* unify per-run token + cost accounting across agents
every agent harness now logs the same 5-column (or 6 with cost) table and
populates the same AgentUsage contract, regardless of agent or upstream
provider. previously OpenCode and the Claude fallback path emitted a 3-col
table whose "Input Tokens" was actually only the non-cached delta, silently
dropping cache read/write — real runs were being reported at ~0.4% of their
true input (e.g. one baseline showed Input=30 while step_finish events
summed to cache_read=724,753).
changes:
- add logTokenTable helper in action/agents/shared.ts with stable columns:
Input | Cache Read | Cache Write | Output | Total | Cost ($). cost
column renders only when a value is known.
- action/agents/opencode.ts: accumulate step_finish.part.tokens AND
step_finish.part.cost (sourced from models.dev inside opencode —
confirmed working across Anthropic, OpenAI, Google, xAI, DeepSeek,
Moonshot, and OpenRouter). drop the event.stats.total_tokens fallback
since that payload has no cache breakdown.
- action/agents/claude.ts: success-path now treats input_tokens as the
non-cached field (matching OpenCode semantics), carries
cache_read_input_tokens / cache_creation_input_tokens separately, and
captures total_cost_usd from the final result event. the per-message
fallback accumulator now captures cache fields too so it's no longer
lossy when the result event never fires.
- formatUsageSummary gains a Cost ($) column that matches the stdout
table row-for-row; missing values render as "—".
- scripts/token-usage.ts parses all three historical formats (new 5-col,
legacy 4-col Claude success, legacy 3-col lossy) and explicitly flags
the lossy runs instead of averaging misleading values.
validation (pnpm play --local, identical "say hello" prompt):
agent+model Input CacheR CacheW Output Total Cost
OpenCode + Anthropic Sonnet 4.6 4 41,177 20,735 129 62,045 $0.0921
Claude CLI + Anthropic Sonnet 4.6 9 80,133 11,611 389 92,142 $0.0766
OpenCode + OpenAI codex-mini 10,893 46,976 0 606 58,475 $0.0059
OpenCode + Google Gemini 3 Flash — — — — — $0.0114
OpenCode + xAI Grok 4 Fast — — — — — $0.0035
OpenCode + DeepSeek Chat 18,854 0 0 1 18,855 $0.0053
OpenCode + Moonshot Kimi K2.5 — — — — — $0.0106
OpenCode + OpenRouter→Anthropic — — — — — $0.0617
OpenCode + OpenRouter→OpenAI — — — — — $0.0038
* isolate play.ts from developer gitconfig
play.ts is a CI-emulator but inherits the developer's user- and system-scope
gitconfig. a common local convenience — url."git@github.com:".insteadOf
"https://github.com/" to force SSH auth — gets applied at read time on every
git call inside the temp repo, causing `git remote get-url --push origin`
to return an SSH URL instead of the stored HTTPS one. pullfrog_push_branch's
validatePushDestination (correctly) treats that as tampering and blocks the
push. the agent then burns the full MAX_COMMIT_RETRIES budget trying
workarounds that can't beat a user-scope insteadOf rule, turning a trivial
"say hello" run into a 1.35M-token session.
point GIT_CONFIG_GLOBAL and GIT_CONFIG_SYSTEM at /dev/null inside run() so
the play process and its spawned agent see the same empty gitconfig that
a real CI runner would. CI has no rewrites, so this is a no-op there; dev
machines get CI-identical git state. SSH client config (~/.ssh/config and
keys) is separate from gitconfig and is unaffected, so setupTestRepo's SSH
clone still works locally. setupGit only writes --local scope, so nothing
downstream depends on user-scope values.
verification: with the scratch repo cleaned up and this isolation in place,
OpenCode + Anthropic on the same "say hello" prompt goes from 1,349,654
tokens / $2.00+ to 62,045 tokens / $0.0921 — no retry loop, no push blocks.
* persist aggregated token + cost usage to WorkflowRun
AgentUsage has been memory-only — rendered into the GitHub step summary
and then discarded when the runner tears down. that made questions like
"avg cost per customer per day" require log-spelunking. persist it:
- add Int? columns for inputTokens / outputTokens / cacheReadTokens /
cacheWriteTokens and a Decimal? costUsd column on workflow_runs.
Int4's 2.1B ceiling is ~200x larger than any realistic run so BigInt
would be overkill. costUsd uses the same default Decimal precision
as existing money columns (accounts.usageUsd, proxy_keys.hwmUsage).
- extend PATCH /api/workflow-run/[runId] to accept the new numeric
fields alongside the existing artifact strings. per-field type
validation ensures the allowlist stays scalar-safe and rejects
negative / non-finite values.
