Phase 1 · Agent RSS v3 · FULL REPORT · 2026-09-16

Every Number,Every Failure.

238 runs — the 176-run matrix plus an n=3 replication of the two heaviest classes — across six harnesses, three gateway models and ten task classes. 228 verified, zero timeouts, zero idle-kills. Median own-process RSS, whole-tree cost, kernel footprint and wall time per harness and per model below.

238runs (176 matrix + 62 replication)

228/238verified completions

0timeouts · idle-kills

483corpus files · hash 205507698927dd67

Harness totals, all 238 runs

Own RSS, Tree, Footprint, Wall.

Median of the harness’s own process, with observed range. Tree adds every helper it spawns; footprint is the kernel’s phys_footprint (leader-only). Sorted by own-process memory.

HarnessnVerifiedOwn RSS median (min–max)Tree medFootprint medAux medWall med

anv (--no-mind)4949/4925.6 MB (20.5–30.5)26.2 MB12.0 MB0.0 MB15.7 s

jcode4444/4445.0 MB (42.6–49.6)45.2 MB11.0 MB0.0 MB25.1 s

codex1717/1797.2 MB (91.1–291)127.6 MB22.0 MB36.5 MB68.0 s

aider4638/46248.9 MB (241–300)251.3 MB200.0 MB0.0 MB22.8 s

claude3432/34432.1 MB (386–500)439.6 MB180.5 MB27.1 MB55.9 s

opencode4848/48582.3 MB (523–655)585.9 MB440.5 MB7.2 MB21.3 s

Idle floors (trivial-prompt median own RSS): anv 21 · jcode 43 · codex 93 · aider 242 · claude 390 · opencode 537 MB. Tree adds every helper the harness spawns; footprint is the kernel’s phys_footprint, leader-only — compare footprint to footprint and tree to tree, never across. Node (opencode, claude) and Python (aider) carry runtime floors no context management removes. n=1 per cell in the matrix, n=3 on the two heaviest classes; ±6% measured session-to-session noise, and differences below that are not claimed.

Per model

Same Harness, Different Models.

Memory is a runtime property, not a model property: each harness’s median barely moves across qwen-3.8-27b, deepseek-v4.1-flash and glm-5.3-flash. Wall time moves a lot — that column is the model.

HarnessModelnMedian own RSSVerified

anvglm-5.3-flash1623.6 MB16/16

anvqwen-3.8-27b1725.0 MB17/17

anvdeepseek-v4.1-flash1626.8 MB16/16

jcodeglm-5.3-flash1544.8 MB15/15

jcodedeepseek-v4.1-flash1345.0 MB13/13

jcodeqwen-3.8-27b1645.1 MB16/16

codexqwen-3.8-27b1797.2 MB17/17

aiderdeepseek-v4.1-flash15248.5 MB12/15

aiderglm-5.3-flash15248.9 MB12/15

aiderqwen-3.8-27b16249.5 MB14/16

claudeqwen-3.8-27b18428.4 MB16/18

claudedeepseek-v4.1-flash16438.6 MB16/16

opencodeglm-5.3-flash16573.8 MB16/16

opencodeqwen-3.8-27b18582.2 MB18/18

opencodedeepseek-v4.1-flash14588.8 MB14/14

codex ran on qwen-3.8-27b only (it requires the Responses API and cannot set opencode’s session header); claude cannot use glm-5.3-flash (opencode’s Anthropic endpoint returns 500 for it); aider reaches opencode through the benchmark’s local header-injecting pass-through. Every skip is printed by the runner and recorded.

Stability · n=3 replication

What Was Real, What Was One-Off.

The two heaviest classes — pressure_read and edit_tool — were re-run at n=3 across every harness/model pair. Every earlier outlier became a classification.

01codex edit_tool +214% → one-offA single serial-anchor run peaked at 291 MB; three follow-ups landed at 92–146 MB. The cell is not reproducibly memory-hungry.

02claude pressure_read abort → intermittent2 of 4 attempts aborted (843 s, 546 s) with the auto-compact guard; 2 passed (628 s, 85 s). A coin-flip at a 128K window, not a deterministic ceiling.

03opencode rc=1 → transient, still verifiedOne attempt in four exited non-zero after writing the correct answer; three follow-ups exited cleanly. Recorded as verified with clean_exit=false, not treated as a pattern.

Failures

228 Of 238 Verified. All Ten Listed.

Zero timeouts and zero idle-kills across the campaign. Every miss below is a capability or context result, published with the same prominence as the wins.

HarnessCellWhat happenedKind

aiderwide_read ×3 modelsNo filesystem search: read the repo-map summary, asked the user to paste files.behaviour

aideraggregate_scan ×3 modelsWrote and applied a compute script, told the user to run it — answer file never appeared.behaviour

aiderbuild_pkg (glm)Did not produce a passing implementation in one attempt.behaviour

aideredit_tool (deepseek, t3)Stopped to ask “Which would you like?” instead of acting — assistant behaviour, not agent behaviour.behaviour

claudepressure_read (qwen, 2 attempts)Auto-compact guard aborted at 843 s and 546 s; two other attempts passed (628 s, 85 s) — intermittent at a 128K window.context

Rerun It. Or Dispute The Method.

Everything above is derived from the published records by the report script. Criteria, corpus generator, runner, and every JSONL record — including failed attempts and superseded runs — are public.

records/238 Phase 1 runs + 66 earlier runs, append-only JSONL · github.com/NaMan6122/Anvaya-Benchmark

205507698927dd67corpus v3 hash (483 files)

42542492c9d0a058anv build fingerprint (pre-release --no-mind)

MIT · CC BY 4.0code · data licenses

Claim policy: what is supported is memory, speed and verified completion on these models, corpus and machine — not a quality verdict, and not “the world’s lightest” until the mode ships, the comparison is run on default configurations, more harnesses are covered, and a non-macOS platform is tested. Disclosure: anv is our product.

Stop Starting From Zero.

One binary. 13+9 Rust crates. 545 tests. Hand-written HNSW index. Three transport modes. Four providers, Ollama, Anthropic, OpenAI, Siemens. Zero API keys required to start. Mind remembers everything after the first session.

Get Started →Read the FAQ

Requires Rust/cargo to build from source. Linux and macOS today, Windows not yet supported. Pre-1.0, public beta — pricing TBD.