Weekly Research— 2026-05-18

[github-actions[bot]]

Anthropic Platform Signals

The June 15 billing split is the most consequential Anthropic policy event in months — and it directly undercuts the plugin's current core pitch.

On May 13, 2026, Anthropic announced that effective June 15, all programmatic usage moves off subscription rate-limit pools into a separate monthly credit bucket metered at standard API list prices. Affected: claude -p, Claude Code GitHub Actions, and Agent SDK consumers. Explicitly excluded and staying on subscription limits: interactive Claude Code in the terminal, Claude.ai chat, and Cowork.

Credit amounts by plan:

Plan	Monthly credits
Pro	$20
Max 5x	$100
Max 20x	$200

Credits expire monthly, do not roll over, and cannot be pooled across team members. The community math (from MagnaCapax's canonical gist) shows a "one PR review per push, 4 pushes/day" CI workflow burns ~$25/month at Sonnet 4.6 rates — a Pro subscriber exhausts credits before month-end.

Why this matters for the plugin: auth.md:127 currently states the Anthropic-side policy is "stable" and the decision tree at auth.md:24 describes the OAuth path as "preferred — free for subscribers." Both claims become false on June 15. The plugin needs updated documentation before that date.

Background on the OAuth crackdown timeline (for context):

• Jan–Feb 2026: Anthropic enforced ToS against tools (OpenClaw, Roo Code, Goose) that extracted OAuth tokens from Claude Code to bypass API billing.
• Apr 2026: Outright ban on third-party Agent SDK consumers on subscription plans.
• May 13, 2026: Reversal with metering — ban lifted, but programmatic use now draws from the capped credit pool. The "reconciliation" position.

The official @anthropic-ai/claude-code CLI path documented in auth.md:7 remains permitted. What changed is that it is no longer effectively free — it now burns from a finite monthly pool.

Sources: [The Register]((www.theregister.com/redacted), [InfoWorld]((www.infoworld.com/redacted), MagnaCapax gist, [devtoolpicks.com]((devtoolpicks.com/redacted)

---

Claude Code Plugin Ecosystem

The ecosystem has grown dramatically; structural consolidation is happening around the official plugin model.

• Anthropic now maintains anthropics/claude-plugins-official with a formal architecture (slash commands + hooks + sub-agents + skills per plugin). A new blockedMarketplaces enterprise config key lets orgs restrict to approved sources.
• v2.1.143 (May 15): Plugin dependency enforcement — claude plugin disable refuses if another enabled plugin depends on the target; projected per-turn context cost now shown in /plugin marketplace browse.
• v2.1.139 (May 11): Agent view (Research Preview) — unified list of all Claude Code sessions across running/blocked/done states.
• --plugin-url flag added: fetch a plugin archive from a URL for the current session without marketplace registration.
• quemsah/awesome-claude-plugins (May 16) now indexes 17,279 total Claude Code plugin repositories. hesreallyhim/awesome-claude-code is at 36.8k stars as the canonical hand-curated list.
• Notable third-party releases: claude-forge v3.0.2 (11 AI agents, 36 commands, 15 skills), ashlr-plugin (57% token reduction on Read/Grep/Edit/Bash), maestro-orchestrate (22-subagent orchestration), obra/superpowers (94k stars, accepted into the official marketplace).

Structural note: Skills, commands, and hooks are now formally codified as distinct primitives that a single plugin can bundle together. This plugin currently uses pure-instruction skills only — no hooks or commands. Adding a hook (e.g., to re-apply the OAuth tweak automatically after a gh aw compile event) would be structurally possible.

Sources: Claude Code changelog, [claudemarketplaces.com]((claudemarketplaces.com/redacted), [Releasebot]((releasebot.io/redacted)

---

gh-aw Upstream Activity

Very active: 9 releases in the last 14 days. Several items are directly relevant.

Version	Date	Notable for this plugin
v0.74.4	May 17	Submodule credential cleanup, repo-memory file cap → 100 KB, integer overflow security hardening
v0.74.3	May 15	slash_command and label_command triggers go stable; glob patterns in add-labels safe outputs
v0.74.2	May 14	gh aw compile --staged flag; gh aw fix codemods; max-runs default raised 100 → 500
v0.74.1	May 13	Pre-activation author gating, concurrency queue support
v0.72.1	May 7	gh aw lint command; first-party coding-agent skill for gh aw CLI
v0.72.0	May 6	Inline sub-agents now default-enabled; context-aware Docker availability checks

Retirement notice still active: Versions v0.68.4–v0.71.3 are being retired due to a billing bug — any users on those should upgrade.

