context_indicator plugin forks agents/_history token estimation to dodge a monkeypatch — same MCP/tool-schema walkers maintained twice

Theme

plugins/context_indicator/usage.py deliberately re-implements the token-estimation walkers from code_puppy/agents/_history.py so that /context stays immune to the token_ratio_learner plugin monkeypatching _history.estimate_tokens. The motive is documented and sound (usage.py:6-28), but the mechanism chosen is a fork:

• usage.py:163-189 _raw_tokens_for_mcp_servers ↔ _history.py:163-201 _estimate_mcp_tool_tokens — byte-identical traversal of _cached_tools / tool_prefix / name / description / inputSchema (jscpd: usage.py:163-176 ↔ _history.py:176-189). Only the leaf call differs (_raw_estimate_tokens vs estimate_tokens).
• usage.py:127-159 _raw_tokens_for_pydantic_tools ↔ the corresponding loop in _history.estimate_context_overhead — the docstring even says "Mirrors _history.estimate_context_overhead". It already imports _extract_tool_description / _extract_tool_json_schema / stringify_part from _history, so the fork is partial — the walkers are cloned, the leaf helpers aren't.
• usage.py:103-112 _raw_estimate_tokens clones the len(text)/2.5 heuristic with a comment admitting it "mirrors the original _history.estimate_tokens before any plugin patches it".

Why it matters

When MCP tool serialization changes (e.g. pydantic-ai renames _cached_tools, or tools gain output schemas), _history.py gets fixed and /context silently drifts — there is no test tying the two walkers together. The traversal logic and the estimator function are two separate concerns that have been entangled: the fork duplicates the traversal just to swap the estimator.

Proposed shared refactor (dependency injection instead of fork)

Parametrize the canonical walkers in _history.py with an estimator callable:

# agents/_history.py
def estimate_mcp_tool_tokens(mcp_servers, *, estimator: Callable[[str], int] = None) -> int:
    est = estimator or estimate_tokens   # patched name resolves at call time
    ...one traversal, leaf calls go through est()...

def raw_estimate_tokens(text: str) -> int:
    """Unpatched char/2.5 heuristic. token_ratio_learner must NOT patch this."""

Then context_indicator/usage.py becomes:

from code_puppy.agents._history import estimate_mcp_tool_tokens, raw_estimate_tokens
total = estimate_mcp_tool_tokens(servers, estimator=raw_estimate_tokens)

One traversal implementation, two estimators, zero drift. The "raw" estimator living in _history.py under an explicit do-not-patch name also turns the implicit contract with token_ratio_learner (which today is "hope the plugin only patches estimate_tokens") into an explicit one.

Filed by Zen Reviewer C (code-puppy-60635a) — DRY review round

Start by reading plugins/context_indicator/usage.py:103-189 and the corresponding walkers in code_puppy/agents/_history.py, including the usage.py:6-28 monkeypatch rationale. Trace how token_ratio_learner patches estimate_tokens, then verify the refactor leaves one traversal implementation with an explicitly unpatched estimator and no behavioral drift in /context.