06 — Make `--estimate` honest about the backend

Priority: P2
Depends on: none
Size: S (30–45 min)
Files: sessionwiki/cli.py, tests/test_synthesize_estimate.py
PR title: fix: label synthesize estimate with the active backend

Read first

sessionwiki/cli.py, _synthesize_estimate (around line 558) and cmd_synthesize (around line 472)
sessionwiki/synth/estimate.py, synthesize_estimate_report
sessionwiki/cache.py, MODEL_PRICING and estimate_cost
tests/test_synthesize_estimate.py

Problem

_synthesize_estimate() calls:

report = synthesize_estimate_report()

estimate_cost() rejects unknown models with ValueError. The report therefore always uses DEFAULT_MODEL (currently a Claude model) and prints:

Model: claude-sonnet-4-6
Incremental sync: $55.4754

when the configured backend is openai with model: qwen3.8, which runs on a local server. The dollar figure is meaningless and misleading.

Do not pass qwen3.8 directly to estimate_cost; it will raise ValueError: unknown model.

Required change

Pass the loaded config into _synthesize_estimate. In cmd_synthesize the config is already loaded before the estimate branch:

if args.estimate:
    return _synthesize_estimate(config)

Change the signature:

def _synthesize_estimate(config: dict | None = None) -> int:

At the start of the function, inspect the backend:

from sessionwiki.cache import MODEL_PRICING

synth = (config or {}).get("synthesis", {}) or {}
backend = str(synth.get("backend") or "dummy")
model = str(synth.get("model") or "")
priced = model in MODEL_PRICING

Print one note when the active backend is not priced by MODEL_PRICING:

if backend not in {"anthropic", "claude"} or (model and not priced):
    print(
        f"note: cost estimate uses Claude API pricing; "
        f"backend {backend!r} may bill differently "
        f"(local models are usually free)."
    )

Keep the existing report and numbers unchanged. The note is enough for this card. Do not build a new pricing table.

If model is in MODEL_PRICING, pass it through:

report = synthesize_estimate_report(model=model if priced else None)

Tests

Add to tests/test_synthesize_estimate.py:

One test that captures output for config = {"synthesis": {"backend": "openai", "model": "qwen3.8"}} and asserts the note is printed.
One test for config = {"synthesis": {"backend": "dummy"}} that asserts the note is printed.
One test for a Claude model that is in MODEL_PRICING and asserts the note is not printed. Use a model key from MODEL_PRICING rather than a hardcoded name.

Use capsys and call _synthesize_estimate(config) directly. Stub the report if the test becomes slow:

monkeypatch.setattr(
    "sessionwiki.cli.synthesize_estimate_report",
    lambda **kwargs: {...},
)

If the function is imported inside _synthesize_estimate, patch the module attribute it actually looks up.

Verification

.venv/bin/python -m pytest tests/test_synthesize_estimate.py -q
.venv/bin/python -m sessionwiki synthesize --estimate

Expected in this repository:

note: cost estimate uses Claude API pricing; backend 'openai' may bill differently ...

The existing Corpus, Already synthesized, and cost lines remain.

Acceptance criteria

[ ] _synthesize_estimate receives the config.
[ ] A non-Claude backend prints the note.
[ ] A priced Claude model prints no note.
[ ] Existing estimate tests still pass.
[ ] CHANGELOG.md updated.

Out of scope

Adding pricing for local or OpenAI-compatible models.
Changing synthesize_estimate_report.
Changing the default model.

Rollback

Revert the signature, the note, and the tests.

06 — Make --estimate honest about the backend