What changed, and which pull request changed it.

Turn rows survive a merge. human_turns is registered in SESSION_CHILD_TABLES with its vocab_json inside the row, so archie merge carries a merged session's turnaround facts and the insights vocabulary block keeps working without a second table keyed by a volatile rowid.

Antigravity CLI sessions carry their models, and the harness can wake them. The agy conversation store records the model that generated each step in field 19 of its gen_metadata blobs; the adapter read everything else out of that store and never that field, so every antigravity row indexed with models_used: [] while claude_code and codex rows carried theirs. The adapter now reads it (parser version 3). Later tip work also reads gen_metadata usage varints into token counters where the store actually carries them (see Added above). archie session wake --inject antigravity turns an agy PreInvocation hook payload into the injectSteps wake envelope, and archie hook print antigravity prints the ~/.gemini/config/hooks.json entry that points at it. Because agy exposes no SessionStart hook, an on-disk marker makes the per-invocation PreInvocation fire the wake document exactly once per conversation; see apps/cli/src/antigravity_hook.rs.

Wake resumes the worktree you are standing in, not the newest session for the repo. extract_repository_or_workspace prunes the --claude-worktrees- suffix of a Claude Code project slug on purpose, which is right for carry-forward but meant that with several agents in worktrees under one repository, "newest for repo" woke agent A into whichever worktree ran last. Selection now prefers sessions whose recorded workspace.cwd equals the checkout's git root and falls back to the repo ordering only when none match (Storage::list_sessions_for_repo_preferring_workspace).

Claude / Codex discovery scope. Non-session files rejected at discovery; stale rows pruned so the index does not advertise ghosts.

cost_weighted_tokens beside total_tokens. total_tokens adds cache_read_tokens at face value, so on a long session it measures how large the context grew rather than what the session spent -- 98.7% of the transcript below was cache reads of a prompt already paid for. total_tokens keeps that meaning for compatibility; the four counters are now exposed separately everywhere, alongside input + output + 1.25 x cache_creation + 0.1 x cache_read (Anthropic's published prompt-caching multipliers, verified 2026-09-10 -- not dollars, and not exact for a 1-hour cache write; a comparable weight for ranking sessions). Surfaced in session_list, session_show, /api/traces, stats_usage (new CACHE WR and WEIGHTED columns, width-gated so the 80-column table is byte-for-byte unchanged) and stats_ladder.

docs/specs/token-accounting.md -- the measurement table, the ledger's contract, why Codex is excluded, and what cost_weighted_tokens deliberately is not.

Redactor::for_publication() -- the default rules plus path-tail and forge-repo stripping, for the one path where text leaves the machine. Additive and opt-in: Redactor::new() is unchanged, because repo_blame and session_show exist to tell you which repo a session touched. It raises the floor rather than making a payload provably clean -- a project name inside a relative path still survives, and the test suite asserts that limit rather than claiming otherwise.

The dashboard shows weighted tokens beside raw ones, flags cache-inflated sessions, and can sort by weighted spend. UsagePeriodSummary and ModelUsagePeriodSummary gained cost_weighted_tokens server-side so no client re-derives the multipliers.

archie policy init [--repo] [--force]. Writes a starter policy.toml with the thrash and loop rules on, and a [spend] block left commented out, carrying this machine's own primary-session token percentiles (p50/p90/p99/max) as comments — read from the index with the new Storage::session_token_percentiles.

archie policy check warns on a spend cap below your own p99. Names the cap, the machine's (or repo's) p99, and how many already-indexed sessions would have tripped it. Still exits 0 — a warning, not an error.

Sessions have a kind: conversation or fleet_snapshot. One column, one migration (every existing row is a conversation). The stub predicates (total_events > 1 AND total_tokens > 0, and the scanner's near-empty rule) apply to conversations only; a snapshot is outside that population, hidden from the default session list, stats and the session aggregate, and shown whole by session list --kind fleet_snapshot or --adapter <snapshot adapter>. Adapters declare it through AgentAdapter::trace_kind. --all-stubs and the include-stubs aggregate relax only the events/tokens bar, never the kind, so neither surfaces a snapshot; the raw "rows in the SQLite index" count behind archie scan's Total Indexed line and archie doctor moved to Storage::total_row_count, which counts every row (snapshots included) and can exceed the conversation aggregate.

identity_sightings: names are display, the harness session id is the key. A pane label, a Herdr agent name, a session title can all be renamed. Every trace can carry identities (session id, name, time seen, via), stored append-only keyed on (session, name, via): a rename adds a row, never rewrites one. Storage::identity_sightings(session_id) reads them back. Written by the Herdr adapter today; the field is open to any adapter, and Claude Code session titles do not land there yet.

