StackUnderflow ingests session data from more than one coding agent. Twenty adapters are registered and every one is on by default. The registry in stackunderflow/adapters/__init__.py is self-discovering: it walks the package at import time and registers every adapter module, so there is no opt-in flag and no import list to maintain — an adapter whose source directory is absent on a machine simply yields nothing. Curated per-adapter fidelity (the Status column below) lives in stackunderflow/adapters/capabilities.json; a beta status means pending broad validation, not opt-in.
| Provider | Status | Source format | Default state |
|---|---|---|---|
| Claude Code | stable | per-project JSONL under ~/.claude/projects/<slug>/ (+ legacy ~/.claude/history.jsonl) |
on |
| Codex | stable | rollout JSONL under ~/.codex/sessions/ |
on |
| Cursor | stable | SQLite state.vscdb at ~/Library/Application Support/Cursor/User/globalStorage/ |
on |
| Cline | stable | per-task JSON in ~/Library/Application Support/Code/User/globalStorage/saoudrizwan.claude-dev/tasks/ |
on |
| OpenClaw | stable | first existing of ~/.openclaw/, ~/.clawdbot/, ~/.moltbot/, ~/.moldbot/ (agents/) |
on |
| Pi + OMP | stable | ~/.pi/agent/sessions/ and ~/.omp/agent/sessions/ (one adapter, both roots) |
on |
| Hermes | stable | recursive JSONL under ~/.hermes/sessions/ |
on |
| KiloCode | beta | per-task JSON in …/kilocode.kilo-code/tasks/ (Cline parser) |
on |
| Roo Code | beta | per-task JSON in …/rooveterinaryinc.roo-cline/tasks/ (Cline parser) |
on |
| OpenCode | beta | SQLite under $XDG_DATA_HOME/opencode/ (or ~/.local/share/opencode/) |
on |
| Cursor Agent | beta | text/JSONL transcripts in ~/.cursor/projects/{p}/agent-transcripts/ (+ ~/.cursor/ai-tracking/ai-code-tracking.db) |
on |
| Qwen | beta | JSONL in $QWEN_DATA_DIR/projects/{p}/chats/*.jsonl (default ~/.qwen/) |
on |
| Gemini | beta | JSON / JSONL in ~/.gemini/tmp/{p}/chats/session-*.{json,jsonl} |
on |
| Copilot | beta | ~/.copilot/session-state/{sid}/events.jsonl + VS Code workspaceStorage/{h}/GitHub.copilot-chat/transcripts/ |
on |
| Codeium | beta | ~/.codeium/ (discovery stub — protobuf decoding deferred; yields nothing today) |
on |
| Continue | beta | ~/.continue/*.{db,sqlite,sqlite3} (defensive SQLite parser) |
on |
| Droid | beta | $FACTORY_DIR (or ~/.factory/sessions/{projectHash}/) |
on |
| Kiro | beta | ~/Library/Application Support/Kiro/User/globalStorage/kiro.kiroagent/*.chat |
on |
| Antigravity | beta | ~/.gemini/antigravity/ + antigravity-ide/ summary .pb files and antigravity-cli/history.jsonl (plaintext metadata only — message text and tokens are encrypted) |
on |
| Grok | beta | JSONL in ~/.grok/sessions/<url-encoded-cwd>/<session>/chat_history.jsonl (no token usage in the source; tokens estimated from content length) |
on |
Cursor and Cline have the deepest coverage of the non-Claude adapters: both carry test coverage, fingerprint-based caching for the Cursor vscdb (stackunderflow/infra/cursor_cache.py), and have run against real local user data across several releases.
Every adapter is registered automatically — there is nothing to enable. stackunderflow/adapters/__init__.py walks the package at import time and registers every module that satisfies the SourceAdapter shape (a non-empty name plus callable enumerate / read), so adding a new agent means adding one module file: no opt-in flag, no import list to extend, no way to ship an adapter that silently never registers. An adapter whose source directory is absent on a given machine simply yields nothing from enumerate(), so shipping the full set is safe and cheap everywhere; a module that fails to import raises loudly rather than disappearing.
Curated per-adapter fidelity lives in stackunderflow/adapters/capabilities.json (loaded and validated by stackunderflow/services/support_matrix.py). Each entry records a status (supported / beta / partial), whether the source can emit billable usage events, and per-field fidelity (content_text, tokens, cost, tool_calls, tool_output, reasoning, file_touches). A beta status means pending broad validation — the adapter is on like every other; the label only flags that its output hasn't been checked against as much real-world data yet.