Auth-related items (directly relevant to this plugin's tweak):

• CLAUDE_CODE_OAUTH_TOKEN is still not natively supported by gh-aw (removed in PR #16757, Feb 19, 2026). The post-compile two-pass sed tweak in auth.md:52–62 remains the required bridge. Issue #16498 tracks re-introduction; PR #20473 (merged Mar 11) added AuthDefinition scaffolding that could host proper OAuth re-enablement — no ETA.
• CLI bundle update (Issue #31172, May 9): fixed a rare login loop from concurrent credential writes overwriting a rotated OAuth token. Low risk, positive for Claude engine users.
• OIDC token support for custom HTTP MCP server auth (Issue #23566, open since Mar 30): feature request to eliminate static long-lived API keys via OIDC workload identity federation. No resolution yet but relevant to the direction gh-aw auth is heading.

Catalog activity (githubnext/agentics):

• New workflow: dictation-prompt (PR #325, May 5) — runs weekly or on manual dispatch; scans project docs, builds an NLP word-frequency histogram of project vocabulary, generates DICTATION.md (~256 terms) for speech-to-text correction accuracy. Auto-merges the PR.
• PR #332 (May 12): Issue triage template revamped with structured quality gates and automated labeling.
• PR #330 (May 9): repo-assist now supports direct task invocation via gh aw run -F command=....
• PR #335 (May 13): README expanded with non-coding use cases (product, ops, compliance, leadership).

Engine news: OpenCode is now a first-class gh-aw engine option (engine: opencode), alongside Copilot, Claude, and Codex.

Sources: github/gh-aw releases, githubnext/agentics, gh-aw Issue #16498, gh-aw Issue #31172

---

Competitive Landscape

Two new competitive signals this week.

microsoft/apm (Agent Package Manager) — the most direct competitive signal. Capabilities: apm install deploys agents/skills/prompts to Claude Code, Copilot, Cursor, Codex, Gemini, and OpenCode targets via a single apm.yml; auto-discovers harness targets from filesystem signals (.claude/, .github/copilot-instructions.md, etc.); supports versioned packages; auto-discovers .prompt.md files as runnable scripts. Updated May 15, 2026. It installs agentic content into repos — the same core operation this plugin performs. Differentiation to clarify: this plugin is gh-aw-catalog-specific and subscription-auth-aware; APM is multi-target but auth-agnostic.
Source: github.com/microsoft/apm

anthropics/claude-code-action v1.0 — Major milestone released recently. Key changes: unified prompt input replaces legacy separate inputs; all CLI options now flow through claude_args; automatic mode detection (interactive vs. automation); enhanced AWS Bedrock and Google Vertex AI compatibility. This is the competing "official" path for Claude in GitHub Actions. Updated May 15, 2026.
Source: anthropics/claude-code-action releases

zircote/github-agentic-workflows (aw-author): Active, well-scoped plugin for authoring gh-aw markdown files (/aw-author new|generate|validate|improve|debug). Complementary rather than competing — they write workflows, this plugin installs them.

Broader harness space: Harness AI Worker Agents now auto-generates pipeline YAML from chat. OpenCode emerging as a reference CI harness. Stripe's "Minions" (internal) cited at 1,000+ merged PRs/week as the benchmark for what's possible.

---

Subscription-Backed CI Signals

Community sentiment has shifted sharply negative following the May 13 announcement. The risk window is June 15.

• The math is clear: after June 15, running Claude in CI on a Pro account ($20 credit/month) for a typical "4 pushes/day, 1 review each" workflow costs ~$25/month — exceeding the credit allotment before month-end.
• Anthropic's Lydia Hallie framed the change as "free credit"; the post was Community-Noted within hours. Theo (T3 Code), Jeremy Howard, Matt Pocock, and Omar Sanseviero were among the prominent critical voices. DHH called an earlier related move "very customer hostile."
• A related Hacker News thread received 245+ points; a anthropics/claude-code-action GitHub issue received 147+ reactions.
• Zed published a direct user-facing post explaining what the changes mean for their users — confirming they're an affected third-party.
• Partial concession: Anthropic announced a 50% increase in Claude Code interactive weekly limits through July 13, stacked on a prior 2x increase. Read widely as a goodwill gesture to offset the programmatic backlash — but this doesn't help CI usage.
• Operationally safe path post-June-15: A direct ANTHROPIC_API_KEY with a hard monthly spend cap. The plugin's OAuth path survives but is no longer "free."

