Code

Progressive reply publication: decomposition and aggregation follow-ups

Daily ironclaw failure taxonomy — 2026-08-31

feat(attestation-cli): report which upcoming TCB evaluation data set demotes the platform

tool schemas: flatten_top_level rebuilds from a whitelist, silently discarding every non-forbidden top-level constraint

perf(github, tools): a raw list_repos payload plus an unhinted result_read schema cost 64 tool calls and 3m01s to list repos

feat(llm): surface NEAR AI model capabilities across model-selection UI

feat(gmail): enforce HTML complexity during DOM construction

test(learning): observability, evaluation, and provider migration gates

feat(skills): route shared learning review into skill distillation

feat(memory): bounded active recall from admitted provider memory

feat(memory): native, mem0, and Mnesis learning capability adapters

feat(memory): deterministic admission and auto-or-approval promotion

feat(memory): stable commit, feedback, and forget capabilities

feat(learning): shared router, settings, and durable candidate store

MCP OAuth 2026-07-28: send resource indicator and prefer CIMD over DCR

Salvage still-relevant work from superseded abbyshekit PRs

feat(webui): stream large thread-artifact downloads

Daily ironclaw failure taxonomy — 2026-08-27

Nightly 2026-08-27: Reborn Playwright (stale landing-copy fixtures + attachment flow) and Postgres stress API 503 readiness gate failing

Judge foreign chain health against the whitelist

Foreign TX Validation security hardening

Look into solutions for exporting logs from TDX nodes

design(webui): select and plan the voice-to-text architecture

design(webui): host-normalized voice-to-text architecture

design(webui): browser-normalized voice-to-text architecture

Provide script/tool to configure unattended upgrades

perf(tools): allow tool arguments to cite a prior result by reference instead of re-emitting it

obs(loop): instrument per-run model-call, prefix-cache, and tool-churn metrics so perf work is decidable

feat(tools): declare apply_patch as a grammar-constrained freeform tool to eliminate JSON-escaped diffs

perf(llm): OpenAI-family backends send no prompt_cache_key — measured ~82%→29% cache-hit collapse past ~200 calls