To see what a given machine actually ingests per provider, run stackunderflow doctor (alias stax doctor). Its delivery scoreboard walks disk sessions → base messages → usage_events → marts and flags any provider that has sessions on disk but nothing loaded; stackunderflow doctor --fail-on-gap turns such a gap into a non-zero exit for CI.
Cursor. Reads the cursorDiskKV table in ~/Library/Application Support/Cursor/User/globalStorage/state.vscdb (opened via SQLite read-only URI). Two key prefixes are walked: bubbleId:% for chat bubbles (with text, modelInfo.modelName, tokenCount, createdAt) and agentKv:blob:% for agent KV blobs (with content, providerOptions.cursor.modelName). One SessionRef is yielded per conversationId; source_kind="database" and seq is the SQLite rowid so resumable reads use the rowid as a high-water mark. macOS only in v1 — Linux and Windows path constants are present in stackunderflow/adapters/cursor.py but untested.
Cline. Reads each task directory ~/Library/Application Support/Code/User/globalStorage/saoudrizwan.claude-dev/tasks/{taskId}/. Two files per task: ui_messages.json (a flat array of UI events — api_req_started events become assistant records and carry tokensIn / tokensOut / cacheWrites / cacheReads) and api_conversation_history.json (Anthropic-shape messages whose first user message embeds <model>...</model>). One SessionRef per task; seq is the event index in ui_messages.json rather than a byte offset, so resume means "skip first N events." macOS only in v1.
Adapter source: stackunderflow/adapters/cursor.py, stackunderflow/adapters/cline.py. See docs/adapters.md for the contract these adapters implement.
Provider chips render on the session table and project cards, color-coded per provider. When the same project slug is ingested through more than one adapter (the schema's UNIQUE(provider, slug) constraint allows one row per provider), the project card renders one chip per provider. Implementation: stackunderflow-ui/src/components/common/ProviderChip.tsx. The API surfaces the field on /api/projects (as provider plus a providers array) and on /api/jsonl-files (as provider).
Anthropic and GLM model identity and rates are data, not code: they live in stackunderflow/data/models.toml and load via stackunderflow/infra/model_manifest.py. AnthropicPricer delegates to the manifest — there is no hardcoded rate dict or token-matching ladder in infra/providers/anthropic.py anymore.
Each [[model]] entry has:
family— the canonical keycanonicalize()returns (e.g.OPUS_48).match— a token list; an id matches when every token is present after lowercasing and splitting on-/.. Order matters: more-specific entries come first and matching stops at the first hit, so["opus","4","8"]must precede["opus","4"].fast_multiplier(optional) — input/output multiplier for the Anthropic priority/fast tier (Opus =6.0; cache rates are never multiplied).fallback = true(exactly one entry) — the family returned when nothing matches (Sonnet 3.5).- one or more
[[model.price]]rows:input/output/cache_write/cache_readin $/M tokens, plus optionaleffective_from/effective_until(ISOYYYY-MM-DD).
Pricing is effective-dated. rates_for(family, provider, at_ts=...) returns the row whose [effective_from, effective_until) window contains at_ts; omit at_ts for the current rate, and a row with no dates is always-current. This lets a historical event be priced at the rate in effect when it ran, and lets a vendor price change be recorded as a new dated row instead of overwriting history.
To add a model or correct a price, edit models.toml — no code change. For Anthropic, also add the canonical id to _CANONICAL_IDS in infra/costs.py so RATE_CARD membership flips the normalizer's cost_source from unknown to rate_card (the two lists are slated to be unified with the DB-backed registry — see the note there).
Rates are grounded in published sources: the cached LiteLLM feed (~/.stackunderflow/cache/pricing.json, the same feed the live overlay uses) and Anthropic's pricing page. The LiteLLM overlay still takes precedence over the manifest at compute time for models it covers.
