[tarekziade]

Two things, both about how much of this reaches the model.

The instruction was being truncated

#114 pushed the output_mismatch instruction to 8,821 chars. serge's build_task_user_prompt head-truncates the instruction at MAX_INSTRUCTION_CHARS (8,000 at the time), so the 821 chars that fell off the end were the two bullets that PR existed to add — "the ACTUAL side of the assertion is off limits" and "a SHAPE or LENGTH mismatch is never expectation drift". Written, reviewed, merged, never delivered.

Every category's build_instruction output is now checked against that cap, with a failure message saying what truncation would cost. serge#118 raises the cap to 12,000 for headroom; this is the half that belongs here, where the blocks are written and where the next one will grow.

Compression

Only these strings are sent to the model — the # comments around them are not — so the rationale and dated precedents stay put and only the delivered text shrinks. 21,141 → 18,886 chars. Same rules, same thresholds, same PR numbers and examples: the 112 phrases the test module pins are all still there, and 328 passed.

block	before	after
_MISMATCH_GUIDANCE	6,178	5,031
_OOM_LOAD_GUIDANCE	4,624	4,333
_INSTRUCTION	2,641	2,192
_CULPRIT_GUIDANCE	2,110	2,041
_CRASH_GUIDANCE	2,058	1,969
_OOM_RETENTION_GUIDANCE	1,881	1,734
four small blocks	1,649	1,586

The largest single win is not prose. The trunk told the agent to go and read the root AGENTS.md / CLAUDE.md / .ai/AGENTS.md — which serge already injects verbatim into the system prompt as REPO CONVENTIONS. That paragraph paid for the conventions twice and bought three tool calls; it now says they are already in the prompt and points only at what the agent does need to open (the failing model's test file and its docs page).

Delivered instruction, worst case: output_mismatch 8,821 → 7,225 chars, inside even the old 8,000 cap.

Where the first load actually goes

For the worst-case task, with the compressed .ai/AGENTS.md:

system  9,876 + user 47,720 = 57,596 chars (~14.4k tokens)
  failure report (context)   40,000 chars — 69%
  conventions + instruction  10,363 chars — 18%

So the prompts are now the small half. If first-load size is still a concern, the lever is MAX_CONTEXT_CHARS (40,000) and how much traceback the triage packs into the report — not these blocks.