Respect explicit full-source intake requests for large planning/spec corpora

What feature would you like to see?

A reliable, explicit "full-source intake" mode for Codex where, when the user asks it to read large source files or planning/spec corpora completely, Codex preserves source fidelity instead of substituting lossy deterministic extraction, summarization, or script-generated digests.

Today, Codex often appears willing to read moderately sized files directly, but for larger files or corpora it tends to avoid actually ingesting the material in model context. Even when instructed to manually read/intake the files, it frequently writes scripts or pipelines to extract headings, snippets, summaries, counts, or other deterministic projections, then treats that compressed representation as equivalent to having read the source.

That may be efficient for data analysis, but it is not equivalent for tasks where the source material is the specification. Examples include compiling a wiki, reconciling architecture plans, preserving nuanced design constraints, or executing long-horizon implementation plans where fidelity to the original planning text matters.

Desired behavior:

• Provide a user-controllable mode or instruction that Codex treats as binding: read the full specified source material, not just an extracted summary.
• Clearly distinguish full source intake from deterministic extraction or summarization.
• If full intake is impossible because of context limits, file size, cost, or runtime constraints, Codex should say so explicitly and offer a bounded fidelity-preserving intake strategy rather than pretending the source has been read.
• Preserve traceability from downstream plans/docs/wiki output back to the original source files and sections.
• Avoid collapsing detailed user-provided plans into a lossy "agent plan" when the user explicitly requires source-fidelity planning/spec compilation.

Why this matters

This is a major pain point for long-horizon agentic work. Some users write extensive architectural, product, system-design, and implementation plans up front. These can easily exceed 10k lines across one or more files. That size is intentional: the files contain constraints, obligations, ordering, rationale, edge cases, and scope boundaries that need to remain authoritative.

When Codex replaces those sources with a deterministic extraction pipeline or compressed summary, it can introduce compounding drift:

• original obligations get omitted or weakened;
• adjacent constraints are merged incorrectly;
• planning context becomes lossy;
• later implementation follows the agent's reduced plan rather than the user's source material;
• wiki/spec output loses nuance that was present in the source.

This feels like a known limitation of agentic workflows: long-horizon work can collapse as the agent repeatedly compresses source material instead of maintaining source fidelity.

Repro shape

1. Create one or more large planning/spec/source files, for example a 10k+ line project plan with detailed architecture, product requirements, unit sequencing, constraints, and closure requirements.
2. Ask Codex to read/intake the full corpus and compile a wiki/spec/implementation plan from it.
3. Explicitly instruct Codex not to summarize first and not to rely on deterministic extraction as a substitute for reading.
4. Observe that Codex often chooses to write scripts that extract headings/snippets/counts/summaries, then proceeds from that compressed representation.
5. Downstream output misses or flattens obligations that were present in the source material.

Expected behavior

When the user explicitly requests full-source intake, Codex should either:

1. actually intake the requested material with source-fidelity guarantees and traceable coverage, or
2. stop and disclose the exact limitation preventing full intake, then propose a strategy that preserves fidelity and makes the unresolved frontier explicit.

It should not silently substitute a lossy extraction pipeline and proceed as though the full source was read.

Additional context

Related community discussion: https://www.reddit.com/r/codex/comments/1ugyvez/half_of_your_highstakes_codex_requests_may_be/

This is not primarily about wanting Codex to spend context wastefully. It is about having a dependable contract for cases where the source material itself is the authority, and lossy compression changes the task.