PARSER_VERSION bumps: Claude Code 2 to 3, Gemini 1 to 2, OpenCode 1 to 2. An incremental scan reparses these sessions even though their bytes never changed. An existing index keeps its inflated token totals until archie scan runs.

The [spend] example in crates/loop/src/policy.rs's doc comment dropped its token number — [spend] now ships commented out, pointing at archie policy init. A 20M-token example would have halted this machine's own 728M-token primary sessions dozens of times; the rule stands, the number is now yours to set from your own history.

A streamed message's tokens are counted once, not once per record. Claude Code writes one assistant message as several JSONL records -- one per content block, apiBlockIndex 0..n -- and every one repeats the same message.usage block verbatim. The parser summed usage per record. On a real 1.08 MB subagent transcript that read 33,777,792 tokens where the true figure is 19,294,496; across a 120-file sample of ~/.claude/projects the machine-wide ratio was 2.194x. Gemini CLI has the same shape keyed on id (1.90x, measured). One shared ledger (crates/adapters/src/usage_ledger.rs) credits each message id at most once, last usage block wins -- only the usage is suppressed, never the record's own content, so a tool call on block 1 is still parsed. Codex is deliberately untouched: its token_count events are cumulative running totals and codex.rs already takes deltas of them. OpenCode's SQLite path gets one row per message from its primary key; its legacy file path now carries the same guard.

Gemini CLI sessions no longer index as zero tokens. extract_token_usage knew promptTokenCount-style and OpenAI-style names but not the ones Gemini actually writes -- input / output / cached / thoughts / tool inside a tokens block -- so every Gemini CLI session read as zero despite real usage. thoughts and tool fold into output, matching Gemini's own total.

archie session list --json carries the token breakdown. It builds its own JSON and never touched SessionSummary, so the four counters and cost_weighted_tokens never reached it.

The maintainer's system username and a third party's project name are out of the public tree. This repo carried one person's real machine layout in 101 places, and -- worse -- a third party's project name in shipped code and user-facing copy: is_leaked_katana_var was a function name and "The Missing local Weapon (The Katana Incident)" printed to anyone who tripped that rule. Now is_leaked_loop_var and "The Missing local Weapon", which describes what the detector actually finds. Product doc comments that addressed the maintainer by name now name the role instead. Fixture paths come from the new agentworth_schema::fixtures rather than literals, so the convenient option is also the safe one.

archie blunder shows you the bytes before it posts them. The exhibit printed to your terminal is the local view, with full absolute paths; what a submission POSTs is a redacted payload you never saw. One redacted_payload() now feeds both, so they cannot drift, and --dry-run prints exactly what would be sent and sends nothing. Submission also switched to the new Redactor::for_publication() -- blunder --submit redacts loose strings, so unlike redact_trace it never picked up the per-trace repository_identity_rule, and a captured payload still carried three repositories' names after full default redaction.

The Herdr adapter reads the file Herdr actually writes. It used to look for coordination_traces, parent_agent_id and delegation_id, fields no Herdr file has ever carried, so the real ~/.config/herdr/session.json parsed to nothing while five fixture tests stayed green. It now decodes the v3 shape (workspaces[].tabs[].panes{}) as a fleet_snapshot: one herdr.pane event per pane with pane_key (Herdr's internal number), pane_number (what the user sees, from public_pane_numbers), cwd, label, agent name and agent_session.value, the harness session id that is the only key. The live pane id (w9:pA) is a runtime allocation, not in the file, and is never stored. Every label and agent name goes to identity_sightings with the time seen. Scalars decode leniently (a "version": "3" string, a non-numeric active), active_tab and active gate focused so one pane per file is focused, one bad workspace, tab, pane or agent_session costs only itself, unknown fields are ignored and a newer version is a warning. PARSER_VERSION 2. Capabilities report none of the seven. Fixture: a redacted real herdr 0.8.2 file.

install.sh runs the binary once and explains a Gatekeeper kill. The macOS binaries are ad-hoc signed and not notarized. A fresh curl | sh install runs (measured on macOS 27: valid signature, no quarantine flag); a copy that carries the flag is killed on exec with nothing but zsh: killed, exit 137 (measured by adding the flag; removing it clears it). The script now removes any existing file before copying (a cp through the npm wrapper's agentworth symlink would have overwritten the wrapper's JavaScript), drops the flag from what it wrote with xattr -dh (act on the file, not a symlink target), runs archie --version, and on exit 137 exactly prints the xattr -d line for the installed files. Other failures are reported as what they are. README carries the note.

archie doctor and agent show name Herdr's real directory, ~/.config/herdr/, not ~/.herdr/.

