[leggetter]

What this is

Docs-only changes for the v0.5.0 release: two improvement loops measured by the 5 and 6 November runs, and corrections to statements the 5–6 November runs and merges made stale.

What changed

• LOOPS.md, Loop 3 (Does documenting hookdeck ci stop agents giving up on CLI authentication? #27, Does documenting hookdeck ci stop agents giving up on CLI authentication?): event-gateway: document hookdeck ci for non-interactive environments agent-skills#27 reached the benchmark through Moved the agent-skills pin to 84e59f4, and recorded it per result #102's pin move (6ea10a4 → 84e59f4). Per attempt, over attempts that invoked the CLI at least once, "the first invocation other than help or a version was hookdeck ci" went from 0/52 across the skills arms on 28 September (run 36386178039, runs=1) to 47/58 on 5–6 November (runs 42044644850 and 42220222687, --runs 3): Claude Code 0/19 → 14/18, GPT-5.6 0/15 → 17/18, GPT-5.4-mini 0/18 → 16/22. Baselines read 0/65. No score is attributable: CLI 2.5.0's error message already names hookdeck ci, and the baselines recover through it.
• LOOPS.md, Loop 4 (The Outpost skill never says you need a key for an Outpost project #40, The Outpost skill never says you need a key for an Outpost project): inconclusive. Skills arms already used OUTPOST_API_KEY on 28 September. outpost-002 on codex-gpt-5.4-mini failed all three attempts on 5 November by working in the Event Gateway project after two of them had read the new skill section, then passed on 6 November. That failure was not the Agent-created Event Gateway resources in the Outpost project survive every reset #111 leak.
• LOOPS.md, Loop 1: the single-attempt warning now says the September matrices were runs=1 too.
• AGENTS.md, Status: the 6 November snapshot at best of three (1/19 scenarios with a failing cell; 12/114 cells, and 102 first-attempt passes, as in the restored latest.json) set against 28 September's single-attempt 9/19; the weak-pair table gains 5 November (−1) and 6 November (+1) and says the earlier six rows were single-attempt; the latest.json pairing note matches its composition; the hookdeck listen TTY and guest-session sentence says what CLI 2.5.0 is known to do.
• AGENTS.md, Costs: the measured Claude Code cost for 5 November ($35.46 for 38 cells, a floor because only final attempts record usage), measured attempts per cell (1.12), and a correction to the "4 of 114 rows carry costUsd" line.
• .plans/delivery-plan.md: the default issue triggers are declared and restored since Reset issue triggers to declared defaults, not snapshot ids #118 (Reset issue triggers to declared defaults, not snapshot ids), with the alerting-001 has run without the default delivery trigger since 19 August #117 (alerting-001 has run without the default delivery trigger since 19 August) history, instead of "verified on evals-ci"; the single-attempt warning covers every scheduled matrix up to 28 September.

How it was verified

• Run composition, attempts and pass counts from results/runs/*.json, the restored results/latest.json, and the raw-results artifacts of runs 36386178039, 42044644850, 42220222687 and 42256415982.
• runs inputs from the prepare-job logs (13 August, 5 and 6 November) and the workflow at each scheduled commit.
• Loop 3's figures come from the same parser as the release notes and the Does documenting hookdeck ci stop agents giving up on CLI authentication? #27 comment, re-run on the artifacts. Guest-session absence and Outpost key use were checked in the transcripts, superseded attempts included.
• pnpm format:check: biome and prettier pass. Neither checks Markdown here.

Not done, and not checked

• The plan's historical TTY sections and the localdev-001 scorer comment are left as written.
• Does documenting hookdeck ci stop agents giving up on CLI authentication? #27 closes with this release; The Outpost skill never says you need a key for an Outpost project #40 stays open.
• Run 42250228870's transcripts were not downloaded.

Observations outside this diff

• The results/latest.json mismatch with its run snapshot (attempts on 11 rows, docs/skills on 3 alerting-001 rows) is fixed by Restore latest.json and the web data from the run's own snapshot #121 (Restore latest.json and the web data from the run's own snapshot). The publish race that caused it is Two overlapping merge runs publish against a stale results file #127.
• Run 42250228870's post-Reset issue triggers to declared defaults, not snapshot ids #118 alerting-001 rows never reached latest.json; the published alerting-001 rows predate Mute issue trigger notifications on every reset #116 and Reset issue triggers to declared defaults, not snapshot ids #118.

🤖 Generated with Claude Code