Final validation: tiny-harness: A tiny agent harness! (issue-3)

The evidence node's proof: what is presented to the user to show the acceptance criteria are met. Summarised from the executed verification — testing-plan.md's Verification results — and mapped onto the acceptance criteria rather than re-derived. Raw captures sit beside this file under evidence/.

Everything below ran on 2026-10-10 at the head of the stacked series (PR #15), after the review rounds; the counts are the results table's.

Final validation evidence

Acceptance criterion	How it was proved	Where
R1 entities and registry	unit tests of the registry (PEP 440, *, not-found, remote protocol, override); API snapshot	unit.md, contract.md
R2 hooks on every operation	unit tests of the chain, pre/in/post replacement, abort, remote executors; integration: activity.failed/activity.retried between attempts	unit.md, integration.md (durable scenarios)
R3 plugins	unit tests of the loader; security tests of path escape and ${VAR} rules; the demo plugin loaded live	unit.md, security.md, e2e.md
R4 system prompt, R5 skills	unit tests of sections, front matter, disclosure levels; the demo's support-agent skill loaded by load_skill in the e2e trace	unit.md, e2e.md
R6 tools through MCP	unit tests of validation and idempotency; security tests of unknown tools and schema drift; the demo's two stdio servers called live	unit.md, security.md, e2e.md
R7 remote agents	integration: delegation to a second harness tracks the remote task; security: a card requiring unsupported extensions refused	integration.md, security.md
R8 task, R9 plan	unit tests; task extension and intrinsic schemas pinned; plan visible in both renderers	unit.md, contract.md, ui/axe.md
R10 context window	unit tests of ordering, never-compact set, compaction record, untrusted framing; cached_tokens from call 2 in the live run	unit.md, e2e.md
R11 persistence	unit tests of the store; push tokens encrypted; retention sweep in the heartbeat scenarios	unit.md, security.md, integration.md
R12, R13 participants and channels	integration: help request → INPUT_REQUIRED → reply resumes; non-member refused; fail-closed assertions	integration.md, security.md
R14 A2A server	integration: every verb, streaming, replay after restart, unadvertised extension refused; agent card snapshot	integration.md, contract.md
R15 inbox and events	integration: a message to an executing task drained next iteration, returnImmediately, push configs, task envelopes on an existing task	integration.md
R16 heartbeat	integration: schedule, poll, snapshot, sweep	integration.md
R17 observability	integration: one span per operation with GenAI attributes; redaction tests; the live trace	integration.md, security.md, e2e/trace.json
R18 models	unit tests of both adapters, retries off, wire names; gpt-6.1-sol live	unit.md, e2e.md
R19 durable execution	integration: crash mid-activity resumes without a second LLM call, non-idempotent failure not retried, continue-as-new; live kill -9 at two points with the activity attempt tables from Temporal history	integration.md, crash-recovery.md
R20 surfaces and renderers	TUI snapshots and keys; web unit, visual and axe suites; both surfaces captured live, the second-surface subscription	ui/axe.md, evidence/ui/
R21 configuration	unit tests of secrets and unknown keys; the demo configuration	unit.md, gates.md
R22 code structure and typing	pyright strict 0 errors, no Any; layout tests	gates.md, unit.md
R23 contract tests	29 contract tests, 14 snapshots	contract.md
R24 the demo	the e2e run: plan, model, MCP tools, A2UI card with an action, a two-turn help request, COMPLETED; both renderers; crash recovery on Temporal Cloud	e2e.md, crash-recovery.md, ui/
Security considerations (abuse cases 2–9)	one negative test per case	security.md, security-review.md

Not run: T11, the owner's manual walkthrough, closed by the approver's decision (manual-walkthrough.md).