• 🐋 DeepSeek Code (deepseek):
  • Discovers ~/.deepseek/, ~/.deepseek-coder/, and .deepseek/ traces.
  • Full reasoning token stream accounting for DeepSeek R1 and V3 (reasoning_content, thought).
  • Tracks prompt cache hits (prompt_cache_hit_tokens) and cache creation (prompt_cache_miss_tokens).
  • Normalizes file editing (str_replace_editor, write, edit) and shell execution (bash).
• 🌙 Kimi Code (kimi):
  • Discovers Moonshot Kimi Code sessions in ~/.kimi-code/ and ~/.kimi/sessions/wire.jsonl.
  • Parses streaming wire JSONL protocols, subagent delegations (subagent_delegation), and tool calls.
• ⚡ MiniMax (minimax):
  • Discovers ~/.minimax/ and ~/.minimax-agent/ coding plan trajectories.
  • Normalizes multi-step planning milestones, tool executions, and token expenditures.
• 🐉 Qwen Code / Qwen-Agent (qwen):
  • Discovers Alibaba Qwen Code and Qwen-Agent trajectories in ~/.qwen/ and ~/.qwen-agent/.
  • Extracts reasoning CoT, code_interpreter executions, and tool calls.
• 🧠 Zhipu / CodeGeeX (zhipu):
  • Discovers ~/.codegeex/ and ~/.zhipu/ IDE extension and GLM-4 session histories.
• 🛠️ Aider (aider):
  • Discovers .aider.chat.history.md and ~/.aider/ git-driven trajectory markdown/JSON logs.
  • Maps git diff edits and commit messages directly into verified outcome evidence (OutcomeKind::CommitObserved).
• 👁️ Cline & Roo-Code (cline):
  • Discovers VSCode global storage task logs (saoudrizwan.claude-dev/tasks/ and rooveterinaryinc.roo-cline/tasks/).
  • Parses task UI messages, API conversation histories, token cache metrics, and tool execution trees.
• 🌊 Windsurf / Cascade (windsurf):
  • Discovers ~/.codeium/windsurf/ and Cascade execution caches.
  • Normalizes multi-turn code edits, terminal outputs, and test validations.
• 🦾 Manus (manus):
  • Discovers ~/.manus/ autonomous agent browser actions and coding trajectories.

• Deep usage analytics by timeframe with --period day|week|month and --limit.
• Aggregates sessions, input tokens, output tokens, prompt cache reads, and estimated USD spend.
• Real-time rolling pacing telemetry with --pacing (default 5-hour window via --hours 5):
  • Token burn velocity (tokens/hour).
  • Prompt cache hit ratio percentage.
  • Active agent adapters and models within the pacing window.
  • Estimated dollar expenditure tracking.
• Machine-readable JSON output via --json.

• Trace code alterations and file edits back to the specific AI agent session, model, sequence timestamp, and user prompt that created or modified them.
• Full support across all 11 supported agent adapter ecosystems.

• Added native SQLite aggregation views: v_daily_usage, v_weekly_usage, and v_monthly_usage with pre-computed token math and cache metrics.

• Added native agwt command-line alias alongside agentworth.

• Fixed an issue where corrupt session timestamps or epoch zeroes (1970-01-01) distorted aggregate date bounds by enforcing MIN(CASE WHEN started_at > '2020-01-01' THEN started_at END) in SQLite aggregate queries.

• Configured optimized SQLite WAL mode, busy_timeout = 5000ms, synchronous = NORMAL, and cache_size = -64000 (64MB) to prevent database locking during rapid parallel parsing.

• Optimized SHA-256 fingerprint checks to instantly skip unchanged multi-gigabyte session JSONL transcripts.

• Standalone script: curl -fsSL https://agentworth.dev/install.sh | sh
• Homebrew: brew install unfoundbox/tap/agentworth
• Cargo: cargo install agentworth-cli
• NPX: npx agentworth or npx agwt

• 100% offline, local-first discovery and streaming JSONL ingestion engine.
• Unified AgentWorthTrace and NormalizedEvent canonical schema.

• Claude Code (claude_code)
• Cursor Composer (cursor)
• Google Antigravity (antigravity)
• OpenAI Codex (codex)
• Block Goose (goose)
• Pi (pi)
• Herdr (herdr)
• Nous Hermes (hermes)
• OpenClaw (openclaw)
• xAI Grok (grok)
• OpenCode (opencode)

• Deterministic outcome verification (DoneClaimed < ArtifactChanged < TestOrBuildPassed < CommitObserved < CiOrDeploymentVerified).
• Explainable 5-factor TraceScore rating.

• agentworth scan — Discovers and indexes local session logs.
• agentworth stats — Machine-wide token expenditures and top model usage.
• agentworth traces — Tabular session directory with filters.
• agentworth inspect — Step-by-step ASCII trajectory timeline.
• agentworth doctor — System health and adapter discovery diagnostics.
• agentworth export — ATIF v1.0 and JSON export with 13-rule offline privacy scrubber.
• agentworth serve — Local embedded Axum API server and monochrome receipt explorer UI.