v0.20.1: free-quota budget — halve auxiliary call volume, keep the daily daily

[hxddh]

Summary

Implements the agreed token-budget plan (owner approved): cut ~50% of call volume with minimal product damage, instead of thinning the daily cadence.

Change	Was	Now
source-sweep	every day	Mon + Thu
promote-candidates	Wed + Sun	Sun
Screening shards (free mode)	4 calls	2 calls (SCREENING_SHARD_GROUPS=2: discussion+official merged with discussion first; github+packages merged; =4 restores the full split)
daily / weekly / monthly	—	unchanged

Estimated effect: ~6 → ~3.5 gateway calls per day, ~3M → ~1.6M tokens/month, free-tier 429 pressure roughly halved. Merged screening ceiling drops 96 → 48 candidates — still ample for 8–10 published signals, and the deterministic Radar Sweep (v0.20.0) guarantees everything screened still reaches the report.

Rejected alternatives (analysis in conversation): dropping the monthly saves ~2%; a 3-day daily cadence breaks the 24–48h freshness machinery and inflates single-run output exactly where free models are weakest.

Validation

• 259 tests pass: new regressions for the Mon/Thu/Sun schedule and the 2-group shard merge; two stale schedule tests updated.
• python scripts/agent_radar.py validate passes; workflow YAML parses.

🤖 Generated with Claude Code

https://claude.ai/code/session_01JXjb8D4NrsgUXAvQodQSum

---

Generated by Claude Code

[chatgpt-codex-connector]

💡 Codex Review

agent-radar/scripts/cloud_agent_runner.py

Line 4015 in edaf3e1

inject_deterministic_radar_sweep(result)

Keep sweep URLs out of repeat-url repair

When a screening candidate in the auto-generated Radar Sweep has a URL already published in a recent daily, this injection runs before repair_repeated_url_freshness(). That repair uses split_daily_signal_bullets(), which does not recognize - [class] ... sweep lines as standalone bullets, so the sweep section is treated as part of the preceding signal and the Freshness: follow-up sub-bullet can be appended under #### 8. Assessment & Gaps instead of to the stale sweep item, corrupting normal day blocks with resurfaced candidates. Exclude Radar Sweep from that repair or teach the repair to handle sweep lines explicitly.

---

agent-radar/scripts/cloud_agent_runner.py

Line 5612 in edaf3e1

attempts = [(r, m) for r in range(rounds) for m in models]

Limit extra retry rounds to 429 responses

With the default AI_GATEWAY_429_ROUNDS=3, attempts repeats the entire fallback chain for every retryable failure, not just HTTP 429. In a gateway/model outage that returns 500s, invalid JSON, or transport timeouts, a daily synthesis now can make up to 3x as many gateway attempts and wait through multiple long MODEL_TIMEOUT periods for non-cheap models before failing, which undermines the quota-saving goal and can push the workflow past its runtime budget. Build additional rounds only after a 429, or stop after the first full pass for other transient errors.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".