Skip to content

fix: TS2352 cast in engine:'pi' loader test - #406

Merged
LinuxDevil merged 17 commits into
mainfrom
integration/harnesses
Oct 6, 2026
Merged

fix: TS2352 cast in engine:'pi' loader test#406
LinuxDevil merged 17 commits into
mainfrom
integration/harnesses

Conversation

@LinuxDevil

@LinuxDevil LinuxDevil commented Oct 5, 2026

Copy link
Copy Markdown
Owner

Summary

Follow-up to #405 — that PR merged while CI was still failing on a typecheck error in loadAgentDir.test.ts:233 (PerRun<ToolsOption> can't be directly asserted to an array type; needs as unknown first). This lands the one-line fix so main typechecks again.

Test plan

  • npm run typecheck clean locally
  • npx vitest run src/agentDir/loadAgentDir.test.ts — 28/28 pass

Generated with Devin

LinuxDevil and others added 17 commits October 5, 2026 18:43
@LinuxDevil @devin-ai-integration
test: fix TS2352 cast in engine:'pi' loader test
3308988 
Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
fix: self-contained piAgent doc snippet + 8GB heap for CI builds
60c0c56 
The merged DTS graph OOMs tsup's worker at the 6144MB NODE_OPTIONS cap
on zod4 peer jobs - bump to 8192 matching the build script. The
docs:verify-snippets gate also flagged sub-agents.md's piAgent example
for undeclared projectDir/ask/allow - import the helpers and use
process.cwd().

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
fix: 12GB heap for DTS builds + self-contained piAgent doc snippet
e1f2689 
The zod4 peer matrix still OOMs tsup's DTS worker at 8GB - zod@4/ai@6-7
grow the single-pass ~30-entry declaration graph well past the light
zod3 local set. NODE_OPTIONS does propagate to the worker thread
(verified: heap_size_limit follows it), so 12288 raises the real cap;
ubuntu-latest runners have 16GB. The docs:verify-snippets gate also
flagged sub-agents.md's piAgent example for undeclared
projectDir/ask/allow - import the helpers, use process.cwd().

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
refactor: satisfy fallow dead-code, dupes, and complexity gates
95087f3 
- pi.worker.ts added to dynamicallyLoaded (bundler shim, not statically reachable)
- unexport internal-only symbols (familyInstructionsFor, AUTH_FILE, loopGuard/outputCap); drop dead PiContent union
- registry/** + coding-harness example excluded from dupes: installable kits are self-contained by design
- split zod3ToJsonSchema/applyChecks into per-kind dispatch tables (cyclo 55/38 -> ~5)
- split piAgent.run into openSessionManager/startSession/deliver/settle (cyclo 40 -> ~10); the tool_call gate handler is a named function so coverage matching attributes it
- split chunkStream's event switch and toPiContext's role chain below the cognitive ceiling
- thresholdOverride for piAgent.ts: nested-callback CRAP matching gap, same as registryEnforce's documented case

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
test: cover the pi zod3->JSON-Schema converter (fallow CRAP)
bc5a0f4 
applyChecks/applyBound/wrappedType sat at 0-27% coverage - CRAP breaches
on untested complexity. Hand-built {_def.typeName} fixtures exercise
every converter branch deterministically under both zod3 and zod4 peer
installs.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
docs: agent-landscape research + replication plan
f071823 
Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
feat: archetype replications + SDK seams they exposed
e2e985c 
Six deliverables replicating the most-built agent archetypes, each
offline-testable and OpenRouter-live-capable:

- examples/deep-research: orchestrator-workers fan-out, cited synthesis
- examples/support-desk: triage -> handoff() to specialists, refund approval
- examples/evaluator-loop: writer <-> llmCritique revision loop
- examples/data-analyst: text-to-SQL with read-only defense in depth
- registry/inbox-triage: channel intake, classify, draft, approved send,
  per-sender memory, hourly sweep
- registry/incident-response: webhook alert -> diagnose -> risk-tiered
  approvals -> remediate -> verify -> audit

SDK changes the builds exposed:

- llmCritique(): judge score + actionable feedback in one call
- testToolContext() + toolResultText() testing helpers
- handoff calls now pass the permission/hooks/approval gate; structured
  routing args reach the target; run-level target options warn once
- approval TTL: approvalTtlMs (agent/run/rule), expiresAt propagates
  through pause/resume, stale decisions deny with 'approval expired'
- 'defer' approval verdict; resumedAfterApproval hook flag; transient
  tool descriptors skip agent-drift noise on resume
- declarative when matchers (eq/ne/lt/lte/gt/gte/matches) + approvalTtlMs,
  rule ttlMs and store { dir } in agent directories
- ChannelInbound.metadata flows to run/memory scopes; webhookChannel
  principal() derives an authenticated caller
- handoffRun deep-imports ToolRegistry so the Worker bundle stays node-free

