cli: stats totals differ from stored session usage #57878

Activity

1. maharshi365

   changed the title [-]fix(cli): reconcile stats with recorded token usage[/-] [+]cli: stats totals differ from stored session usage[/+] on Oct 21, 2026
2. opencode-agent commented on Oct 21, 2026

   

   opencode-agentBot

   on Oct 21, 2026 – with OpenCode Agent

   Contributor

   More actions

   Thanks for the detailed analysis. I reproduced this on 2.1.2 and it still happens on the latest dev (863ea41).

   Setup: a local mock OpenAI-compatible provider that reports 1000 input + 100 output tokens for each completion.

   1. Send two prompts. opencode stats shows 2 prompts · 2 steps · 2.2k tokens.
   2. Stage a revert to the second prompt, then send a new prompt, which commits the revert and runs another model step.
   3. opencode stats still shows 2 steps · 2.2k tokens, although three completions (3.3k tokens) were used. The session's stored totals still include the reverted step (input 4000, output 400, cost $0.008).

   So the reverted step's usage drops out of stats once the revert is committed. That matches your reading: stats add up the remaining message rows, and committing the revert deletes them.

   One more thing I noticed: the stored session totals also include the session-title generation call, which stats never count (related: #46371). That is probably part of your measured gap besides reverts.

   The script below reproduces it.

   Reproduction script (repro.sh)

   #!/usr/bin/env bash
   # Reproduce anomalyco/opencode#57878: committed reverts drop consumed usage from `opencode stats`.
   # Usage: ./repro.sh            (uses `opencode` from PATH)
   #        OC_DEV=/path/to/repo ./repro.sh   (runs CLI from a source checkout)
   set -euo pipefail
   W=$(mktemp -d /tmp/opencode/repro57878.XXXX)
   export HOME=$W/home XDG_DATA_HOME=$W/home/data XDG_CONFIG_HOME=$W/home/cfg XDG_STATE_HOME=$W/home/state XDG_CACHE_HOME=$W/home/cache
   mkdir -p $HOME $W/proj
   export OPENCODE_SERVER_PASSWORD=repro57878
   oc() {
     if [ -n "${OC_DEV:-}" ]; then
       bun --conditions=browser --preload=$OC_DEV/packages/tui/node_modules/@opentui/solid/scripts/preload.js $OC_DEV/packages/cli/src/index.ts "$@"
     else opencode "$@"; fi
   }
   # Mock OpenAI-compatible provider: every completion reports 1000 input + 100 output tokens.
   cat > $W/mock.ts <<'TS'
   Bun.serve({ port: 4011, async fetch(req) {
     const body: any = await req.json().catch(() => ({}))
     const usage = { prompt_tokens: 1000, completion_tokens: 100, total_tokens: 1100 }
     const base = { id: "c1", created: 1, model: "mock" }
     if (!body.stream) return Response.json({ ...base, object: "chat.completion", choices: [{ index: 0, message: { role: "assistant", content: "ok" }, finish_reason: "stop" }], usage })
     const c = (o: any) => `data: ${JSON.stringify({ ...base, object: "chat.completion.chunk", ...o })}\n\n`
     return new Response(c({ choices: [{ index: 0, delta: { role: "assistant", content: "ok" }, finish_reason: null }] }) + c({ choices: [{ index: 0, delta: {}, finish_reason: "stop" }] }) + c({ choices: [], usage }) + "data: [DONE]\n\n", { headers: { "content-type": "text/event-stream" } })
   } })
   TS
   bun $W/mock.ts & MOCK=$!
   cat > $W/proj/opencode.json <<'JSON'
   { "model": "mock/mock", "small_model": "mock/mock",
     "provider": { "mock": { "npm": "@ai-sdk/openai-compatible", "name": "Mock",
       "options": { "baseURL": "http://127.0.0.1:4011/v1", "apiKey": "x" },
       "models": { "mock": { "name": "Mock", "cost": { "input": 1, "output": 10 } } } } } }
   JSON
   cd $W/proj && git init -q
   oc serve --port 4097 >$W/server.log 2>&1 & SRV=$!
   trap "kill $MOCK; pkill -f \"serve --port 4097\"" EXIT
   for i in $(seq 60); do curl -s -o /dev/null http://127.0.0.1:4097/ && break; sleep 1; done
   DB_DIR=$XDG_DATA_HOME/opencode
   q() { DB=$(ls $DB_DIR/opencode*.db | head -1); bun -e "import {Database} from 'bun:sqlite';const db=new Database('$DB',{readonly:true});$1"; }
   oc run --server http://127.0.0.1:4097 "first" >/dev/null 2>&1
   oc run --server http://127.0.0.1:4097 -c "second" >/dev/null 2>&1
   echo "== stats after two model steps"; oc stats --server http://127.0.0.1:4097 2>&1 | grep -E "prompts"
   S=$(q "console.log(db.query('select id from session_v2').get().id)")
   U2=$(q "console.log(db.query(\"select id from session_message where type='user' order by seq\").all()[1].id)")
   echo "== stage revert to second prompt ($U2), then commit it by sending a new prompt"
   oc api --server http://127.0.0.1:4097 session.revert.stage --param sessionID=$S -d "{\"messageID\":\"$U2\"}" >/dev/null
   oc run --server http://127.0.0.1:4097 -s $S "third" >/dev/null 2>&1
   echo "== model steps actually consumed: 3 (3.3k tokens, excluding title generation)"
   echo "== stored session counters:"; q "console.log(db.query('select tokens_input, tokens_output, cost from session_v2').get())"
   echo "== stats after revert commit:"; oc stats --server http://127.0.0.1:4097 2>&1 | grep -E "prompts"

