Graceful Tool-Output Truncation: The PARTIAL Signal¶

When a tool's output would exceed the model's token budget, return a useful prefix, a structurally distinct truncation signal, and a continuation handle — instead of failing hard and forcing the agent to recover blind.

Graceful tool-output truncation is a contract: a tool that can produce variable-sized output returns the most useful state it can fit, plus an explicit signal the result is incomplete and a continuation handle. Claude Code v2.1.145 (2026-05-19) shipped this for the Read tool — a whole-file read past the token limit now returns a truncated first page with a PARTIAL view notice instead of a hard error (Claude Code changelog). The same shape generalises to log readers, search results, directory listings, and MCP responses.

The Contract¶

Three load-bearing parts. Drop any one and the pattern degrades into the failure it tried to fix.

Part	Why it matters
Useful prefix	The agent gets work done on the first turn instead of consuming a retry round-trip.
Structurally distinct marker	The agent recognises the result is incomplete instead of treating the prefix as the whole answer.
Continuation handle	The agent has a defined next call (offset, cursor, file handle) rather than guessing at retry parameters.

The marker is load-bearing. A trailing [PARTIAL] line is the easiest implementation and the most ignored — the model reads the prefix as the document and the trailing line as commentary. Anthropic's own Read tool surfaced this as a bug: the agent treated the preview as the complete file, rules in the un-read portion were silently absent from context, and no signal reached the operator (anthropics/claude-code#28783). The marker belongs in a structurally distinct slot — a leading banner, a separate JSON field, or schema-typed metadata.

Why It Works¶

Turn-cost argument: every tool call costs a model latency turn, so error-then-retry doubles the cost of any large-output call. A tool that returns useful state plus a signal saves the retry whenever the prefix is enough and only spends it when continuation is required. The same logic underlies RFC 9457 error responses and observation contracts: encoding recovery into the response shape lets the agent branch on structure instead of parsing prose. MCP has an open proposal for client-negotiated response size caps but no normalised _meta.truncated field today (modelcontextprotocol#2211). Anthropic recommends the same shape: "pagination, range selection, filtering, and/or truncation with sensible defaults" (Writing tools for agents).

Where the Lever Lives¶

This is a tool author layer that composes with three adjacent layers: harness compression (Terminal Tool Output Compression), harness rewriting (PostToolUse Output Replacement), and server-side durability (MCP Result Persistence Annotation). The four stack: a tool can ship a PARTIAL contract while a harness compresses the prefix and an MCP server marks the result durable.

Diagram¶

graph LR
    A[Tool call<br/>oversized output] --> B{Within budget?}
    B -->|Yes| C[Full output]
    B -->|No| D[Truncate to prefix]
    D --> E[Attach PARTIAL marker<br/>+ continuation handle]
    E --> F[Agent reads marker]
    F -->|Prefix sufficient| G[Proceed]
    F -->|Need rest| H[Call with continuation handle]
    C --> G

When This Backfires¶

Trailing-prose marker. A bottom-of-output [PARTIAL] is the most ignored signal — the model reads the prefix as a complete document and the suffix as commentary (anthropics/claude-code#28783). Put the marker in a structurally distinct slot.
Security-critical reads. Guardrail files (CLAUDE.md, .cursorrules, policy configs) under PARTIAL semantics drop un-read rules silently. Fail-closed beats return-prefix here.
No continuation handle. PARTIAL without a cursor or offset is a polite hard error — recovery collapses back to retry-with-smaller-range (modelcontextprotocol#2211).
Downstream summarisation. A PostToolUse hook that summarises a PARTIAL prefix loses both the missing content and the marker — see PostToolUse Output Replacement §When This Backfires.
Agents without a reaction policy. A graceful tool paired with an agent that ignores the marker is worse than a hard error. The system prompt must say what to do (paginate, summarise, ask user, proceed).

Example¶

The Read tool's pre- and post-v2.1.145 contracts on a file that exceeds the token budget:

Before — fail-hard:

Error: File too large to read (exceeds token limit). Use offset/limit to read a range.

The agent has consumed a turn, gained no content, and has to guess at offset values it cannot inform without a size.

After — graceful with marker (v2.1.145):

[PARTIAL view — file exceeded token budget. First N tokens shown. Use offset to continue.]

<first N tokens of the file>

The marker is a leading banner — read before the prefix — and names the continuation parameter. The agent can act on the prefix immediately, decide whether the remaining bytes matter, and only then issue a second call (Claude Code changelog).

Authoring Checklist¶

For any tool whose output can grow past a fixed limit:

Return the prefix, not an error, when the full result overflows
Place the marker before the payload, in a structurally distinct slot — not as trailing prose
Name the continuation parameter (offset, cursor, page_token, file handle)
Specify which bytes were dropped (head, tail, middle) so the agent can judge whether the prefix is enough
Document the contract in the tool description so the model plans for the marker
Test both truncated and full paths — regressions on the truncated path are silent until production

Key Takeaways¶

Graceful truncation is a tool contract, not a harness workaround — the tool author owns the prefix-marker-continuation shape.
The marker is load-bearing. Trailing prose is ignored; structurally distinct slots (leading banner, separate field) survive the model's reading order.
Pair the marker with a continuation handle. Without it, the contract collapses back to fail-hard with extra steps.
The pattern composes with harness-side compression, post-hoc rewriting, and MCP persistence annotations — it does not replace them.
For security-critical reads (policy files, guardrails), fail-closed beats partial: a silently truncated rules file is the same as a missing rules file.