The design doc behind every feature.

• Archie bencharchie-bench.md

  A leaderboard built only from this machine's own sessions: model by effort by repo, ranked on verified rate, with tokens, tool calls and an API-equivalent cost beside every row an…

  proposed, measured 2026-09-02.
• Archie — the memory agents do not havearchie.md

  MCP server shipped 2026-09-02 (#56); this spec's precondition is met.

  proposed.
• Asks: a questions-to-answers indexasks.md

  In a long session Saurabh asks a question. The answer lands several messages later, after tool calls, subagent notifications, and other assistant text — buried, not missing. He re…

  built, PR #97 (2026-09-02).
• Beliefsbeliefs.md

  AgentWorth already extracts what a session decided and verifies what it claims it did. It never checks whether a stated fact is still true. Give it a claim_check a session can cal…

  proposed, measured 2026-09-02.
• Cache economics — what returning cold actually costscache-economics.md

  A session's own logs already record what every turn paid to re-read its own history. Nobody surfaces it, so nobody learns from it. This spec turns that into a number a developer c…

  built, #42.
• CLI grammar, completions, and the cockpitcli-grammar.md

  Thirty-two top-level commands is a list, not a grammar. Nobody can guess blind-spots or cache-doctor from having used handoff. The fix is noun-then-verb: four nouns carry everythi…

  sections 1, 2 and 4(1) built 2026-09-02 in #118.
• Compaction diffcompaction-diff.md

  Four things the build settled that this document left open:

  built, PR #83 (2026-09-02).
• Compaction, and what it costscompaction.md

  Measured on a real index of 612 sessions, not estimated.

  built, #57 (dashboard pane) and #62 (per-session tracking).
• Context composition — what fills the window before the work doescontext-composition.md

  Working name in conversation was "chewed". This document argues for context composition, on the grounds below.

  built, #46.
• Convergenceconvergence.md

  Where in a session the last verified progress happened, how much was spent after it, and whether anything that could fail ran at all.

  proposed, measured 2026-09-02.
• Desktop App — feasibility specdesktop-app.md

  Question this answers: would a Tauri wrapper around AgentWorth produce a working .dmg, what would that actually cost, and what has to be decided first.

  proposal, not started.
• Dropped commitmentsdropped-commitments.md

  An agent says it will do something and then does not. It is not lying; the intent was real when it was written and the conversation moved on. But the commitment is now in a transc…

  built, #44 and agentworth loose-ends.
• Efficiency receiptsefficiency-receipts.md

  Two weeks of local transcripts, four deterministic detectors, and a receipt per five-hour window saying which repeated work was provable, how long it sat in the window afterwards,…

  proposed, P0 measured 2026-09-02.
• Fleet viewfleet-view.md

  Nobody can currently look at AgentWorth and answer "what's running right now, across everything on this machine." OverviewPane.tsx shows VerdictBoard (aggregate outcome distributi…

  built, #41 (fleet strip) and the SSE live-tail endpoint (v0.
• Handoffhandoff.md

  The rest of this document is the design as measured on 2026-09-02.

  built, PR #TBD (2026-09-02).
• Local search and the small-model anglelocal-search.md

  Correction, verified 2026-09-01. An embedding pipeline already exists in this repo — crates/storage/src/{chunker,embedder,vector} and a working agwt search command. This spec was…

  draft spec, not yet built (mostly).
• Loose endsloose-ends.md

  The command is agentworth loose-ends.

  proposed.
• Market map: observability that opens PRsmarket-autofix.md

  Research date: 2026-09-01. Every claim below is dated to what the cited page said on or near that date — pricing and positioning in this space change often. Where I could not veri…

• MCP servermcp-server.md

  The owner runs many long coding-agent sessions across several machines and repos. Every day he hand-writes handoff files so the next session knows what the last one did. A better…

  built.
• The questions worth answeringquestions.md

  Build the answers before designing screens for them. Most of these will turn out to be boring, and finding out which ones are boring is cheap.

  proposed.
• The AgentWorth and SpacePilot loopspacepilot-loop.md

  AgentWorth feeds SpacePilot's registry dream, and SpacePilot feeds AgentWorth's missing intelligence layer.

  direction, decided 2026-09-02; endpoints verified against spacepilot fc…
• Suspect commitssuspect-commits.md

  Shipped as the MCP tool suspect_commits and the CLI agentworth suspect, with agentworth suspect --hook printing the pre-push script. Two things changed between this spec and what…

  built, PR #84, measured 2026-09-02.
• Trajectory scrubbertrajectory-scrubber.md

  apps/dashboard/src/shell/TimelineStrip.tsx draws three rows of ticks (messages, model, tools) across the width of the inspector. Bucket count comes from measured container width,…

  built, #40.
• Verified outcome rateverified-outcome-rate.md

  Measured 2026-09-02, corrected the same day. Rung 3 used to be granted from the command string alone: is_test_or_build_command("cargo test") was true, so the session was recorded…

  proposed, measured 2026-09-02.

24 files, read straight out of docs/specsSource on GitHub