- generalize patchWorkflowRunFields in the action so it accepts a
mixed string/number payload, and add an aggregateUsage(entries)
helper that sums per-agent AgentUsage records into a single patch.
- call the reporter from main.ts's outer finally block, gated on
toolContext. this is the shared cleanup path that every agent
implementation flows through — claude.ts, opencode.ts, and any
future harness all push their AgentUsage into toolState.usageEntries
via the same line 468, so one finally-block call covers them all.
running in finally also means partial usage gets persisted even
when the agent errored out mid-run.
* anneal token + cost accounting
follow-up polish from a review pass:
- aggregate usage across commit-retry iterations inside each agent harness.
previously runClaude / runOpenCode returned only the final retry's usage,
so any run that hit the dirty-tree retry loop under-counted tokens and
cost in both the stdout table and the WorkflowRun row. added a shared
mergeAgentUsage helper in agents/shared.ts; both harnesses now fold each
iteration's usage into a running total and return the sum.
- scripts/token-usage.ts now handles the unified format with or without
the Cost ($) column. previously the int-only number regex rejected
decimals and the 5-cell length check rejected 6-cell rows, so logs
from post-cost-tracking runs fell through to "no token table". the
parser now accepts both 5- and 6-cell unified rows, splits int vs
decimal cells, and averages reported Cost alongside the tokens.
- PATCH /api/workflow-run/[runId] now rejects INT field values above
INT4_MAX (2_147_483_647) so a malformed payload gets a clean 400
instead of propagating a Prisma error. also defends against a
compromised runner sending a deliberately huge value.
- clarifying comments: opencode.ts documents that step_finish.part.cost
is a per-step delta (empirically verified), main.ts explains that
toolState.usageEntries already carries merged per-retry usage so
aggregateUsage just sums entries (one per agent.run()).
- tests for aggregateUsage and mergeAgentUsage — 12 new cases covering
empty / partial / multi-agent inputs and the "keep undefined" semantic
that prevents spurious zeros from being persisted.
- drop `as number` cast in logTokenTable — narrow via const instead.
* anneal: clamp INT overflow + guarantee mergeAgentUsage immutability
second review pass surfaced two defensive gaps:
- a single token field exceeding INT4_MAX would pass the client but be
rejected by the server's per-field validator, writing a partial row
with some NULLs where sums belonged. clamp in aggregateUsage so the
wire payload is always self-consistent across all numeric columns,
with a loud warning so the clamp doesn't silently swallow weirdness.
- mergeAgentUsage's single-sided branches returned the input reference.
callers treat AgentUsage as immutable but future callers might not;
always return a fresh shallow copy instead. two new tests guarantee
the no-mutation-leak property.
no behavior change in the happy path — INT4_MAX is ~200x the largest
realistic per-run token count.
* anneal: resilient usage persistence + cross-platform null device
third review pass surfaced three small issues:
- main.ts finally block: writeGitHubUsageSummaryToFile throwing would
skip the WorkflowRun usage PATCH. both are independent best-effort
cleanup tasks — wrap the former in catch so a filesystem failure
doesn't block DB persistence.
- AgentUsage.inputTokens had no jsdoc explaining that it's the full
billable input (cached + non-cached). the same word "Input" means
"non-cached only" in the stdout/markdown tables (derived by
subtraction). document the semantic so dashboards querying
WorkflowRun.inputTokens don't misinterpret it.
- play.ts gitconfig isolation was hard-coded to "/dev/null" which
doesn't exist on Windows. use `os.devNull` for cross-platform
parity (resolves to `\\.\nul` on win32). the project is Linux-only
in CI so this only helps local Windows contributors, but it's a
zero-cost swap.
also updated the finally-block caveat comment: usage is only pushed
to toolState.usageEntries when agent.run() returns an AgentResult,
not when the timeout race rejects — so timed-out runs don't
persist partial usage. documented instead of trying to thread state
through Promise.race.
* anneal: NaN-guard cost accumulators + clarify inputTokens docs
final polish from review round 4:
- guard both cost accumulators (opencode step_finish.part.cost and claude
result.total_cost_usd) with Number.isFinite. `typeof x === "number"`
accepts NaN, and one NaN `+=` would poison the running total for the
whole session.
- reword prisma schema comment on WorkflowRun usage fields to call out
that cacheReadTokens / cacheWriteTokens are SUB-totals within
inputTokens (not additional tokens on top). prevents future dashboards
from double-counting by ~2x when summing "total tokens used".
Colin McDonnell
f662b1a0c8