", and the user has to re-run or scroll up to find the real failure.

The model also plans against the wrong budget: told 60KB, it doesn't paginate or filter commands that would exceed 16KB, so the problem compounds on anything verbose.

Shape of the fix

1. Pick one number. Either bump outputHelpers.ts to 60000 (or make it a setting like chat.tools.terminal.maxOutputBytes), or change the description strings to 16KB. The description and the truncation limit must agree.
2. Delete sanitizeTerminalOutput — it's dead code, and having a second MAX_OUTPUT_LENGTH constant in the same module is how this drift happened.

[anthonykim1]

Revised after re-reading the linked issues (previous version labelled a few "Strong" that are really same-symptom-family, not same mechanism):

Relevance	Issue	Title	State	Why related
Strong	#299486	run_in_terminal timeout silently truncates output with no signal to agent	Closed	Same class of bug: tool returns partial output with no flag, agent can't tell. That one is about the explicit timeout param; this one is about the 60KB-vs-16KB mismatch on the normal (non-timeout) path
Strong	#296887	Terminal tool returns partial output for commands >~11s on terminals without shell integration	Closed	Same class: silent truncation with no indication to the agent. NoneExecuteStrategy's 10s extendedTimeoutMs returns early; this issue is about getOutput capping at 16KB
Medium	#297109	run_in_terminal tool non-deterministically truncates command output	Closed	Different mechanism (render-race between command-end marker and renderer flush) but same user-visible result: agent gets less output than the user expects
Medium	#282904	Copilot Chat / Agent misses terminal output (last lines) on Windows (Git Bash + WSL)	Closed	Different mechanism (tail-dropping on Windows) but again same class — if the 16KB cap also drops content from the tail in some cases, these compound
Weak	#299588	Copilot Chat repeatedly fails to read terminal output, causing retries and excessive token usage	Closed	End-user symptom (model retries because output looks empty) — could be exacerbated by the 16KB cap in some large-output cases, but no direct evidence
Weak	#299205	Blind to terminal output	Open	Vague "command has output but model says none" report. Possibly amplified by the 16KB cap but not specifically about it

The specific 60KB-in-tool-description vs 16KB-in-outputHelpers.ts mismatch is not covered by any of these.