projects: the panel ignores every filter and cannot show cost

[pitimon]

The Projects panel is where the README's central argument lands, and it is the least finished surface in the product.

Verified

The client sends filters. The server reads none of them.

dashboard/src/lib/api.ts:90-109 builds from, to, source, limit, timeZone into the request. The handler at src/lib/local-api.js:1232-1305 reads project.queue.jsonl and aggregates it unconditionally — awk over the whole handler finds zero references to searchParams or url..

Concrete failure: pick "24h". Every other card narrows. Projects keeps showing all-time totals, with nothing on screen saying so. A user comparing "this week's spend by repo" is reading lifetime numbers.

There is no cost, and none can be computed. The handler accumulates total_tokens and billable_total_tokens only. Project rows are keyed project|source|hour (src/lib/rollout.js:1596-1611) and carry no model — and cost is per-model. So per-repo cost is not merely unimplemented, it is absent from the data model.

README:33 says the product answers "which repo, which model, and which hour". Hour resolution is stored on every project row and thrown away at the API. The repo×model join has never existed.

Coverage is partial. projectBucketsQueued appears in 7 parsers (codex, claude, gemini, opencode ×2, openclaw, antigravity). Cursor, Copilot, Zed, Goose and Kiro usage is invisible to the per-repo story entirely — and the panel does not say which sources it can account for.

Proposal

1. Honour the filters. from/to against the stored hour_start in the request timezone, plus source and limit. Pure surfacing — the data is already on disk.
2. Add model to newly written project rows, moving the key to project|source|model|hour. Per-repo cost then computes through the existing computeRowCost.
3. Rows written before the change price as unattributed — the tier added in fix(pricing): stop reporting the "unknown" placeholder as an unpriced model #94 handles exactly this and renders honestly rather than as a silent $0.
4. Say what is not counted. If a source has no project attribution, the panel should name it rather than let its absence read as "that tool cost nothing here".

The risk to get right

Mixed-era rows. Legacy project|source|hour and new project|source|model|hour describe the same bucket at different granularity; summing both double-counts. Needs one explicit keep-last rule and a test over a fixture containing both eras — the same shape as the migration in sync.js:1836.

Out of scope

Backfilling model onto historical project rows. The source logs may no longer exist, and unattributed is the honest rendering of "we recorded this before we recorded models".

[pitimon]

The surfacing half is done in #118 (main @ f5331c2). Leaving this issue open for the cost half.

Closed by #118

• Filters are honoured. from/to against the stored hour_start in the request timezone, plus source and limit. The handler previously read no query parameter at all.
• The per-source fallback is filtered too — the path a user sees before project attribution has ever synced ignored the window as well.
• Sources with no per-repo attribution are named. Cursor, Copilot, Zed, Goose and Kiro have no project story; omitting them looked identical to "that tool cost nothing here". The list follows the same window as the entries.

One thing worth recording, because it is not in the issue text: the other range-filtered handlers compare against "" directly, which is safe there because the client always sends both bounds. The project endpoint's client does not, and day <= "" is false for every day — a naive copy of the existing pattern would have blanked the panel whenever only one bound was set.

Still open — the cost half

Points 2 and 3 of the proposal, unchanged:

• Add model to newly written project rows, moving the key from project|source|hour to project|source|model|hour, so per-repo cost computes through computeRowCost.
• Rows written before the change price as unattributed (the tier from fix(pricing): stop reporting the "unknown" placeholder as an unpriced model #94).

The risk the issue flags is the whole job. Legacy project|source|hour and new project|source|model|hour describe the same bucket at different granularity, and summing both double-counts. That needs one explicit keep-last rule and a test over a fixture containing both eras — the same shape as the migration at sync.js:1836.

Two things that landed since this issue was written make that easier:

• queueRowKey / planCompaction in src/lib/queue-compact.js (feat(queue): reclaim the 84% of the queue nothing ever reclaimed #117) established the pattern for "the readers define the key, and compaction only has to produce what they already compute". The project queue has its own dedup at readProjectQueueData, and any key change has to move both together.
• buildProjectUsageSummary is exported from src/lib/local-api.js, so the mixed-era aggregation can be tested against a two-era fixture without standing up a server.

Backfilling model onto historical rows stays out of scope, per the original text.

[pitimon]

Cost half shipped in #121. Both halves of this issue are now closed.

The transition rule, which was the risk this issue flagged: the bucket state in cursors.json is re-keyed on load, so a pre-model bucket becomes the unknown slot for its hour and keeps its running total and its queuedKey. It stops receiving new usage — that goes to the model-specific bucket — so it is frozen at its pre-upgrade total and the two sum to the true number. The reader keys model-less rows to that same slot, so a legacy row and its re-keyed bucket are one identity rather than two.

Cost is honest about what it cannot price: a model-less row prices at 0 and is counted as unattributed_tokens rather than folded in. A repo that is entirely pre-model shows "no model recorded" instead of "$0.00"; a partly-priced one shows "$3.00+".

Backfilling model onto historical rows stays out of scope, per the original text.