Cursor publishes no general per-token rate card — users pay a flat subscription and the IDE multiplexes Anthropic / OpenAI / Google / Cursor-trained models behind the scenes. CursorPricer (stackunderflow/infra/providers/cursor.py) prices each cursor model id one of four ways:
| Class | Examples | Rate source |
|---|---|---|
| Vendor-prefixed | claude-4.5-sonnet-thinking, claude-4.6-sonnet, gpt-5-codex, gpt-4o, gemini-2.5-pro-preview-05-06, gemini-3-pro |
Delegates to AnthropicPricer / OpenAIPricer / GeminiPricer. A Claude-via-Cursor record costs the same as a native Claude record. Gemini ids with -preview-MM-DD or -experimental suffixes strip back to the base id and retry. |
composer-1 |
composer-1 |
Cursor's published rate: input $1.25/M, output $10.00/M, cache-write $1.5625/M, cache-read $0.125/M (cursor.com/docs/models-and-pricing). |
composer-2 |
composer-2 |
ESTIMATED at Anthropic Sonnet 4.x rates: input $3/M, output $15/M, cache-write $3.75/M, cache-read $0.30/M. Cursor publishes no per-token rate for composer-2; Sonnet-tier is the closest acknowledged analogue for an agentic Sonnet-class model. |
| Autoselectors | cursor-auto, cursor-fast |
Same ESTIMATED Sonnet-tier rates. When Cursor picks the model the actual engine is unknown, and the estimate keeps these records out of $0 territory in compare / cost reports. |
An id no delegate recognizes falls back to the same Sonnet-tier estimate rather than returning None, so a record with non-zero token counts always contributes a real dollar figure. The source comments flag every estimated rate.
When a record's record.raw["cost_source"] == "estimated", the UI prefixes its cost with ≈ and exposes a tooltip ("estimated cost — provider does not surface per-message tokens"). The Cursor adapter sets this flag whenever it falls back to the len(text) // 4 heuristic because the bubble has zero tokenCount.{inputTokens, outputTokens} (Cursor v3 returns zero counts on every bubble; see stackunderflow/adapters/cursor.py).
The flag is set on the adapter Record and persists in messages.raw_json. The ETL usage_events pipeline carries cost_source as a first-class column: the normalizer layer (stackunderflow/etl/normalize/) stamps every event with one of live, rate_card, estimated, or unknown (the COST_SOURCE_* constants in etl/normalize/base.py). The ≈ marker on the Sessions table reads the messages-level flag.
flowchart LR
user[User: stackunderflow start] --> ingest[run_ingest]
ingest --> claude[ClaudeAdapter]
ingest --> codex[CodexAdapter]
ingest --> cursor[CursorAdapter]
ingest --> cline[ClineAdapter]
ingest --> openclaw[OpenClawAdapter]
ingest --> pi[PiAdapter]
ingest --> hermes[HermesAdapter]
ingest --> betas[13 more adapters — all always-on]
claude --> store[(SQLite store.db)]
codex --> store
cursor --> store
cline --> store
openclaw --> store
pi --> store
hermes --> store
store --> api[/REST API/]
api --> ui[React dashboard]
Cursor / Cline show no data. Check that the source file exists. Cursor: ls ~/Library/Application\ Support/Cursor/User/globalStorage/state.vscdb. Cline: ls ~/Library/Application\ Support/Code/User/globalStorage/saoudrizwan.claude-dev/tasks/. If either path is missing the adapter exits cleanly without logging an error — the tool is not installed or has not been used on this machine. After confirming the path, re-run stackunderflow reindex.
A provider I use shows no data. Same pattern: confirm the on-disk source exists for that provider (paths are listed in the table above), then re-run stackunderflow reindex. stackunderflow doctor reports, per provider, whether sessions found on disk actually made it into the store.
My Cursor sessions show $0 (or a ≈ marker). As of the cursor-pricing fix, cursor records with non-zero token counts always price at a real dollar figure: vendor-prefixed ids (claude-*, gpt-*, gemini-*) delegate to the upstream pricer; composer-1 prices at Cursor's published rate; composer-2 and the cursor-auto / cursor-fast autoselectors use ESTIMATED Anthropic Sonnet 4.x rates (see "Cursor pricing" above). A record still showing $0 means its token counts are zero — Cursor v3 stores zero tokenCount.{inputTokens, outputTokens} on every bubble, and the adapter's len(text) // 4 fallback estimate also returns 0 when the bubble's text payload is empty (the v3 bubble shape stores rich JSON with diffs and code chunks instead of a top-level text field, so some assistant bubbles legitimately have no estimable text). The ≈ marker on Sessions table rows reflects the estimated-tokens flag; cost is estimated even when the dollar figure is non-zero.
How do I skip an adapter I don't want? There's no runtime toggle — the registry self-discovers every module, so an adapter is disabled by removing or renaming its file in stackunderflow/adapters/ (or excluding it in a custom build). In practice there's rarely a reason to: an adapter with no source data on your machine yields nothing and costs nothing.