fix(agent-platform): read the janitor's args_schema in tool schemas

[Aidan945]

Problem

Models see custom tools with no parameters. The janitor writes tools/<id>/schema.json as {description, args_schema} (writeToolSourceAndSchema), but the runner's loadCustomSchema parses {description, args} — so every typed-pipeline tool's schema silently resolves to a bare {type: "object"} and the model can only guess arguments from prose. A session transcript shows the model concluding "the tool's schema shows no explicit input parameters".

Both sides pass their own tests because the runner's fixtures hand-write schema.json with the runner's key — two hand-maintained copies of one wire contract. Among the readers of this file, the runner was the only one parsing args; readTypedBundle already reads args_schema.

Changes

File	Change	Why
agent-runner/src/loop/build-agent-tools.ts	loadCustomSchema reads args_schema	Direct fix; aligns the one outlier reader with the only production writer. No production bundle uses the legacy args key (confirmed by the maintainer), so there is a single path, no fallback
same file	warn logs on both silent paths: unreadable schema.json (custom_tool.schema_unreadable, with the error) and a parsed schema missing both keys (custom_tool.schema_missing_args_schema)	The degradation was invisible by construction; the second event covers the exact failure mode of this bug if the key ever drifts again
build-agent-tools.test.ts	regression test authors through the real writeToolSourceAndSchema	Welds writer to reader: if the canonical shape drifts again, this fails
4 files in agent-tests/src/cases/	hand-written args fixtures → args_schema	e2e now rehearses the shape production writes — removes the masking that let this ship

The key rename is the direct fix; the two warn logs and fixture flips are the durable half — they make this class of drift visible and tested rather than silent. Deliberately not in this PR: single-sourcing the schema.json codec in agent-shared (natural follow-up; kept out to avoid colliding with in-flight storage-boundary work).

Sibling fix in the same authoring loop: #69201.

How did you test this code?

Ran locally against a full dev stack (all four agent services + Postgres/Redis/Kafka/SeaweedFS):

• New regression test custom tool parameters survive the writeToolSourceAndSchema → loadCustomSchema round-trip, authored through the real writer against the real test bundle store. With the fix reverted it fails (parameters collapse to the bare {type:"object"} fallback) — the writer→reader weld no existing test provided.
• pnpm vitest run src/loop/build-agent-tools.test.ts → 25 passed; full @posthog/agent-runner suite → all assertions pass; typescript:check, oxlint, oxfmt --check → clean on agent-runner and agent-tests.
• End-to-end: before the fix, the model calls a schema'd tool with {}; after, it passes correct structured arguments. Reproduced both directions.
• The fixture flips are behavior-neutral by construction (the reader accepts both keys); they execute fully in CI's e2e job — I couldn't run that suite on this machine (kafka host alias), so that's the one place CI adds coverage beyond what I ran.

Automatic notifications

• Publish to changelog?
• Alert Sales and Marketing teams?

Docs update

No docs change needed.

🤖 Agent context

Autonomy: Human-driven (agent-assisted)

I hit this while building agents against a local stack — a model calling my custom tool with empty arguments — and traced it from the transcript to the key mismatch; the fix and tests were developed in a Claude Code session (Fable 5) under my direction, with every executable claim above verified by running it. Key decision: fix the reader, not the writer — a census of every schema.json consumer showed the runner was the single outlier, and frozen bundles are immutable, so only the reader side heals revisions that already exist.

🤖 Generated with Claude Code

[dmarticus]

thanks for this fix! I left some comments around just committing to a single direction for this change rather than supporting the legacy behavior, since we don't have any bundles in production that are actually malformed. I'll approve once that's addressed!

[Aidan945]

Makes sense. Good to get ahead of it. Dropped the fallback and the legacy test, so args_schema is the single path now.

[dmarticus]

nice work; thank you!

[deployment-status-posthog]

Deploy status

Environment	Status	Deployed At	Workflow
dev	✅ Deployed	2026-07-09 15:12 UTC	Run
prod-us	✅ Deployed	2026-07-09 15:43 UTC	Run
prod-eu	✅ Deployed	2026-07-09 15:40 UTC	Run