| TE-10 · Content-route before consuming — classify a large input (code, log, JSON, test output, diff, markdown, prose, transcript, generated/vendored output, binary/asset metadata) before deciding how much of it to read | Routing decides depth, never whether to verify; an authoritative artifact still gets read in full when the task needs its invariants |
| TE-11 · Extract before reasoning — for large logs, test output, JSON, traces, transcripts, and search results, pull the relevant slice or a summary first instead of loading the whole body into reasoning context | The slice is for triage; if it is ambiguous or the decision is load-bearing, widen it or open the source |
| TE-12 · Preserve the retrieval path — every summary or compressed note carries a handle to reopen the original: source path, command, report path, line/window pointer, request ID, or test name | A note with no retrieval handle is a dead end — you must always be able to get back to ground truth |
| TE-13 · Original-before-edit — before editing code, changing product logic, approving acceptance criteria, deciding security / billing / permissions / client-data, or closing verification, reopen and read the authoritative source | A summary may point you AT the change; it never substitutes for reading the file you are about to edit or approve |
| TE-14 · Summary is not authority — compressed/summarized context is for exploration and triage | NEVER the sole basis for implementation, security, permissions, billing, migrations, compliance, public API contracts, client-data access, or acceptance criteria — those decide against the original |
| TE-15 · Stable context floor — don't rewrite always-on rules, agent files, command files, handoffs, or governance docs just to improve phrasing | A small, stable instruction surface is cheaper every session than a "better-worded" one; governance edits are explicit, reviewed, diff-based tickets — never an optimization side effect |
| TE-16 · Cross-agent dedupe — handoffs between PO, AI Developer, Architect, Deploy, QA, and other roles carry decisions, IDs, paths, and pointers, not duplicated full artifacts | Dedupe never drops a decision or an ID the next role needs; every pointer must resolve |
| TE-17 · Large-output hygiene (anticipate, then narrow) — when a command's output is plausibly large or unbounded (full logs, whole-tree listings, unfiltered queries/dumps), scope the FIRST invocation — limit/offset, tail, grep/jq, targeted test, shorter traceback, or write-to-report-then-read — rather than running it wide and narrowing only on the rerun | Scoping serves the question; never scope so tight that needed evidence is excluded, and an estimated size is never a hard cap or a reason to skip a needed read |
| TE-18 · Compression is not verification — never use compression/summarization to skip reproduction, skip smoke tests, thin acceptance evidence, or hide uncertainty | Quality outranks tokens — on any conflict the correctness/safety floor wins, exactly as in the Guardrail |
