phoenix-context

phoenix-context — make every token earn its place

Overview

A model has no memory between turns, so the harness re-sends the whole transcript every turn — every file you read is re-paid for on every subsequent turn. The cost that matters is not tokens sent but tokens processed, summed across all turns until the task is done. phoenix-context collapses that: structural questions go to a prebuilt code graph (the bundled TokenMasterX, graphify-backed), answered in one bounded query, instead of grep-and-read loops that compound turn after turn.

Pay once to understand structure, then never again.

When to use

• Any non-trivial / unfamiliar codebase.
• Before a change whose blast radius is unclear ("what breaks if I change this signature?").
• During phoenix-plan (scope the real impact) and phoenix-debug (isolate the cause).
• Whenever you catch yourself opening a 3rd+ file to answer a structural question.

When NOT to use: a short, local question grep answers in ~one turn, or you already have the snippet.

Two ways to assemble the same answer

   GREP its way there                    GRAPH lookup
turn 1  grep            2,000         turn 1  query graph   2,500
turn 2  read auth.py    6,000         turn 2  answer        3,000
turn 3  read server    11,000         ─────────────────────────
turn 4  read handlers  17,000         TOTAL PROCESSED       5,500
turn 5  grep narrower  19,000
turn 6  read config    24,000         Same answer.
turn 7  answer         26,000         19x less context processed.
─────────────────────────
TOTAL PROCESSED       105,000   ← it is the AREA UNDER THE CURVE, not the last turn

Route structural questions to the graph

Question	Graph query
"Who calls X?" / reverse dependencies	callers (or impact for the transitive blast radius)
"What does X call?"	callees
"What inherits / overrides X?"	inheritors
"Orient me on this unfamiliar symbol"	explain (or find to disambiguate a definition first)
"What breaks if I change X?"	impact