IronClaw
IronClaw proposes reading documents on demand instead of inlining every attachment
IronClaw’s document-pointer proposal would replace automatically inlined document text with a stored path and an explicit read instruction. A companion context-limit override addresses larger models; both remain opt-in design work, with no measured end-to-end savings established by the inspected evidence.
IronClaw has an open September 8 proposal to change what attaching a document means for the model’s initial context. Instead of copying extracted text into every request, an opt-in pointer mode would present the stored project path and let the agent read the document when needed. The source describes a live same-question comparison in which inline mode answered directly while pointer mode called read_file before answering.
A companion proposal makes the prompt-context ceiling configurable rather than fixed at 128,000 tokens. These are separate controls: one changes when document content is introduced, while the other changes the budget against which compaction is planned. Neither demonstrates that a larger context window or an extra file read automatically produces a cheaper, better answer.
The facts
- PR #8082 proposes IRONCLAW_ATTACHMENT_DOCUMENT_TEXT with inline as the default and pointer as the opt-in value. - Pointer mode renders a stored project path and read_file instruction; an unstored document is labeled as unstored rather than given a fictitious path. - Unset and unrecognized values keep inline behavior, so a typo does not silently suppress document text. - Audio transcripts remain inlined in both modes, and existing credential-redaction behavior remains in place for inlined text. - The author reports correct answers in both modes for a same-attachment comparison and cites 18 attachment-context tests. - Companion PR #8087 proposes a context-limit override included in the family fingerprint and compaction strategy. - The pointer switch is global; per-model and per-file-type selection remain outside the proposal.
Why it matters
Automatic attachment inclusion makes a chat convenient, but persistent transcripts can turn one large document into repeated context pressure. Pointer mode offers a clearer retrieval decision at the cost of an additional tool path and the possibility that an agent reads too much anyway. A context-limit override solves a different problem: operators should not have to carry a source edit to describe their chosen model’s budget. Including that override in runtime identity matters because replay assumptions change with the configuration.
Current
Inspected on 2026-09-10. The IronClaw stable-release baseline is ironclaw-v1.4.0 published 2026-08-28T11:19:24Z. The main source was open (not merged) when captured. Release metadata is a version boundary; an open proposal, a merged mainline change and a packaged release remain different states.
Evidence
The primary evidence is nearai/ironclaw PR #8082 (https://github.com/nearai/ironclaw/pull/8082). Supporting context comes from nearai/ironclaw PR #8087 (https://github.com/nearai/ironclaw/pull/8087); IronClaw — current official README (https://raw.githubusercontent.com/nearai/ironclaw/main/README.md). These pages were inspected directly. Related project records explain the source cluster; they are not independent replications of contributor test claims.
Operator take
Compare complete task traces rather than initial prompt sizes. Measure the file-read output, later turns, accuracy and total model usage with the same document and question. Confirm missing-file behavior and keep audio behavior in the test matrix. The PR’s phrase that the document never enters the prompt should be read narrowly: it is not automatically inlined as an attachment, but retrieved content still reaches the model through tool results.
Caveat
Both changes remain open pull requests against a stable ironclaw-v1.4.0 baseline. The example PDF size and comparison are author claims, not representative benchmarks. No latency, token-saving percentage or production-quality improvement has been independently established; pointer mode is not a guarantee that document contents never reach inference.
Both changes remain open pull requests against a stable ironclaw-v1.4.0 baseline. The example PDF size and comparison are author claims, not representative benchmarks. No latency, token-saving percentage or production-quality improvement has been independently established; pointer mode is not a guarantee that document contents never reach inference.