feat: add shelltime cc backfill and shelltime codex backfill

[AnnatarHe]

Summary

Adds two commands that upload past Claude Code and Codex usage from the transcripts both tools keep on disk. They fill gaps live OTEL tracking missed: usage from before cc install / codex install, or from while the daemon was down. Events are uploaded as the same kinds the live pipeline sends (user_prompt, api_request, tool_result, Codex sse_event/response.completed, conversation_starts), so dashboards, session summaries and MCP tools pick them up.

Requires shelltime/server#494. Against an older server the command stops with "server does not support backfill yet".

shelltime cc backfill --dry-run     # preview per day, upload nothing
shelltime cc backfill               # upload Claude Code sessions
shelltime codex backfill --since 2026-09-01 --no-prompts

How it works

1. Parse local transcripts. The approach follows ryoppippi/ccusage, with these choices:
   • Claude Code reads ~/.claude/projects, ~/.config/claude/projects, or the CLAUDE_CONFIG_DIR list.
     • One API response is written as several lines that repeat its usage, so api_request is deduplicated globally by message.id:requestId, keeping the most complete usage.
     • Prompts come only from human lines. Meta, subagent, compaction, local-command, interruption and notification lines are excluded.
     • Each tool call is paired with its result. Only file_path / path from the parameters is kept; commands and file contents are never uploaded.
     • Subagent transcripts add to their parent session. History copied into resumed sessions is counted once.
   • Codex reads ~/.codex/sessions, archived_sessions, or CODEX_HOME.
     • Token counts come from last_token_usage, or from the change in the cumulative total. Null and repeated counts are skipped.
     • Tool results carry exit code and duration.
     • History copied into forked rollouts is dropped.
     • Old rollouts with no user_message events fall back to the conversation's own user messages.
2. Ask the server which sessions it already has. Sessions it tracked live, archived, or already received in full are skipped. Sessions still running (activity in the last 30 min) are held back.
3. Upload batches of up to 500 events, 50 completed sessions and 8 MiB. Long prompts and tool arguments are capped. A session is listed as completed only in the request carrying its last event, so the server builds its summary once everything is stored. 5xx, 429 and network errors are retried with backoff; other 4xx errors stop the run.
4. Call /complete so the server refreshes the activity heatmap and caches for the uploaded range.

Event ids are bf1: plus a sha256 of a natural key (message/request id, prompt uuid, tool call id, and so on). They don't depend on the session, so re-runs and copied history are idempotent.

Other changes

• SendHTTPRequestJSON returns a typed HTTPStatusError, so callers can tell retryable errors from fatal ones. The error text is unchanged.
• Backfill events use their own wire type with pointer fields. The live AICodeOtelEvent has success,omitempty on a bool, which drops failed tool calls; backfill sends "success": false.
• aiCodeResourceIdentity() is extracted, so backfilled events carry the same user.name / machine.name / team.id as live ones.
• README: "Backfilling AI Usage" section and command table rows.

Test plan

• Unit tests for the Claude parser:
  • partial streamed usage and repeated usage lines
  • resumed copy credited to the original session; subagent file
  • every skipped prompt kind, slash-command normalisation, is_error sent as false
  • --no-prompts, prompt cap, CLAUDE_CONFIG_DIR list, stable ids
• Unit tests for the Codex parser:
  • null and repeated counts, last vs delta usage
  • fork dedup when the copy appears before any turn context
  • model fallback, both tool output formats, legacy rollouts
• Unit tests for helpers: batch packing (event, byte and completed caps), session selection, a 10 MB JSONL line, file discovery with a symlinked root, a golden event id checked against Python.
• Command tests against an httptest server:
  • only new or partly uploaded sessions are sent
  • large backfills are split; --ai-summary is passed through; --dry-run only checks status
  • a 503 or 429 is retried; a 400 aborts; a 404 explains the server is too old; a missing token is reported
  • Codex end to end; flag parsing
• go vet ./... and go test ./... pass.
• End-to-end run against shelltime/server#494 locally (Postgres + Redis), using a real Claude Code 2.1.289 transcript with 4 subagent files:
  • the dry run showed 1 session, 519 events, 2 prompts, 222 requests
  • the upload sent 2 batches; all 222 requests were costed and the summary matched (tokens, 295 tool calls with 5 failed, cwd, version)
  • a second run reported "already uploaded" and stored nothing new
  • Claude token totals matched an independent Python implementation of ccusage's dedup exactly

🤖 Generated with Claude Code

https://claude.ai/code/session_01LMCwtubXzBYhQTnX44vhzF

---

Generated by Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[claude]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.

[codecov]

Codecov Report

❌ Patch coverage is 89.42308% with 99 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
model/aicode_backfill_codex.go	85.76%	39 Missing ⚠️
model/aicode_backfill_claude.go	89.43%	26 Missing ⚠️
commands/aicode_backfill.go	90.99%	20 Missing ⚠️
model/aicode_backfill_common.go	91.72%	12 Missing ⚠️
model/api_aicode_backfill.go	95.00%	2 Missing ⚠️

Flag	Coverage Δ
unittests	84.57% <89.42%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
commands/cc.go	78.57% <ø> (+10.82%)	⬆️
commands/codex.go	100.00% <ø> (ø)
model/aicode_otel_claude_settings.go	85.35% <100.00%> (+11.82%)	⬆️
model/api.base.go	92.20% <100.00%> (+3.74%)	⬆️
model/api_aicode_backfill.go	95.00% <95.00%> (ø)
model/aicode_backfill_common.go	91.72% <91.72%> (ø)
commands/aicode_backfill.go	90.99% <90.99%> (ø)
model/aicode_backfill_claude.go	89.43% <89.43%> (ø)
model/aicode_backfill_codex.go	85.76% <85.76%> (ø)

... and 108 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.