   Run it: bash repro.sh
3. opencode-agent

   added

   reproduced

   on Oct 21, 2026
4. projectescape commented on Oct 23, 2026

   

   projectescape

   on Oct 23, 2026

   More actions

   For the workaround, I'd keep a response-usage record separate from the editable conversation. A revert changes which messages remain in context, but shouldn't remove already-consumed model usage. The title-generation call in the reproduction is another reason to count completed provider requests rather than only visible assistant turns.

   Disclosure: I build Torana, a local proxy. Its usage_logger plugin writes content-free JSONL with request ID, model, input/output and cache counts for completed responses routed through it. That gives a separate record to compare against stats after a revert, without reading prompts. It starts when enabled and is best-effort, with bounded file rotation; it isn't a billing ledger or a fix to OpenCode's counters.

   I haven't run this particular revert reproduction through Torana, so I'm suggesting an independent measurement path, not claiming a verified fix.
5. maharshi365 commented on Oct 23, 2026

   

   maharshi365

   on Oct 23, 2026

   ContributorAuthor

   More actions

   5. For the workaround, I'd keep a response-usage record separate from the editable conversation. A revert changes which messages remain in context, but shouldn't remove already-consumed model usage. The title-generation call in the reproduction is another reason to count completed provider requests rather than only visible assistant turns.
   Disclosure: I build Torana, a local proxy. Its usage_logger plugin writes content-free JSONL with request ID, model, input/output and cache counts for completed responses routed through it. That gives a separate record to compare against stats after a revert, without reading prompts. It starts when enabled and is best-effort, with bounded file rotation; it isn't a billing ledger or a fix to OpenCode's counters.
   I haven't run this particular revert reproduction through Torana, so I'm suggesting an independent measurement path, not claiming a verified fix.

   Why tho? You could just read from durable event stream. No need to duplicate data.
6. projectescape commented on Oct 23, 2026

   

   projectescape

   on Oct 23, 2026

   More actions

   fair, for fixing stats itself the durable event stream is the right source. i suggested an outside record because you mentioned your event table was empty, so there was no event history to check the gap against. useful as a cross-check, not as the fix.