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
test: live harness tests gated on OPENROUTER_API_KEY (never in CI)
a670944 
*.live.test.ts is already excluded from the default vitest config and no
workflow runs test:live, so these are opt-in regression checks a human runs
with a real key to prove a change did not break a live model's ability to
drive a harness:

- examples/harnesses.live.test.ts: createAgent tool call, approval
  pause/resume, deep-research fan-out + citations, support-desk handoff to
  billing with a gated refund, evaluator-loop through llmCritique,
  data-analyst read-only SQL - all on openrouter/openai/gpt-4o-mini
- registry/kits.live.test.ts: inbox-triage and incident-response loaded
  through loadAgentDir() exactly as installed kits, state isolated via the
  kits' own env-var homes
- vitest.live.config.ts include widened to examples/ + registry/
- .env.example documents the key

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
fix: scaffold a resolvable SDK dep range for docker/node-server builds
51eec4b 
deployDependencies() pinned @lousho/build-ai-agent to the exact local
package version; a scaffold built from an unpublished dev version
(1.0.0-alpha.4) made the image's `npm install --omit=dev` fail with
ETARGET. Emit `^version` instead, the same convention sync-version.mjs
uses for the create-lousho-agent pin, so an in-development build resolves
to the nearest published release.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
docs: regenerate llms-full.txt for agent-directory when-matcher/store…
b55821f 
… docs

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
refactor: satisfy fallow dead-code, dupes, and complexity gates
ac8361e 
- extract toolOutcomeMessage() so normal + handoff paths share the
  tool-result transcript shape (undefined -> null, error payloads,
  replacedByHook metadata)
- share the resumed-hook context builder across resume paths
- validateConfig: OPERAND_KIND dispatch + operand/store/rule helpers;
  numeric comparisons via a table (fail-closed semantics unchanged)
- loadAgentDir: split assembleConfig (null-override semantics preserved)
- createAgent: extract resumeRequester/agentSession factories and the
  runLevelOptions/subagent-options/approval-store/clash helpers
- ignore examples/** for duplication (standalone demos, like kits)
- unexport handoffRun internals only used in-file

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
chore: remove internal plan/research/design docs from the repo
51cf80f 
Internal working docs (docs/plan, docs/research, .design-ref) are not
user-facing; drop them and the README link to the validation report.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
test: refresh OpenRouter cassettes + harden pi live test instructions
432151d 
- Live OpenRouter run re-recorded the provider/execution cassettes;
  replay tests still pass (verified).
- piProvider live test: gpt-4o-mini was replace_all-ing the identical
  add/subtract bodies and then looping on read_file line-number
  prefixes; instructions now require unique-context old_string, forbid
  replace_all, warn about the line-number prefixes, and maxSteps 10->14.
- lockfile: create-lousho-agent template dep synced to ^1.0.0-alpha.4.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
fix: load .env into process.env for npm run test:live
507d50a 
Vitest only exposes .env on import.meta.env, so live tests' skipIf(
process.env.OPENROUTER_API_KEY) checks skipped even with a populated
.env. loadEnv() in the live config merges it, as the comment promised.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
docs: list inbox-triage/incident-response kits + new archetype examples
25d6eff 
The registry table still said seven items and the changelog's install
flags named network,env while both kits' manifests declare env +
filesystem: write. README's example table now lists the four new
archetype examples and plan-mode.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
docs: README for data-analyst + support-desk; bump tarball size budget
0d4fefd 
docs-links.test.ts requires every README example link to resolve to a
non-empty README.md. The two missing example READMEs also feed
llms-full.txt, which is shipped in the tarball; with them the packed
size crossed the 4.5 MiB budget by ~11 KB, so it goes to 4.6 MiB (the
unpacked budget was already raised the same way in 4070302).

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil @devin-ai-integration
docs: regenerate llms.txt for the new example READMEs
21d3d0d 
The check gate regenerates both files; llms-full.txt was committed but
llms.txt was missed.

Generated with [Devin](https://devin.ai)

Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
@LinuxDevil LinuxDevil self-assigned this Oct 6, 2026
@LinuxDevil
LinuxDevil merged commit 2cfc1be into main Oct 6, 2026
9 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

@LinuxDevil LinuxDevil

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

1 participant

@LinuxDevil