Sources: MagnaCapax gist, [Zed blog]((zed.dev/redacted), [VentureBeat]((venturebeat.com/redacted), [dev.to analysis]((dev.to/redacted)

---

Strategic Suggestions

1. Update auth.md and the decision tree before June 15 — this is urgent

auth.md:24 currently says the OAuth path is "preferred — free for subscribers." auth.md:127 calls the Anthropic-side policy "stable." Both need updating. The OAuth path remains permitted but is no longer free — it burns from a $20/month (Pro) or $100–$200/month (Max) non-rollover credit pool. Add a cost-estimation note (e.g., "a 4-push/day PR review workflow costs ~$25/month at Sonnet 4.6 rates — verify your plan's credit allotment before committing to the OAuth path for heavy CI use"). Update the decision tree to surface plan tier and estimated usage as a factor. Failing to update before June 15 means the plugin will install workflows that silently fail once subscribers exhaust credits.

2. Add a gh-aw version check to the OAuth tweak step with an exit for native support

auth.md:129 documents that gh-aw PR #20473 (Mar 11) added AuthDefinition scaffolding that could host native CLAUDE_CODE_OAUTH_TOKEN re-enablement (tracked in issue #16498). When that lands, the two-pass sed tweak in auth.md:52–62 can be retired. Add a version-gated check to install-workflow/SKILL.md Step 6: detect if the installed gh-aw version is >= the release that includes native OAuth support, and skip the tweak if so. This prevents the tweak being applied when it's no longer needed (and no longer correct). Document the threshold version in auth.md as a TODO: bump when #16498 ships note.

3. Differentiate clearly against microsoft/apm in the README and discovery output

microsoft/apm now does "install agentic content into repos" with multi-target support. The differentiation for this plugin is specific and worth making explicit: (a) it is gh-aw-catalog-aware — knows the githubnext/agentics catalog structure, proper workflow names, and the engine: claude frontmatter requirement; (b) it is subscription-auth-aware — no other tool guides users through the claude setup-token → two-pass sed → grep verification flow; (c) it lives inside Claude Code natively (no extra CLI to install). Add a one-paragraph "Why not just use apm?" section to the README and update the first sentence of discover-workflows/SKILL.md to position against the multi-target "install anything anywhere" tools that don't know gh-aw's conventions.

---

Enjoyable Anecdote

The newest entry in the githubnext/agentics catalog is dictation-prompt (PR #325, May 5). It runs weekly, scans your project's docs and source comments for unusual vocabulary, builds an NLP word-frequency histogram, and generates a DICTATION.md file of ~256 project-specific terms — names like gh-aw, awf, sk-ant-oat01, lock.yml — tuned to improve speech-to-text correction accuracy for developers who code by voice. It auto-merges the PR. It's a delightfully niche workflow: a coding assistant that helps you talk to your coding assistant more accurately. The catalog continues to expand beyond pure CI automation into territory that would have been too quirky to ship even a year ago.

---

Research appendix — queries, commands, and tools used

Web search queries

• anthropic platform policy changes oauth subscription june 2026
• anthropic agent sdk credit billing split june 15 2026
• anthropic consumer terms of service 2026 update
• claude code plugin ecosystem marketplace 2026 may
• hesreallyhim awesome-claude-code new plugins may 2026
• claude code changelog v2.1.139 v2.1.143
• gh-aw upstream releases may 2026
• githubnext agentics catalog new workflows may 2026
• zircote aw-author github agentic workflows plugin
• anthropics claude-code-action v1.0 release
• microsoft apm agent package manager github
• subscription claude ci community sentiment reddit hackernews
• anthropic solo founder hiring ai tooling engineer 2026
• context engineering portfolio claude code

Repository files read

• /skills/install-workflow/auth.md (lines 1–140)
• /skills/install-workflow/SKILL.md (lines 1–40)
• .github/workflows/weekly-research.md
• /README.md (via explore agent)
• /skills/discover-workflows/SKILL.md (via explore agent)
• /skills/install-agent-team/SKILL.md (via explore agent)
• /catalog/agent-team/README.md (via explore agent)
• /.claude-plugin/plugin.json (via explore agent)

Bash commands

• git log --oneline -20 (via explore agent)

GitHub MCP tools

• mcp__github__list_discussion_categories — verkyyi/github-agent-runner (to find the Ideas category ID)

Generated by Weekly Research · ◷

To install this agentic workflow, run

gh aw add githubnext/agentics/workflows/weekly-research.md@96b9d4c39aa22359c0b38265927eadb31dcf4e2a

• expires on May 25, 2026, 1:45 PM UTC

[github-actions[bot]]

This discussion was automatically closed because it expired on 2026-05-25T13:45:20.769Z.

Closed by Workflow